Home

supplement - Walter Fontana Laboratory

1. a _ 10 b 1 00 S g g Qa 8 c 7 B 0 75 z a 6 s a 55 l 2 0 50 g 4 i E wildtype 2 3 4 eee g 2 g E 0 25 age 1 hx546 oO r as oO si EE T om o E me T 0 00 rror area aaa 012 3 4 5 67 8 9 10 wildtype age 1 hx546 Posture span days Figure 6 2 Stationarity posture span a The time spans between cessation of translatory motion and death stationarity spans are shown with overlaid averages wide bar and standard deviations narrow bar b Comparison of the cumulative stationarity span distributions 1 0 minus this curve would be a survival curve whose terminal event is death and whose initial time point t 0 corresponds to cessation of translatory motion Nature Methods doi 10 1038 nmeth 2475 32 Supplementary Note 7 A death phenotype After an animal ceases to crawl the image analysis of the Lifespan Machine produces a time series of registered images of that animal This image stack typically used to quantify local movement and identify death times Note 5 section 5 2 can also be used to longitudinally monitor morphological b 1 15 w i i a wildtype 25 C wildtype 35 C a i an fo GE ai i o p Worm area fraction of area at death Worm area fraction of area at death r 4 T Me es e 0 9 cz r T r r T 1 70 1 2 3 4 Y 0 0 5 1s 1 5 gt _ rm Time relative to death a Time relat
2. 1x Soldering Iron or Gun Optional tool for construction For example a Soldering Gun available from McMaster Carr for 58 1x A power drill screw driver Optional tool for construction Unless you have wrists of steel it is easier to use a power drill to add and remove machine screws Available from McMaster Carr for 150 or for less at a home improvement store like Lowes 1x Canned air Optional tool for construction A source of compressed air either from a can often sold to clean computer keyboards sold at any office supply store or from a lab air outlet 10x Additional 80x 80x 25mm DC fans Optional Used to circulate air throughout incubator to achieve best possible temperature calibration Only required for certain experimental designs 10x Metal ductwork Optional Used to circulate air throughout incubator to achieve best possible temperature calibration Only required for certain experimental designs For example 3 Aluminum Bend and Stay Duct Hose available from McMaster Carr for 5 25 Each 2x Plastic bin for precise drying of plates Useful ancillary equipment Exact dimensions are not crucial but long and flat boxes work best For example the 35 5 8 x 18 1 4 x 6 1 4 h Long Underbed Box with Wheels available from The Constainer Store com for 20 2x Sheet of Wire Mesh fitted to plastic bin i e 18 x36 Useful ancillary equipment Mounted in plastic bins for precise drying of plates Ava
3. Fig 4 of the main text even our baseline observation frequency of 24 measurements per day permits relatively few observations potentially eliminating the advantage of large population sizes In such a case however we can increase the effective temporal resolution by using multiple scanners and staggering their scan times to effectively monitor at all times some sub population of the total population residing in the apparatus We use this technique to improve the temporal resolution of 35 C survival assays resulting in smoother aggregated survival curves Fig 4b of main ods doi 10 1038 nmeth 2475 9 Supplementary Note 16 Use cases The lifespan machine captures images of worms using flatbed scanners whose data are analyzed and integrated by software Scanners can be easily added and removed allowing the technique to be applied at small scale in installations including one or two scanners as in Fig 3c e f of the main text where four mutants were monitored in two rounds on two scanners or at large scale as in Fig 2b where one mutant was monitored across ten scanners As such the method reduces labor in a variety of experimental contexts adding scientific value by enabling new or old and recalcitrant questions to be approached with commensurate data Nature Methods One or two scanners Asingle scanner holds 16 plates with 35 worms each for a total of 560 individuals As indicated in Note 9 this capacity is sufficie
4. To prevent the build up of heat within the TPU we hang four fans F4 F5 F6 and F7 which serve both to remove heat from inside the TPU and to circulate cold air across the plates A single 12cm fan F8 is hung from the rear of the TPU which serves to remove heat from the scanner TPU the scanner chassis and moves cold air across the plates Four cuts are made into the scanner chassis and three into the TPU to permit cold air to cycle through both compartments These cuts are easily made using a spiral saw or other cutting device Our experience is that most Dremel tools aren t quite up to the task of cutting through the hard chassis plastic When working with any cutting tool wear safety goggles and always pay attention Clean uncluttered work spaces promote safe practices Three fans F1 F2 F3 are sealed in place using a glue gun Four fans are hung from the scanner TPU are suspended by a loop of nylon twine hanging from machine screws drilled into the TPU at positions S1 S2 S3 and S4 The holes can be cut with a power drill if one is available or else using the spiral saw The exact location and dimension of each modification are presented in Figure 3 1 Photographs of the cut holes and glued fans are provided in several consecutive figures 3 3 2 Modifying the scanner chassis 1 Connect an assembled scanner to a computer and scan a piece of paper to confirm that the hardware functions Scan in transparency mode t
5. black curves are provided as in Fig 3d of the main text Panel b The inclusion of a strain exhibiting a roller phenotype further demonstrates the broad applicability of our approach Fraction survivng 5 10 Age days of adulthood 15 20 25 25 C spe 9 hc88 I fer 15 b26 age 1 hx542 spe 9 hc88 I fer 15 b26 30 35 40 Figure 13 6 Lifespans of temperature sensitive self sterile mutants To avoid the potential for cohort confusion lifespan experiments with self fertilizing hermaphrodites on the same plate must prevent the accumulation of live progeny This is often achieved with 5 fluoro 2 deoxyuridine FUdR 17 To assess the performance of the Lifespan Machine without the use of FUdR we assayed two strains containing temperature dependent sterile mutations in the spe 9 and fer 15 genes Placing the eggs of such mutants at 25 C results in sterile populations Nature Methods doi 10 1038 nmeth 2475 56 0 757 0 507 Fraction surviving 0 257 glp 1 e2141 wildtype hsf 1 sy441 Visual Inspection 0 00 10 ot 15 20 25 30 Age days of adulthood 35 oO Fraction surviving 1 00 0 757 0 507 0 257 unc 64 e246 wildtype ced 1 e1735 visual Inspection 0 00 5 10 15 20 25 30 35 Age days of adulthood Figure 13 7 Visual controls of additional mutant survival curves Fig 3 in the
6. c and d are controls for a case in which the null hypothesis is true in this case the samples being compared are drawn from the same wildtype population at 25 C Note that in both cases the abscissa is linear in scale but tick marks were placed at each size sampled As for panel a the ordinate of panel c is logarithmic See text for details In one case we compare two samples one drawn from the wildtype population and one from the daf 16 population presented in Fig 4b of the main text both at 35 C In that case the null hy pothesis is false by construction In another case we compare two samples both drawn from the same wildtype population of 3578 individuals 25 C presented in Fig 2b of the main text illus trating the situation in which the null hypothesis is true The red line indicates 0 01 significance 99 confidence Each dot above the red line represents a test correctly not rejecting the true null hypothesis whereas each dot below the red line corresponds to a test that turned out signifi cant thus erroneously rejecting the null hypothesis type I error or false positive In Fig 9 1d we plotted the fraction of tests as a function of sample size that reject the true null hypothesis As expected this fraction is very low even at small sample sizes and approaches zero for large samples In contrast Fig 9 1a reports the case in which the null hypothesis is wrong as high resolution data shows daf 16 mutant pop
7. 6 8 10 12 14 16 18 20 22 Age days of adulthood Age days of adulthood Figure 12 1 Absence of a scanner effect To test whether the scanner environment affects the lifespan statistics of worms we manually scored plates residing on scanners thus exposed to the same environment as plates processed by the Lifespan Machine Panel a compares a set of such manually acquired survival curves with curves assembled from the traditional manual scoring of plates stacked in a box inside the same incubator Both curves agree Panel b compares the two manual approaches plates inside a scanner or in a box with the Lifespan Machine The inset of panel b shows the temperature differences between animals on individual scanners monitored by various techniques inside the incubator set at 25 C see also Fig 1 2 All plates were seeded with HT115 DE3 since the data were acquired in conjunction with Fig 13 3 We sought to determine what effects the scanner environment light vibrations etc might have on nematode lifespan For example a stimulation of worm movement in response to light has been noted using standard microscopy techniques 11 Mathew et al 12 suggest that the light produced by flatbed scanners may elicit a similar response We transferred several thousand wildtype ani mals onto plates stored in two environments inside operating flatbed scanners and inside a box sitting inside a neighboring temperature matched incubator For
8. C eaaa aE EAT leas cadeaven teaicadeaieeaneicadeaieatanines 54 13 9 daf 16 RNAV M 25 C renerien EEEE EEE see EAR E CRAEN EERTE 55 13 4 Automated lifespan assay of a severely uncoordinated mutant 0e eee 55 13 5 Survival curves of unc 50 e306 and unc 4 e120 rol 6 su1006 at 25 C 65 56 13 6 Lifespans of of temperature sensitive self sterile mutants cee eee eeeeeees 56 13 7 Visual controls of additional mutant survival CUIVES ccc esse cece eeeeeeeneeees 57 13 8 Error of LM operation at high temperature sccesccescesevsenveeeecesssnnnenes 57 Nature Methods doi 10 1038 nmeth 2475 3 Supplementary Note 1 Lifespan Machine hardware 1 1 A erii caei aaa E Kd OEE TORE EE AE REE E 4 1 2 Temperature control during image acquisition 2 eee e eee eee ee 5 1 1 Image acquisition Our reference installation consists of 50 scanners allocated to 5 incubators A single scanner is sufficient to obtain controlled survival curves of high accuracy while a fleet of scanners enables high throughput In Note 16 we provide a description of use cases for varying numbers of scan ners Instructions for the modification of a commercially available scanner model Epson v700 are provided as a separate supplementary document At www lifespanmachine org we have set up pointers to a public repository for the software components of the Lifespan Machine to facilitate user driven develo
9. Epson com for 540 In the past we have also used the dis continued Epson 4990 Photo scanner Other scanner models that possess appropriate tran sillumination capabilities are likely to be adaptable 70x 80x80x25mm DC fans This is a standardized computer chassis component and as such is widely available from Nature Methods doi 10 1038 nmeth 2475 13 many suppliers For example the Thermaltake DuraMax 8 AFoo58 80mm Case Fan avail able from NewEgg com for 8 20 10x 120x120x25mm DC fans This is a standardized computer chassis component and as such is widely available from many suppliers For example the Thermaltake A2492 120mm Case Fan is available from NewEgg com for 9 1x A 300 Watt source of 12V DC Power We use standard Power Supply Units PSU which are mass produced components of PC computer chasses and as such are widely available at low prices from many suppliers For example the CORSAIR Builder Series CX430 V2 CMPSU 430CXV2 430W is available from NewEgg com for 45 1x Temperature controlled environment for equipment This might be an incubator or a temperature controlled room For example the Thermo Forma Environmental Chamber is a multi use incubator which many labs already possess This incubator in 2010 cost approximately 13 000 new and as such represents more than half the price of an incubator based Lifespan Machine installation 50x Machine Screws For example Stainless Steel Machine Screw and Hex Nuts 11 32
10. High Strength EPDM Rubber Sheet 40A Black available from McMaster Carr at 22 per 3 feet A reference design appropriate for laser cutting is provided 1x Round Hole Arch Punch 2 Diameter Optional tool for construction If a laser cut design is not available rubber mats can be manufactured by hand using this punch Available from McMaster Carr for 70 160x Small Petri Dishes Consumed by each 5600 worm experiment BD Falcon Petri Dishes 50x9mm Tight Fit Lid Sterile BD Biosciences Available from mul tiple suppliers 215 per box of 500 40x Butt splices or wire nuts Optional To avoid soldering joints butt splices or wire nuts can be used to splice fan wiring Available from McMaster Carr in large packs for a few dollars Yes these have a silly name 1x Mini screw driver set for scanner disassembly Tool for Construction Already present in most labs For example a 5 Piece Phillips Screwdriver Set available from McMaster Carr for 22 18 1x Spiral saw for cutting holes in scanner chassis Tool for construction Other types of cutting devices may work well but a spiral saw cuts quickly and will impress your friends For example Rotozip Spiral Saw Electric 5 5 Amp 15000 30000 rpm 9 1 4 L available from McMaster Carr for 106 25 1x Hot Melt Glue Gun Tool for construction Standard Duty Hot Melt Glue Gun available from McMaster Carr for 23 Packs of Hot melt glue sticks available from McMaster Carr for 10
11. Intensity RI e RI Variance e RI Skew e RI Roughness Entropy e RI Roughness e RI Max e RI Average Along SP e RI Average in Nbrhd e RI darkest 20 of pixels e RI darkest 20 area e RI Dist from Nbrhd e RI Variance Along Spine e RI of Region e RI Normalized Average e RI Normalized Max e RI Normalized Avg Along SP e Edge Area e Edge Area Object Area e Intensity Profile at Edge e Intensity Profile at Center e Intensity Profile Max e Intensity Profile Variance Because no single feature is sufficient to accurately discriminate between worm and non worm ob jects information from multiple features must be integrated To assess whether accurate discrim ination is feasible we hand annotated 4 000 images as depicting either a worm or a non worm object We divided this set into two parts a training set 1489 worms 945 non worms and a test set 992 worms 631 non worms Using the 1ibsvm library we built an SVM model of the training data and evaluated its performance on the test set The confusion matrix is shown below demon strating a discrimination accuracy of 97 7 The Worm Browser Note 4 section 4 2 allows rapid inspection of all objects predicted by the machine to be worms Mistakes can therefore be anno tated producing a final false positive rate of nearly zero SVM prediction worm non worm by hand worm 991 1 non worm 36 595 To visualize the feasibility of discrimination we computed the eigenvectors of the feature correla t
12. Q O O to 100 Gompertz i 50 Gompertz 100 Weibull 50 Weibull 5 10 15 20 5 10 15 20 Predicted Age Days Predicted Age Days C g 4 d TA g am 2 B 5 lt N D o 2 50 Gompertz 50 Gompertz 50 Weibull 20 Weibull o 4 50 Log Logistic 50 Log Logistic 100 Generalized F 100 Generalized F 5 10 15 20 5 10 13 a Predicted Age Days Age Days Figure 11 2 Wildtype high resolution quantile quantile plots at 25 C The quantile quantile plots see section 11 1 visualize the extent of dis agreement between data and a variety of models a The parameters of Gompertz and Weibull hazards were determined by MLE using the full range of data except for the 1 crops at the tails Clearly the agreement is poor since the haz ard deceleration present in the data is at odds with functions that cannot decelerate b Model parameters were determined using data only up to the median lifespan This panel is a striking indication that early mortality in wildtype C elegans follows a Weibull more closely than a Gom pertz c Two further models are considered a Log logistic whose parameters were determined using data up to the median and a generalized F distribution 28 that takes into account the whole range of data The Log logistic coincides with the Weibull distribution at early times but evidently overestimates the extent of hazard deceleration The generalized F is a four parameter umbr
13. Technique for High Throughput Phe notypic Analysis of Caenorhabditis elegans PLoS ONE 7 e33483 13 Larsen PL Albert PS Riddle DL 1995 Genes that regulate both development and longevity in Caenorhabditis elegans Genetics 139 1567 1583 14 Morley JF Morimoto RI 2004 Regulation of longevity in Caenorhabditis elegans by heat shock factor and molecular chaperones Mol Biol Cell 15 657 664 15 Maduro MF Gordon M Jacobs R Pilgrim DB 2000 The UNC 119 family of neural proteins is functionally conserved between humans Drosophila and C elegans Journal of Neuroge netics 13 191 212 16 Gems D Riddle DL 2000 Genetic behavioral and environmental determinants of male longevity in Caenorhabditis elegans Genetics 154 1597 1610 17 Hosono R 1978 Sterilization and growth inhibition of Caenorhabditis elegans by 5 fluorodeoxyuridine Experimental Gerontology 13 369 374 18 Reinsch CH 1967 Smoothing by Spline Functions Numerische Mathematik 10 177 19 Herndon LA Schmeissner PJ Dudaronek JM Brown PA Listner KM et al 2002 Stochastic and genetic factors influence tissue specific decline in ageing C elegans Nature 419 808 814 20 Vaupel JW Carey JR Christensen K Johnson TE Yashin AI et al 1998 Biodemographic trajectories of longevity Science 280 855 860 21 Johnson TE Wu D Tedesco P Dames S Vaupel JW 2001 Age specific demographic profiles of longevity mutants in Caenorhabditis
14. a sufficient level of confidence 1 a where a is the rate of false positives type I error but also with a sufficiently high probability 1 b where b is the rate of false negatives or type II errors a is called the statistical significance typically chosen as 0 05 and 1 bis known as the power typically chosen to be 0 8 or 0 9 We used our actual lifespan data to reason about the sample size required to reach a desired power 4 ogee Q 4 e o o a lijis b d oe ae o t t i H Qa foo e e of o Oo 2 aS o a e gt E e g i z lt nu x7 o ge 7 9 o 2 gor 5 al eC c Oo g Fi o 8 oj T mT Tr TT T T T T T T T T T T T T T u Q mM T T T T T T T T T T T T T 2 20 60 100 140 180 220 260 300 340 2 20 60 100 140 180 220 260 300 340 sample size sample size C J 7 d x10 4 vis S 04 z ae Ona a Gare Lio 2 ba ao os e Hee 6 as oe i e bd o et M a las E 2 aane ae so P gani 8 Q Ng gt oO Ms mee s es 7 PT a Tee peed te ta Z Qa o7 gn te 7 Bo eo e O fos i a w 2 oe ee a x a s e i S Lo 7 a a Ta 2 2 4 e e e 8 S J ad n 2 e 5 e 5 5 pt e e e a S i 5 e e e e 8 1 k te i METET AE e e ERA A EEEE ee ne aL a S S S S S S S S S S S 100 200 300 400 500 2 100 200 300 400 500 sample size sample size Figure 9 1 Statistical power The top pane
15. access the bottom port 3 Identify the small metal clip that attaches the scanner bar to the drive belt The drive belt is the long rubber belt running the length of the scanner chassis Take a look at this clip and remember how it is attached Pop it off by pushing down on the exposed lever Locate the nut at the front of the scanner that holds the support bar in place The support bar is the long metal pole that running the length of the scanner chassis Remove the nut Notice that the bar is still held securely by a friction connection at the back of the scanner 4 Nature Methods doi 10 1038 nmeth 2475 21 5 When handling the support bar wear gloves to avoid exposing it to skin oils Rotate the support bar in place to release the friction connection Sometimes pliers may be necessary Only use pliers on the rectangular end of the bar never on the cylindrical part The scanner bar and support bar should now be loose 6 Remove the support bar from the scanner bar and put it in a safe clean place 7 Flip the scanner bar so that rests upside down inside the chassis Be careful not to catch any wires or to crease the ribbon cable 8 You should see several screws on the bottom of the scanner bar securing the bottom lid Remove them and remove the black bottom lid 9 You should now be able to see the fixed lens as shown in Figure 3 5 It is black often with a white dot There is a second lens used for IR measuremen
16. bit pixel intensity bits else ad i 8 bit pixel intensity 30 0 07 0 04 0 007 25 0 06 0 05 0 006 2 2 2 0 00 2 5 2 2 005 z 0 005 5 S aa 3 0 05 0 004 2 15 ae 2 0 0 Z a 5 0 03 5 0 05 5 0 003 3 i 0 02 3 0 002 a a a 0 00 a 5 0 01 0 06 0 001 0 T T pee 0 00 0 00 0 000 01 02 03 04 05 06 OF 08 0 10 20 30 40 50 60 70 010 30 5060 80 100 120 140 0 1000 2000 ratio of perimiter length and medial average median filtered intensity absolute object intensity at eroded Variance In absolute object intensity 4 f lt Perr i aahi pi around perimiter after repeated erosions axis length unitless along perimeter 8 bit pixel intensity animal center 8 bit pixel intensity squared 8 bit pixel intensity Figure 2 2 Feature palette for worm recognition The panels show the frequency distribution of the values for each feature in the case of worms blue and non worms red Because feature value distributions of worm and non worm objects overlap in all cases no single feature is sufficient for accurate discrimination 1 pixel width corresponds to 8m Nature Methods doi 10 1038 nmeth 2475 9 non worms Figure 2 3 Worm and non worm objects The images inside the red rectangle exemplify worm objects the images on the left outside the rectangle illustrate various non worm objects ranging from fibers to dust to lot numbers stamped on plates The two classes of objects are well separated by the Support Vector Machine Fig 2 4 based on
17. elegans show segmental effects J Gerontol A Biol Sci Med Sci 56 B331 339 22 Baeriswyl S Diard M Mosser T Leroy M Maniere X et al 2009 Modulation of aging pro files in isogenic populations of Caenorhabditis elegans by bacteria causing different extrinsic mortality rates Biogerontology 11 53 65 eth 2475 63 Natira Matt Nature Mett 23 Curtsinger JW Fukui HH Townsend DR Vaupel JW 1992 Demography of genotypes fail ure of the limited life span paradigm in Drosophila melanogaster Science 258 461 463 24 Carey JR Liedo P Vaupel JW 1995 Mortality dynamics of density in the Mediterranean fruit fly Exp Gerontol 30 605 629 25 Weitz J Fraser H 2001 Explaining mortality rate plateaus Proc Natl Acad Sci USA 98 15383 15386 26 Steinsaltz D Evans SN 2004 Markov mortality models implications of quasistationarity and varying initial distributions Theoretical Population Biology 65 319 337 27 Steinsaltz D Evans SN 2007 Quasistationary distributions for one dimensional diffusions with killing Transactions of the American Mathematical Society 359 1285 1324 28 Cox C 2008 The generalized F distribution an umbrella for parametric survival analysis Statist Med 27 4301 4312 ods doi 10 1038 nmeth 2475 4
18. from a can or from a hose hooked up to house air blow away all the plastic chips produced by the cuts 5 Insert machine screws into each hole Si S5 The head of the screw should be inside the TPU Secure the screw by tightening a nut on the reverse side the outside top of the TPU 6 Grab the scanner TPU bottom and snap it back into place Screw in all screws 7 Cut five 16 5 inch segments of nylon twine Put the TPU of a scanner on a flat surface with the machine screws facing upward Thread one end of a piece of thread through the two holes at the top of a 8cm fan Tie a knot at the appropriate piece of the thread such that the fan hangs snug at the TPU s side when hung from machine screw S1 Repeat three more times hanging fans from machine screws S2 S4 Remember to orient the fans correctly Fans F4 and F5 face inward fans F6 and F7 face outward 8 In the same way tie a 12cm fan F8 to machine screw S5 It should face inward 9 Finished ods doi 10 1038 nmeth 2475 7 Scanner Chassis Transparency Unit TPU 5 157 13 1cm 25 6 5 5 657 Additional Details Legend 14 4cm All holes in sthe TPU are 5 32 in diameter F1 F7 8 cm fans Side fans cuts on scanner body run from the F8 12 cm fan bottom of the scanner to 25 from the top H1 H7 Holes cut into scanner walls S1 S5 Holes drilled into scanner lid and fit with mounting screws Figure 3 1 Scanner modification diagram The locati
19. i we write Set w p1 p2 Pn i Nature Methods doi 10 1038 nmeth 2475 27 Natur e Methods doi 10 1038 nmeth 2475 This allows us to express that no location can belong to more than one time track for all i 4 j Set w N Set w We refer to the most recently added component youngest animal age t of the time track of worm i as w that is w p of w Throughout the process we maintain a set of time tracks W fw each track tentatively belonging to a single worm Corresponding to W there is a set of anchor points wi wi containing the most recently identified points of each track Finally we maintain a set of positions that have not been assigned to a time track retrospectively up to time t F p p P r gt tand Vi p Set w NY Given sets W and F we need to identify which positions P derived from scan t 1 extend backwards in time existing time tracks in W or initiate a new Hine This is accomplished by finding an assignment f wou F 14 P that minimizes the total Euclidean distance between positions p in wit U F and their i images under f X lior f pe minimum 2 where as many elements as possible must be assigned If either wi U F has fewer elements than P or vice versa some positions must be left unmatched i e they are matched to nil This combinatorial optimization problem is of a well known form that can be solv
20. method This however is a high standard for our equipment since the Lifespan Machine can monitor large populations at short intervals enabling it to detect very small effects from environmental variation We therefore pur sue a different strategy detecting and correcting such environmental variation computationally by using a linear regression model in which scanners are treated as values of a categorical vari able device This model makes a few assumptions that each scanner has the same effect on the lifespan of all individuals it measures and that it alters lifespan linearly either shifting it by a constant offset or scaling it by a constant factor The latter case is known as the Accelerated Failure Time AFT model 10 2 The device effect regression model Let t be the lifespan of the jth individual on scanner i i 1 s The effect of environment on lifespan can be modelled in two simple ways First additive effects might result in time shifts between survival curves such that t is 6 with 6 a constant offset specific to scanner i and Nature Methods doi 10 1038 nmeth 2475 45 cij a residual the corresponding lifespan absent the treatment by scanner i Second multi plicative effects might result in a proportional scaling of lifespans such that t becomes e 6 the AFT model For the sake of brevity we describe additive and proportional effects using the same additive model with the unde
21. over time as plates dry slightly or the scanner bar drifts The r criterion may allow two time tracks to drift close to each other and cross over or failure of the SVM model to detect worms may produce time tracks with excessive intermittency or very long gaps Spurious events may be introduced by fast moving worms that are coincidentally caught close to the anchor point of a worm time track To handle all of these cases paths are discarded or merged cleaned up according to a panel of criteria prior to being submitted to local movement analysis e A w is discarded if it is too short total duration of less than 6 hours too sparse i e it con tains points at only half of all its observation times or it drifts quickly the average distance between consecutive scans is greater than 5 pixels If two or more w are located very close to each other less than 50 pixels near the beginning 28 Natu of one w and near the end of the other it is likely that a single stationary animal was subjected to a one time displacement causing it to be incorrectly split into two w If the corresponding ends of the two w s do not not overlap for a very long time more than two hours and are not separated by a large gap in time more than twelve hours the two w s are merged The end result of this analysis is a set of time tracks as shown in Fig 1d of the main text Each straight track corresponds to a stationary worm with an associat
22. results of analysis Debugging image processing pipeline steps The Worm Browser acts as a platform for exper imenting with new image processing techniques This is often easier in a client than working with the image processing server directly E Inspect Worm Selection Clipboard Masks Data Annotation Testing Config Frame 7 of 712 Day 11 74 Date 03 49 11 07 2011 N2 25C 25C OP50 dawn_a 1 worm 14 1864 3830 Machine Changing Posture Human Changing Posture Figure 4 1 Worm Browser screen shot See also Supplementary Video 6 26 Supplementary Note 5 Movement analysis 5 1 Identification of stationary WOTMS ccc cece cece eee ee eee ee nees 27 5 2 Posture analysis of stationary WOrMS cece eee eee cent eee 29 5 1 Identification of stationary worms Once foreground objects have been classified into worms and non worms Figs 2 2 and 2 5 the task is to identify when individual worm objects have permanently stopped moving The Lifespan Machine distinguishes two kinds of motion i spontaneous locomotion across the agar surface and ii subtle posture changes that do not involve whole body displacement also referred to as local motion Accordingly the identification of death events occurs in two phases First we iden tify worms that have stopped locomotion Such worms are termed stationary Second images of stationary worms are further analyzed to identify cessation of spontaneous posture
23. stuck together Note 8 section 8 3 These cases we resolve through rapid visual inspection using the Worm Browser Note 4 section 4 2 For many users it may be permissible to instruct the survival curve assembly software to discard such pairs entirely as doing so does not appear to cause problems of mis estimation see Fig 8 5 presumably because such paired dying occurs randomly At high densities however we see an increasing number of animals dying in groups of four or more at which point even careful human inspection cannot disentangle the varoius death times Ultimately we prefer to minimize the potential scope of such problems by maintaining low animal densities We find that with 35 animals per plate the frequency of worms dying in close juxtaposition is very tolerable Lowering worm density can further reduce issues associated with worm aggregation if desired Deaths from causes unrelated to aging The current version of the Lifespan Machine cannot discern causes of death that are unrelated to aging such as bagging The risk of bagging and perhaps of non aging deaths in general decreases as egg production slows with age As a result bagging occurs almost exclusively during a nematode s reproductive window In our by hand data sets the frequency of such events is less than 5 and in some cases 0 agreeing with previous characterization 6 Bagging and vulval ruptures can be identified in the visual data validation step using t
24. than a Gompertz distribution up to times slightly beyond median lifespan when deceleration becomes noticeable 01 4 0 1 4 oO g 2e oO 0 01 0 01 S 0 01 01 50 Gompertz 50 Gompertz 0001 50 Weibull 9 001 50 Weibull T T T T l l l l 5 10 15 20 5 10 15 20 Age days Age days Figure 11 1 Wildtype high resolution hazard functions at 25 C In this figure we graph for the purpose of visual comparison our hazard rate data black dots and the Gompertz red and Weibull blue model hazard functions as determined by MLE from our lifespan data The black dots are produced by subdividing the duration of the experiment into equal intervals and report ing the average risk of death the number of deaths as a proportion of the population initially at risk within each interval The comparison with model hazard functions suggests that a power law Weibull describes early mortality better than an exponential Gompertz The left panel repro duces Fig 2d of the main text and is plotted using a log log scale in which the Weibull hazard is a straight line while the right panel uses a log lin scale in which the Gompertz hazard is a straight line The hazard fits are drawn solid over the data range used in the fits i e up to median lifespan and dotted elsewhere Nature Methods doi 10 1038 nmeth 2475 48 a g b S a a g g lt lt x mo mo 2 o 2 o or 5 gt n n a
25. the feature set shown in Fig 2 2 PCA 2 Figure 2 4 Support Vector Machine for worm recognition Black worms Red Non worms Each object of a hand annotated gold standard is represented as a dot in a 65 dimensional feature space The figure shows the projection of this point cloud onto the subspace spanned by the three principal components that explain the largest share of total variation 60 4 The previous table was based solely on wildtype The following table illustrates the performance of the SVM when trained on a mix of strains and then applied to each separately rather than trained on each specifically For this analysis we 1 lumped a number of genotypes into one big set 2 assigned randomly 3 5 of animals to the training set and the remainder to the test set 3 trained an SVM on the training set 4 tested it on the test set 5 calculated the confusion matrix for each 475 10 genotype FP false positives FN false negatives all daf 16 mu86 wildtype glp 1 age 1 hx546 FP 84 1550 5 4 23 207 11 1 21 562 3 7 32 453 7 1 8 328 2 4 FN 73 2426 3 0 11 336 3 3 19 777 2 4 31 883 3 5 12 430 2 8 All strains have a false negative rate between 2 4 and 3 5 The false positive rate is more varied but ultimately immaterial since we catch the non worm objects during visual inspection in the quality control phase with the Worm Browser Fig 2 5 provides examples of image segmentati
26. we do not attempt to determine the exact form of these changes Instead we simply determine the overall change A in pixel intensity inside the alignment frame between X and X This strategy also has the advantage of not relying on accurate segmentation allowing subtle head movements to be detected even in the common situation in which heads are not fully retained during segmentation Consecutive images of the same plate region may vary in overall intensity simply because of vari ations in the output of the scanner light source or stray light To compensate for differences in average brightness between subsequent images we use as the local movement score N N Y c X ef X D b Dh g 5t Se 3 i l i 1 where X denotes the mean value in image X and the summation runs over all N pixels in the alignment of images X oe X The above expression yields a time series of local movement scores A t 1 T 1 for a stationary worm see Fig 1e in the main text and Figs 7 1a 7 1b 7 3 Additional forms of scanner noise may cause A to deviate from zero even for dead worms Fig 1e in the main text and Fig 7 2b This necessitates a one time calibration step to define a thresh old value A max below which we consider an animal as non moving However dead worms might appear to sporadically move as live animals push them around We therefore define an animal as permanently non moving i e dead when it is obse
27. worm clusters 00 ccc es cee cesceeeneeevees 41 8 5 Determining total population size 0 cee ee eee cee ereere nn 42 8 1 Right censoring and interval censoring Individuals are lost from a lifespan experiment when supervening events preclude the observa tion of the age induced death of these individuals For example in both manual and automated lifespan experiments worms are occasionally found dehydrated on the plate wall These individ uals are lost from further observation but cannot simply be ignored during statistical analyses as doing so would involve consideration of a biased sample Kaplan and Meier proposed a nonpara metric method for estimating survival curves that takes into account the contribution of individuals to the population at risk up until the moment they are lost 7 We compile our survival curves ac cording to Kaplan Meier KM Many statistical packages like JMP require time to event data to assemble a survival curve 1 0 interval censoring at mean of interval 2 05 2 Z 5 40 6 i Turnbull estimator w 0 5 0 0 0 5 10 15 20 25 Age days of adulthood Figure 8 1 Interval midpoint vs Turnbull estimator See text for details The two curves are effectively indistinguishable if drawn on top of each other Our scans however happen at discrete times Death or loss events are therefore only known to occur within a time interval defined by successive scans Based on
28. 0 70 80 90 0 200 400 600 800 100012001400 1 0 1 2 3 2 3 4 average median filtered object intensity variance of median filtered object intensity skew of median filtered object intensity entropy of median filtered object intensity 8 bit pixel intensity square 8 bit pixel intensity 8 bit pixel intensity bits Tt 0 030 0 025 1 0 gt gt gt 09 gt 0 025 4 0 020 ic B 08 ic 5 S S 0 7 S 0 020 5 gt 0 015 gt gt 06 gt 05 0 015 5 3 3 3 5 3 S 0 010 g 04 0 0104 B 2 03 2 gt 0 005 02 9005 4 0 1 0 000 T T 0 0 0 000 rrr he a a a 0 100 200 0 1000 200 3000 2 4 0 1 2 3 4 5 6 O 20 40 60 80100 130 160 190 average absolute object intensity variance of absolute object intensity skew of absolute object intensity maximum absolute object intensity 8 bit pixel intensity square 8 bit pixel intensity 8 bit pixel intensity 8 bit pixel intensity 0 00045 3 5 0 035 0 012 0 0004 3 0 0 030 2 0 00035 a cos 0o10 2 0 0003 0 008 Y 0 00025 2 0 0 020 2 0 006 0 0002 15 0 015 3 0 00015 1 0 0 010 0 004 a 0 0001 0 002 0 00005 oS 9 003 0 i mrm 0 0 0 000 rrr 0 000 0 10000 20000 30000 1 2 3 4 5 130 100 80 60 40 20 O 20 40 60 100 200 300 400 500 600 700 800 A are P ai P difference between average absolute object intensity variance in absolute object intensity along medial axis entropy of absolute object intensity and average intensity of object neighborhood perimeter length pixel widths square 8
29. 2 Inspect the captured image at 100 zoom The quality of the current scanner focus can be Nature Methods doi 10 1038 nmeth 2475 22 scanner bar clip to drive belt scanner bar it i Figure 3 5 Inside the scanner chassis a A top down view of the scanner chassis with its lid removed b A close up of the clip that connects the scanner bar to the drive belt e The slot in the scanner bar through which the lens can be adjusted d The freed scanner bar can be overturned to reveal a bottom access panel e With the panel removed the bottom of the lens can be seen glued in place f The gum can be removed with a screw driver g The lens freed from gum evaluated through comparison to a collection of gold standard images available in Figure 3 6 13 If optimal focus has not been achieved remove the lid of the scanner chassis along with the Nature Methods doi 10 1038 nmeth 2475 23 entire TPU as one piece With proper organization of your workspace it should be possible to do this without detaching any cables 14 Look through the top access port of the scanner bar and identify the fixed lens With a small flat head screw driver nudge the lens slightly in the desired direction The lens should be ad justed in tiny nearly imperceptible increments If you think you ve made a large adjustment in all likelihood the scanner is now extremely out of focus In this case it will be a long job hunting your way
30. Device specific environmental effects on lifespan 0 cece eee eee 45 10 2 The device effect regression model cece cece cece eee e rrenen no 45 10 1 Device specific environmental effects on lifespan The natural units of aggregation in the Lifespan Machine are its units of confinement plate scanner incubator The procedures described in the previous sections apply to a single plate and yield a survival curve for that plate Each plate produces data with high temporal but low statistical resolution as a plate contains about 30 50 worms Better statistics can be produced by aggregating multiple plates containing animals of the same strain either directly with our software or using a statistical package like JMP 3 Building curves for individual plates provides a highly valuable sanity check as it affords an opportunity for statistical tests of equivalence and for recognizing when entire plates may have to be censored for example due to fungal invasion desiccation or fogging Data from individual plates are aggregated to generate survival curves at the level of in dividual scanners Before pooling scanner aggregated curves into an overall curve at the incubator level they may have to be registered to account for small device specific environmental differences especially local temperature C elegans lifespan is sensitive to environmental conditions and lifespan assays can detect the ef fects on survival that resu
31. Nature Methods doi 10 1038 nmeth 2475 Supplementary Information The C elegans Lifespan Machine Nicholas Stroustrup Bryne E Ulmschneider Zachary M Nash Isaac F Lopez Moyado Javier Apfeld and Walter Fontana Department of Systems Biology Harvard Medical School These authors are co last authors These authors are corresponding authors stroustr fas harvard edu javier_apfeld hms harvard edu walter hms harvard edu Table of Contents BE Oe R E E E E EE Supplementary Note 1 Lifespan Machine hardware ceeeeeecceeeeeeneeceeeeeeees LI Tmageacguisiion FR eea HR eS EA ia a ESRD ae e a a 1 2 Temperature control during image acquisition 0 00 e ee eee Supplementary Note 2 Image Processing ccccssesesvececcsescsseceacsessesvsencseseeens 2 1 Masking background subtraction and segmentation 2 2 Registration of consecutive scanner images 5 0 ee eee ene ees 23 Worm recognitlon oso oR OM HO Ee Re EN OO eck OO eS Supplementary Note 3 Asset ly ccix cawissieesesr cee asavernea doesn teresa aire radio ews at Tntroduehon ssa vd 6 iG be Gee ae ee ed bee Pee ee Gate eds 3a Complete parts bi i o Re ew Ee a e a 3 3 Recommended scanner modifications sso sooo e ee ee eee 3 3 1 Overview of the fan mounting procedure aasa 3 3 2 Modifying the scanner chassis saaa ence ee enne 3 3 3 Modifying the transparency unit a aa eaa 3 3 4 Accessin
32. a t D t in equation 4 The green curve 3 is m t and is obtained by subtracting the red curve from the total population size N at the beginning of the experiment Determination of N is discussed in section 8 5 The red curve shows change at long time scales presumably reflecting the overall clustering behavior of worms in the population Early in the experiment obs t declines because a fraction of worms cluster Worms then leave clusters perhaps because food becomes locally scarce As aging animals become progressively paralyzed a subset co locate into juxtaposed pairs or tuples Close to death these tuples are misidentified by our image processing as single worms see section 8 3 The green curve m t mirrors obs t by construction The orange curve 4 corresponds to the orange curve in the schematic of Fig 8 2 a An example of missing worm detection is shown for a single plate This plate was chosen because an unusually large fraction of animals was lost b For each of 160 plates the equation 5 is applied to individual plates and the results are aggregated c The same data in b is replotted with deaths separated into singletons two worm clusters and three worm clusters In the other scenario starting from t 0 each down step recapture event at time t is randomly matched with an available up step event gone missing that occurred before t whereupon that up step is no longer available for future recaptures As in the
33. action and segmentation 00 7 2 2 Registration of consecutive scanner images 2 ee eee eee cece eee 7 SS Worm recognition secc ceccccisisissnerenenvstei seckisi it siedennehes ean 8 2 1 Masking background subtraction and segmentation In images captured by a scanner only areas corresponding to the agar surface of each plate need to be analyzed Masking is the process of separating multiple plates and removing their edges The Worm Browser Note 4 section 4 2 and Supplementary Video 6 generates a low resolution composite of each image captured The user can then draw on top of this image to produce an overlay the mask that specifies the location of each plate This drawing can be done quickly in programs such as Photoshop or GIMP 1 This task is a candidate for future automation The image processing server uses the resulting mask to extract individual plate images and stores them in a stereotyped folder hierarchy associated with each experiment Prior to analysis all images X t 1 T undergo background subtraction Each pixel ol X is reduced by the median value in a surrounding square of 25 x 25 pixels 198 um x 198 um This median filtering removes variation at spatial scales larger than roughly twice the average width of an adult worm 198 um suppressing background effects including shadows smooth bacterial lawns and uneven agar shading Median filtering sets the stage for image s
34. alf day intervals the first between 0 75 and 0 25 days before death the second within 0 25 days of death and the third between 1 75 and 2 25 days after death The cumulative distribution of each group is shown for each genotype In all cases the middle interval 0 25 days surrounding death showed the smallest median and mean size indicating that animals reach their smallest area just as they cease to move f At 35 C animal area was considered at three single time points two hours before apparent death at apparent death and three hours after apparent death Animals appear to begin increasing in size before their last movement and the size continues to increase afterwards features of worms nearing death This led us to discover that most animals exhibit a stereotyped morphological change at or near the time of their final posture change They first shrink by 10 or more and subsequently expand by more than that amount The phenotype is established in Fig 7 1 its coincidence with the cessation of all spontaneous motion the death criterion of our method is shown in Fig 7 2 and Fig 7 3 indicates its generality a 1 254 b 0 3 a J wildtype wildtype F 1 24 daf 16 m86 daf 16 m86 ee 115 age 1 hx546 2 op age 1 hx546 ov q i Sg 114 8 g J z EG q o o s 1 057 E Zc J S 01 2 1 O J 0 954 o E a 0 21012345 67 8910 Time relative to death d Time rela
35. along the wire a voltage drop of less than 3 along the wire is widely considered safe If the power source is 10 feet away from the scanner this suggests that an AWG gauge no larger than 20 should be used Incorrect wiring can generate excess heat and represents a fire hazard 3 3 4 Accessing and preparing the fixed scanner lens The Epson Perfection v700 contains a single fixed lens whose position sets the focal plane of the device Worms on agar plates are out of focus on stock scanners because they are raised above the standard focal plane To optimize image quality we move the lens position very slightly 1 The scanner lens is located inside the moving scanner bar situated inside the bottom scanner chassis Remove the lid of the scanner chassis as described in Recommended Mounting Procedure Step 3 and identify the scanner bar Refer to Figure 3 5 2 There is a removable plastic flap in the top of the bar Remove it with a screwdriver Put it aside in a safe place If you look into this slot you should see a few inches down a black tube This is the top of the lens Unfortunately scanners ship with this lens glued in place by a piece of hard resin We need to remove this resin before we can adjusting the lens position Removing the resin in turn requires we access the lens directly through a port on the bottom of the scanner bar This in turn requires we disconnect the scanner bar from the scanner chassis to make it possible to
36. animals stored inside flatbed scanners we observed a portion by hand and the remainder using the Lifespan Machine Fig 12 1 We first discuss animals scored by hand We observed that animals located inside scanners lived 18 hours shorter than those housed outside scanners in a separate incubator To place this difference in the context of temperature we identified a significant correlation between measured scanner Nature Methods doi 10 1038 nmeth 2475 52 surface temperature and mean lifespan both through linear regression on plate means R 0 65 p 0 027 and in Cox regression with temperature as a continuous covariate p 0 002 Using our linear regression as a predictor for lifespan we found that the difference between scanner and box groups was reduced to 15 6 hours or 4 4 of lifespan representing our best estimate of the effect of the scanner micro environment on nematode lifespan We suspect that effects of this magnitude despite our attempts at accounting for the effect of temperature might nevertheless result entirely from limitations in our ability to accurately measure temperature as it varies between scanner and box environments We then considered animals housed on scanners scored either by hand or the Lifespan Machine We found that animals observed by hand appeared to live 6 2 hours shorter than those observed by machine or 1 8 of total lifespan This close correspondence between our method and the manual procedure sug
37. as missing random missing duration Figure 8 5 Censoring strategies for terminal worm clusters Curve 1 orange Our best practice method in which multiple worm clusters are manually annotated using the Worm Browser The image processing pipeline determines an interval bounded by the nucleation time of a cluster and its death time The cluster is typically misidentified as a single worm One individual is assigned the cluster death time all others are assigned the interval midpoint This is the curve shown in Fig 2b of the main text Curve 2 blue No manual annotations are considered and only a single death time is de facto recorded for all animals in a cluster This death occurs at the death time of the last animal to cease movement within the cluster Curve 3 red Using manual annotations of multi worm clusters all members of the cluster are assigned the same death time which is the cluster death time Curve 4 purple Using manual annotations of multi worm clusters worms entering into clusters are declared lost at the time of cluster nucleation and right censored Curve 5 green Using manual annotations of multi worm clusters such clusters are completely eliminated from the death record As a consequence all animals that disappeared into terminal clusters are handled according to the missing worm procedure which also handles transient clusters as described in section 8 2 Curve 5 is obtained when the procedure operates accordi
38. ases Such deceleration has been described anecdotally before in C elegans 20 22 and flies 23 24 and humans wasps and automobiles 20 although never in stress regimes e g high temperature nor systematically for a large number of mutants The parameters of model hazard functions Gompertz Weibull and Log logistic were obtained by maximum likelihood estimation MLE using the lifespan data death times of the wildtype population shown in Fig 2 of the main text Gompertz and Weibull hazards do not decelerate and MLE estimates of the parameters of these distributions will yield a poor fit The origins of the decel eration phenomenon deserve an empirical study in their own right which we leave to future work but see 25 27 for computational and theoretical approaches In the present data analysis we focus on the functional form of early mortality which is a natural and easier first task in organizing our volume of data To this end and unless otherwise noted when obtaining MLE estimates of model parameters we only use lifespan data up to the median right censoring the remainder and thus eliminating the late phase in which hazard decelerates This is what we mean in legends or captions when referring to for example 50 Gompertz The phrase 100 Weibull then means that the full range of lifespan data was used except for cropping 1 at the noisy tails The max imum likelihood estimation of model parameters was perfo
39. at went missing most recently This scenario matches each downslope of m t with the most recent upslope see Fig 8 2a Each unmatched upslope step then provides the number of animals that were lost to the experiment at that time and must be right censored in the final assembly of the survival curve This strategy can be formalized as a retrospective right to left propagation of the smallest value Nature Methods doi 10 1038 nmeth 2475 38 of m t z ti min m t z tig1 t n 1 1 with z tn m tn 5 The series of z t is a step function that increases monotonically with increasing i A step at time t triggers a censoring annotation in the lifespan log asserting that A t z ti 1 z t individuals are permanently not accounted for and should thus be right censored at time t b 3500 3000 2500 2000 1500 1000 500 0 Number of animals Number of animals Time d O 3500 3000 2500 2000 1500 1000 500 1 total observed 2 cumulative deaths 3 missing 4 permanently lost 5 cumulative deaths of worms outside clusters 6 cumulative observed deaths of worms inside clusters 7 cumulative estimated deaths of worms inside clusters Number of animals 0 5 10 15 20 25 30 Time d Figure 8 3 Going missing and getting lost data In all panels the red curve 1 shows the total number of observed animals obs t at time t i e obs t
40. back One strategy is to scratch notches into the paint of fixed lens housing These scratches can then be used as a reference to mark your position 15 Reattach the chassis lid and TPU making sure that the lid is seated flush 16 Repeat steps 11 15 until a scanner is in focus It is the authors experience for half the scan ners they have adjusted optimal focus is achieved within two or three adjustments In this case focus is achieved in less than five minutes For the other half of scanners optimal focus is more elusive requiring up to half an hour of iterations Be patient this process only needs to be done once in the lifespan of the scanner A W F ee lt q a wa t Y Pr nd Figure 3 6 Reference images demonstrating sufficient focus These images were collected across several scanners and represent the end point of scanner focal plane optimization for the devices In the authors experience attempts at acquiring images sharper than those presented are generally counterproductive Nature Methods doi 10 1038 nmeth 2475 24 Supplementary Note 4 Software toolset 4 1 Software components 22 266 bec cdaacvas tea viae cess ecu es dee died eaeeaceead 25 4 2 The Worm Browser GUI client 0 ccc cc eee enn een ne nees 25 4 1 1 2 3 4 Software components Statistical software To assemble survival curves we used JMP 3 a commercially available statistical package and R 4 an excel
41. cance ceases to correlate with biological significance Instead the machine is meant to enable investigations into the shape 61 doi 10 1038 nmeth 2475 and functional form of the hazard rate which often requires the evaluation of subtle effects that occur late in life at which time effective population numbers have dwindled as well as early in life where only a fraction of deaths in each population occur Installations of five to ten scanners allow the experimenter to perform multiple replicates of these types of analyses Installations of five to ten scanners also permit experimenters to screen through libraries of mutants and RNAi constructs When population sizes of 100 animals are chosen three strains can be reasonably tested per scanner allowing thirty strains to be measured in a single incubator Consider that a wildtype population should be run on every scanner along with the target populations This will necessarily curtail real estate for mutants that can be co located on the same scanner One should also consider that this use captures entire survival curves for each population increasing the quality but necessarily decreasing the throughput compared to the single time point assays currently used in many high throughput screening contexts Stress resistance assays such as the thermotolerance experiments shown in Fig 4 of the main text also benefit from an incubator full of scanners Being able to spread animals across mul tip
42. cccccccceeceecceecencs 34 7 3 Generality of the volume dip ensensensensenserserserseeserseeseeseessessessessessesse 35 8 1 Interval midpoint vs Turnbull estimator sssssssrssserssessseressessesroserrssees 36 8 2 Going missing and getting lost schematic ccccccesscscevecssevsseeasensvereaevesesnes 38 8 3 Going missing and getting lost data sssesussrssssrsssesseresssosseresesrsssee 39 8 4 Interval censoring for worm clusters Schematic cccccceceeceeeeeeeeeeeeeeecs 40 8 5 Censoring strategies for terminal worm clusters sssssssssssssssssssssssssssssee 41 i Statistical RE so eh sce cent irirna ind tra Tin A EEOAE RARA IE TREERNE 43 11 1 Wildtype high resolution hazard functions at 25 C ssssssssssssesssreseereseeessso 48 11 2 Wildtype high resolution quantile quantile plots at 25 C eeceeecee eee eee eeees 49 11 3 Wildivpe hazard at 25 C scored manually oo dees detesdeeiesedoneen dosti Sedawkvden 50 it 4 Wildtype hazard at 20 C on PET is ie sscoweaxdawrasdaweireeseeserneasestiakersedsenuuned 50 11 5 Population size and hazard parameter estimates cce cee eeeeee eee eeeeeeeeeees 51 12 1 Absence of a scammer effect 6 6 5 c cc ccc ses sdeeessdaeueesns vasdecvesededesdaccessdesse 52 13 1 Wildtype at 25 device corrected oi iiisavinadacacasadinadesiiaeea saa dideriees eae cieeainia ds 54 Nature Methods doi 10 1038 nmeth 2475 13 2 RNAI ice ans Ab 20
43. changes The retrospectively identified cessation of all motion is then recorded as a death event A trained Support Vector Machine SVM see Fig 2 2 and Fig 1d of the main text identifies the position x y of worm objects in each image X Let PO x y object at position x y in X is an SVM identified worm denote the set of these positions Early in a lifespan experiment when most worms move around rapidly the scanning rate of two images per two hour period does not allow determination of which worm in P corresponds to which worm in PY To identify death events however we do not require such early stage detail It suffices that worms become trivially trackable once they have ceased locomotion These worms are then monitored for subtle head and tail displacements as described in Note 5 section 5 2 Identifying the onset of a worm s stationary phase is a retrospective analysis It results in a straight time track Fig 1d of the main text with an associated series of images that is then subject to posture analysis While this is straightforward in principle the intermittent locomotion of worms their spatial interactions and the noisiness of image capture Fig 2 1 and worm recognition create challenges We construct time tracks of stationary worms in a stepwise fashion and clean them up prior to posture analysis The time track of a stationary worm w is defined as a sequence of locations wi pipe Pn wi
44. e 13 2 RNAi Lifespans at 20 C To test the performance of the Lifespan Machine at 20 C we placed a subset of animals prepared for Fig 3a in the main text on scanners in an incubator cal ibrated at 20 C The by hand control for wildtype on daf 2 RNAi was truncated as experimenters left for winter holidays the Lifespan Machine continued unaffected Nature Methods doi 10 1038 nmeth 2475 54 o a 1 0 4 2 2 Z 0 6 2 0 6 3 3 D 4 D 6 0 44 5 04 g g i 0 24 wildtype empty vector i 0 2 wildtype empty vector wildtype daf 16 RNAi wildtype daf 16 RNAi 0 0 i o r a al Ai 0 0 T1111 TT 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 Age days of adulthood Age days of adulthood Figure 13 3 daf 16 RNAi at 25 C RNAi of daf 16 has been reported to shorten lifespan 13 14 Here we note that the RNAi knockdown is noticeably weaker than the mutant shown in Fig 3c of the main text Panel a survival curves by plate Panel b aggregated survival curves 25 C 0 8 7 unc 119 ed3 automated 2 064 unc 119 ed3 visual inspection wildtype automated 0 4 wildtype visual inspection 0 27 0 0 0 5 10 15 20 Age days of adulthood Figure 13 4 Automated lifespan assay of a severely uncoordinated mutant We used unc 119 ed3 animals to assess the performance of the Lifespan Machine on worms with unusual move ment phe
45. e midpoint estimate as our standard practice All our data however are optionally supplied in an interval format suitable for Turnbull estimation 8 2 Tracking population size and right censoring strategies As in the manual lifespan procedure worms whose lifespans are acquired automatically can be lost before their death is observed Such animals must be right censored when their observation ends A classic loss occurs when a worm crawls up the vertical sidewalls of a plate In the Lifespan Machine this causes a decrease in the total count of worms observed on that plate Other events that require censoring are of biological nature such as the rupture of a worm that fails to release fertilized eggs These events are currently not detected automatically and are counted as deaths unless explicitly censored by visual inspection with the Worm Browser during the quality control phase that follows data acquisition and image processing One source of censoring events is specific to the automated method Individuals can aggregate into clusters or clumps of two or more animals where they are usually either misidentified during image processing as a single worm or ignored entirely Fig 2 5 Yet for living animals the re sultant loss is often only temporary as these individuals become again observable once they leave aclump Dead animals inside clumps as well as animals that leave the field of observation will remain missing permanently We thu
46. e t and D t the number of deaths that have occurred up until and including t D t to d i Further let a t be the number of worms alive at t and N the total number of worms present initially At any given time a worm is either alive or it has died sometime in the past including now or it is missing a t D t m t N 4 with m t denoting the number of worms that are unaccounted for missing at time t Automated analysis of the image stack acquired during the lifespan experiment yields a t and D t providing us with m t N a t D t Worms mostly go missing in three ways i A scan does not image Nature Methods doi 10 1038 nmeth 2475 37 a worm properly For example it may have crawled near or up the vertical plate wall Gi A worm is misidentified as a non worm by the classifier iii A worm has disappeared into a clump of worms where it goes undetected In all of these cases a worm may return from the missing and be accounted for at a later time Therefore a plot of m t versus t typically results in a non monotonic step function a schematic of which is depicted in Fig 8 2 green curve and an actual example is shown in Fig 8 3b a b t ts tg ti ts te tg Time Time Figure 8 2 Going missing and getting lost schematic In both panels the green curve repre sents a schematic m t of equation 4 the number of animals detected missing over time Each up step in the direction of increasing ti
47. e that individual j is exposed to the environmental impact of a scanner is therefore not its lifespan t but t t or log t t where t is the age at which animals were placed on scanners The Cox model constitutes an additional possibility to account for a scanner effects In Cox re gression the scanner environment is assumed to have an additive effect on the log of the hazard function We refrain from running a Cox model which requires fitting a baseline hazard for the purpose of the present paper Any of the AFT class of regression models can be applied to censored data using the Buckley James estimator based on modified least squares estimating equations 10 This method replaces each censored observation by the conditional expectation given the observed data and the covariates Because the conditional expectation itself depends on the estimating equations must be solved iteratively Nature Methods doi 10 1038 nmeth 2475 46 Supplementary Note 11 Hazard 11 1 Hazard estimation and rendering cece cece cece eee cee esac 47 11 2 Hazardofwildiyp bc ctvea nena ease ec heheas never nie sand cana heenexeen 48 11 3 Hazard and sample size oi cen ecaca cos bebeas es nds Oeweiceeed ei enewsenras 51 11 1 Hazard estimation and rendering The hazard rate is a fundamental statistical observable in aging It is the instantaneous rate at which individuals that survived to time age t die in the next instant The role
48. ed 5 The new set Ft recall that we proceed from later towards earlier times is formed from all elements of F that were not matched to a p P plus all p P amp that have no match in wi UF Each element in F paired with an element in P initiates a new time track and thus identifies a new worm that is added to W in producing W A p P that pairs up with w extends the time track of worm w i e wi pw In effect p becomes the new wh Gaps occur when a freshly paired wh is several scans old The process begins with t T the last scan setting F PT w and consequently W 0 The mechanism advancing the scanning bar causes jitters of a few hundred microns that make it difficult to exactly map position x y in one image of a plate to the same position in a subsequent image Fig 2 1 Thus for two positions to correspond to each other we require them to be within distance r In anes the assignment f we therefore exclude penalize pairings between a position in wit JU F anda position in P whose Euclidean distance is larger than r This speeds up the optimization process and tunes it to detect stationarity since pairings between more distant positions are unlikely to refer to the same animal in our time lapse context Various sources of noise and measurement error may produce non optimal behavior of the loco motion analysis procedure For example w may drift perceptively
49. ed stack of cropped images 5 2 Posture analysis of stationary worms The input to local movement analysis is a series of cropped images X We t 1 7 ofa stationary worm i with t 1 indicating the image from the first scan and t T the image from the last scan For the sake of less clutter we shall omit the index 7 with the understanding that the following analysis pertains to a single time track All images have undergone background subtraction as described in Note 2 section 2 1 Further more a scanner does not produce perfectly aligned images in consecutive measurements of the same plate Fig 2 1 Each image X is therefore aligned computationally with the consecutive image X to remove machine jitter An image is a set of pixel intensities on an x y grid and represents a discrete sampling of the continuously varying intensities being measured Although pixels are located at integer offsets the scanner alignments errors resulting from the advancing scanner bar can vary continuously To account for such sub pixel errors in alignment we sub sample images to allow registration offsets of distances less than a single pixel A typical stationary worm series with posture changes is shown in Fig 1e of the main text The objective of local movement analysis is to detect these changes by comparing consecutive images of the same stationary worm Since animals may take a variety of postures and various mutants may present novel postures
50. egmentation the separation of worms from their agar and bacterial surface background using a global threshold Two global thresholds are applied to images one restrictive threshold that identifies only dark objects such as worms but tends to break worms into multiple objects and a second more permissive threshold that tends to pick up extra neous features of the bacterial lawn The two are compared to produce a final image segmentation See also Fig 2 5 2 2 Registration of consecutive scanner images The panels in Fig 2 1 show the magnitude of two types of jitter in the acquisition of images by our scanners and presumably consumer electronics scanners in general Displacements like these must be taken into account when processing images leading up to the movement analysis sections 5 1 and 5 2 a b C wo wo N L L D 4 o 1 1 i AY worm position pixels N wo L AY plate position pixels o 1 1 1 1 1 AY worm position pixels 1 A r r r r 0 0 5 1 1 5 2 2 5 0 0 5 1 1 5 2 25 2 1 0 1 2 Time d Time d AY plate position pixels Figure 2 1 Registration of consecutive scanner images a The mechanism advancing the scan ning bar causes offsets of a few pixels relative to world coordinates between subsequent images of the same plate b Slight distortions in plate images cause displacements of a few pixels relative to plate coordinates for stationary worms c T
51. ella that encompasses both Log logistic and Weibull distributions Not surprisingly it can be tailored to the data very well Panel d shows the survival curves implied by the fitted model parameters The black curve with the narrow 95 confidence interval represents the non parametric Kaplan Meier curve based on our data The generalized F sits snug underneath The main differences to the other models are chiefly in the tail Again the blue Weibull curve is seen to match the data well beyond the median Nature Methods doi 10 1038 nmeth 2475 49 The hazard analysis of the manually scored results also indicates a better fit to the Weibull than the Gompertz distributuion 5 4 PG oe J F oa o 0 5 4 soe T 2 T f Pi lt e se g 2 05 4 wee ge 5 a N wee vee 4 of fe 005 oF os o machine 0005 by hand T T T T 5 10 15 20 Age days of adulthood Figure 11 3 Wildtype hazard at 25 C scored manually This figure compares the empirical haz ard of the manually scored wildtype population Fig 2 in the main text with that of the LM scored population The Weibull hazard is seen as the better fit for early mortality data obtained with either method 5 4 05 P S ee amp P e oO Fe w 05 J 2 E ae S a j Oo J T 0 005 4 l 0 0005 10 15 20 25 30 Age days of adulthood Figure 11 4 Wildtype hazard at 20 C on HT115 This figure d
52. epicts a hazard rate plot with Weibull fit obtained from a wildtype population 220 animals feeding on HT115 bacteria at 20 C It shows that the general features of the mortality kinetics at 25 C persist at 20 C Nature Methods doi 10 1038 nmeth 2475 50 11 3 Hazard and sample size In Fig 11 5 we demonstrate the impact of population size on hazard estimates The effects of pop ulation size demonstrates the precision afforded by large populations 2 04 e swe amp 05 PEAR E 2 z gece E 2 0 2 ie es 2 B 4 p N N 7 0 05 amp 1000 Individuals 0 014 T T T 10 15 20 Age Days 2 0 ze a S 0 57 S 2 02 2 sie Bv ye 2 S 2 BS fa N ff 005 A 3 ce a Fi ld J ar bi eames 100 Individuals 0 01 ae J n 10 15 20 Age Days Age Days Figure 11 5 Population size and hazard parameter estimates The four hazard panels show em pirical hazard data dots and Weibull fits MLE parameter estimates to original lifespan data for four random samples colors at the indicated sample size This conveys a sense of how sample size impacts the estimation of hazard parameters Nature Methods doi 10 1038 nmeth 2475 51 Supplementary Note 12 Control for scanner effect C elegans researchers define death in terms of macroscopically observable proxies The most widely used proxy is the failure to move in response to prodding with a platinum wire Our aut
53. erage lifespan of all animals on a given scanner was compared to the temperature measured on that scanner Online Methods A linear regression was run with all scanners r 0 163 red dashed line and with the R 3 outlier excluded r 0 746 red solid line a 34 b 25 6 a ce fans off fans on incubator layout 2557 L5 g L4 32 254 i2 G a G 253 oe MAL RV RRA RY gt 25 27 18 E 3 25 14 P3 29 a 3 J 5 XX MM 3 25 04 L1 28 N J R4 27 E aX cae R5 NNNNA ea 26 24 7 25 24 64 RON ee aT 24 ot ot oo ot 24 5 fe 0 1 2 3 4 3 3 5 4 Time h Time h c d 25 6 25 4 O 25 5 4 z D 25 44 ry t e p 25 2 a Z 253 4 3 7 5 25 4 g 25 2 4 s g 25 1 F 24 8 5 8 23541 8 Rk ry 246 5 24 97 J 5 2 248 4 3 2 244 4 e go 5 247 a 5 2421 RS GB 24 6 4 8 24 5 T T T T T T T T T 24 T T T 7 7 7 7 Li L2 13 L4 L5 RI R2 R3 R4 RS 24 242 244 246 248 25 25 2 25 4 Position inside incubator Row Column Scanner surface temperature C Figure 1 3 Consistency of scanner surface temperature Caption on next page Nature Methods doi 10 1038 nmeth 2475 5 Figure 1 3 Cont d a Local temperature differences between and temperature excursions of 10 scanners operating at a scan frequency of 15 minutes inside an incubator Online Methods On the left the fans are turned off on the right they are turned on The inset diagrams the shelving o
54. f scanners for reference in panels b d b Detail of the rectangular area indicated in panel a Fans reduce local temperature differences due to incubator airflow to within 1 C and temperature excursions from scanning to within 0 2 C e Data shown in panel b are reported against the location of the scanners in an incubator adjusted to maintain 25 C with 5 shelves and 2 scanners Left and Right per shelf see inset of panel a Time series are summarized by the mean with error bars indicating one standard deviation The left side of the incubator has a slightly higher average temperature compared to the right side due to incubator airflow d In a separate experi ment the scanner temperatures of a single incubator were measured at one time x axis and again two months later y axis The temperature measures correlate R 0 85 indicating stability of operating conditions Temperature control is especially important when measuring lifespans outside the room temper ature range as with thermotolerance assays e g Fig 5 of the main text To control temperature gradients that result from a scanner s electronics and light source cooling fans are mounted into the chassis of the scanner as shown in Fig 1 1b and schematically in Fig 1 of the main text Fig 1 3 below shows the efficacy of these modifications Nature Methods doi 10 1038 nmeth 2475 Supplementary Note 2 Image processing 2 1 Masking background subtr
55. first scenario unpaired up steps Nature Methods doi 10 1038 nmeth 2475 39 correspond to lost animals see Fig 8 2b In general we can tune the time horizon within which a recapture event is matched up with a gone missing event Specifically given a time threshold 7 we attempt to pair a down step that occurred at time t with an up step that occurred within t 7 If all up steps in that interval have been paired we choose an up step within t 27 r and so on The first scenario corresponds to setting 7 1 one scan interval and the second scenario sets r T the whole experiment In all of this we need to know the total number of worms N at the beginning of the experiment This is the subject of section 8 5 8 3 Worm clusters and interval censoring The most frequent reason for animals to remain unaccounted for in the Lifespan Machine is the irreversible clumping clustering of some worms near the end of their life In the manual setting clumping presents no problem since the experimenter can physically separate individuals that have aggregated There is no such intervention in the Lifespan Machine Early in life when individuals are highly mobile aggregation is dynamic and transient yielding small clusters with odd shapes such as a Y or T that our image analysis software can occasionally recognize as multi worm objects and disentangle computationally Fig 2 5 Later in life when worms become stat
56. g and preparing the fixed scannerlens 3 4 Instructions for focusing ascanner 0 2 eee ee ee ee eee Supplementary Note 4 Software toolset csccscsesessccecevessescveneceesersvecesseeens 4 1 Software components 5 2 86 oo A SANS LO EMR OO EO 4 2 The Worm Browser GUI client 0 0 00 eee ee ee ee Supplementary Note 5 Movement amalysis cccsesecsvecccsecevesceseceveseeseesvecenses 5 1 Identification of stationaryworms e eee eee eee ee 5 2 Posture analysis of stationary worms 020 eee eee Supplementary Note 6 Statistics of worm movement cece cece cece ee eee eee eeeeeeees Supplementary Note 7 A death phenotype cccccceeeeeeeeeeeeeeeeeeeeeeeeeeseeeees Supplementary Note 8 Censoring ssesesessssssessesessessessessesseesesseesseseeseeseeseo 8 1 Right censoring and interval censoring a so 0c ee eeee 8 2 Tracking population size and right censoring strategies 8 3 Worm clusters and interval censoring osa saa ee we ew 8 4 Censoring strategies for worm clusters a ooe soea e eee eee ee ees 8 5 Determining total population size aaa aaa 2 eee ee ee eee Supplementary Note 9 Statistical POWET c ccc cee cee cess cee eee ees eee eee eee eee eeeeneeees Supplementary Note 10 Aggregation of survival curves cceceeeeeeeeeeeeeeeeeneeees 10 1 Device specific en
57. gests that our proxy for death cessation of movement closely matches that of the manual and that any worm that is capable of spontaneous movement does eventually move Nature Methods doi 10 1038 nmeth 2475 53 Supplementary Note 13 Additional survival curves controls ne machine by hand ell machine device corrected z 0 64 gt N J 2 0 47 O S J L 0 2 0 0437 qo ot 0 5 10 15 20 25 Age days of adulthood Figure 13 1 Wildtype at 25 device corrected In Fig 2 of the main text panel 2b shows the wildtype survival data aggregated from the entire experiment and panel 2c shows survival data from individual scanners with their device corrected versions using the categorical regression of Note 10 section 10 2 This figure shows for completeness sake the aggregated device corrected curve alongside the curves of panel 3b for comparison At the aggregate level device correction is seen to have only a minor impact in this case suggesting that the small variations in temperature across scanners and their effects nearly cancel out 10S 20 C wildtype empty vector Aii ya wildtype daf 2 RNAi 0 6 i Aa daf 16 mu86 empty vector daf 16 mu86 daf 2 RNAi automated acquisition men by hand acquisition Fraction survivng 0 0 DELEDELERESEEEEEE EE 5 TITTY TIT ryriirryt TTT O 5 10 15 20 25 30 35 40 45 50 55 Age days of adulthood Figur
58. he Worm Browser where they can be annotated for censoring In Fig 8 5 we demonstrate how various strategies for handling a much more frequent censoring event the disappearance of worms have measurable but ultimately negligible effects on survival curves We therefore do not expect typical bagging rates to have much of an impact on the quality of survival data produced by the Lifespan Machine However certain environments and specific mutations may increase the rate of bagging rendering automated analysis infeasible In such cases we 58 recommend the following strategy for recognizing and removing if desired bagged animals Because bagging occurs primarily early in adulthood when deaths are very infrequent ani mals need not be placed onto scanners until late in their reproductive span Populations can be manually inspected for bagging or any other visible phenotype for several days of adult hood after which they are transferred into the automated imaging apparatus for survival analysis Inaccessibility of plates and animals during an experiment Aclear limitation of our approach is that animals cannot be handled during automated obser vation It follows that nematodes must be sterile during observation so that progeny do not obscure the analysis of parents One common means for ensuring sterility is to add FUdR to media If FUdR is problematic affecting for example the robustness of specific RNAi effects there are several optio
59. he two types of displacements shown in panels a and b are uncorrelated Nature Methods doi 10 1038 nmeth 2475 7 2 3 Worm recognition The image processing pipeline computes for each foreground object a set of features that are sub sequently used by a Support Vector Machine SVM 2 Fig 2 4 to discriminate between worm and non worm objects Fig 2 3 The following is a list of 65 features used by SVM based worm classifier Their distribution of values is shown in Fig 2 2 e Pixel Area e Rectangular Width e Rectangular Height e Rectangular Diagonal e Spine Length e Average Width e Max Width e Min Width e Variance in Width e Width at Center e Width at Front e Width at Rear e Spine SP Length Pixel Area e SP Length Rect Width e SP Length Rect Height e SP Length Rect Diagonal e SP Length Max Width e SP Length Average Width Nature Methods doi 10 1038 nmeth 2475 e Ratio of end widths e Total Curvature e Distance between ends e AI Skew e SP Length Width at Front e Average Curvature e Curvature X Intercepts e Absolute Intensity AI e SP Length Width at Rear e Max Curvature e Curvature Variance e AI Variance e AI Roughness Entropy e AI Roughness e AI Maximum e AI Average Along SP e AI Average in Nbrhd e AI darkest 20 of pixels e AI darkest 20 area e AI Dist from Nbrhd e AI Variance Along SP e Al of Region e AI Normalized Average e AI Normalized Max e AI Normalized Avg Along SP e Relative
60. hine org 4 2 The Worm Browser GUI client The Worm Browser is designed as a desktop client that complements the web interface to the lifes pan machine The Worm Browser performs five separate tasks Nature Methods c Processing and uploading image capture schedules The Worm Browser accepts and pro cesses a high level description of an experimental schedule written as an XML file This schedule file contains a description of the image acquisitions that should be performed in the style of scan all plates on 10 scanners executing a scan every 15 minutes Do this for 28 days The worm browser compiles this file into a list of specific scan times which it uploads to the central database Generating and processing mask files Individual plates must be identified in images of scans The Worm Browser creates the mask files and saves them to local storage for anno tation The annotated mask file is then loaded back into the Worm Browser and the results submitted to the central database Inspecting and annotating worm image data The Worm Browser enables a user to browse through various representations of processed image data including storyboards that con tain images and videos of each identified animal as it goes through the death transition Sup plementary Video 6 Joi 10 1038 nmeth 2475 25 Output of statistical files The Worm Browser can produce a variety of statistical data ger mane to both the operation and the
61. hough image processing typically misclassifies a persistent cluster of n worms as a single indi vidual the death event of a cluster is a valid event associated with the last worm dying Once a cluster is identified our software finds the first stationary worm that nucleated the cluster thus defining a time interval within which the remaining n 1 worms must have died Fig 8 4 This then is handled by interval censoring either using the Turnbull estimator or the interval midpoint see section 8 1 Nature Methods doi 10 1038 nmeth 2475 40 8 4 Censoring strategies for worm clusters Our current image processing software requires manual intervention to distinguish single animals from pairs or triplets lying in close contact to each at death To evaluate the statistical effects of such annotations we compare 6 different strategies for handling multi worm clusters Fig 8 5 All death times were acquired as discussed in Fig 2 of the main text D amp 2 od 2 5 gt S 0 6 7 gt 0 4 I n j c 12 13 2 0 4 Age days of adulthood 7 Best Practice Method Interval censor clusters 0 2 Fully Automatic Method exclude all but last death in cluster 0 0 0 5 10 15 20 25 Age days of adulthood Alternative Methods exclude all animals in cluster include all at cluster death time minimize missing duration right censor at cluster nucleation exclude cluster members
62. ilable at local hardwarde stores Corrosion Resistant Mesh Fencing is available from McMaster Carr for 4 70 per square foot 1x 25lbs Regular non indicating 8 mesh Drierite Useful ancillary equipment Available from drierite com for 83 38 3 3 Recommended scanner modifications In this section we present the steps required for modifying stock Epson v700 Perfection Photo Scanners into a device capable of producing data of the quality presented in our manuscript The modifications accomplish two goals 1 install fans as a means for regulating scanner surface tem peratures and 2 raise the focal plane of the scanner to optimize its use in scanning agar plates We see great potential for the future development of simpler more effective strategies to accom plish these goals The authors consider the current design primarily as a starting point from which the technology can be adapted according to each individual researcher s needs 3 3 1 Overview of the fan mounting procedure The Perfection v700 Photo Scanner has two parts a bottom chassis which contains the CCD im age sensor and a detachable lid also called the Transparency Unit or TPU which contains the transilluminating light source To prevent the build up of heat within the chassis we mount three 8cm computer fans into the sides of the bottom chassis labelled F1 F2 and F3 in the figures These serve to move the heat produced during each scan out of the scanner body
63. ion matrix of the gold standard and represented each animal as a dot projected onto the subspace spanned by the eigenvectors with the three largest eigenvalues principal components PCA 1 3 as shown in Fig 2 4 0 0012 re enon 0 025 5 0 020 4 0 D 4 4 non worm 0 0204 2 2 2 0 0 015 4 06 4 3 3 3 3 z z oy 2 ooo Z on 5 5 a a ee s 0 0104 S S 2 2 2 2 a amp 0005 5 0 005 5 064 0 000 T T 0 000 T y 0 0 1000 2000 3000 4000 5000 0 100 200 300 0 100 200 300 0 10 20 30 animal area pixels length of medial axis pixel widths distance between head and tail pixel widths width average pixel widths 50 0 40 4 0 1 4 46 0 35 35 124 2 0 30 2 3 0 z 4 z 2 2 2 D 32 3 088 3 S os 2 2 0 20 2 2 0 2 3 20 9 2 0 15 2 15 a oe a a a 3 2 S 0 10 Q 10 p 0A a 10 4 a a a 0 05 0 5 0 2 4 OF aaa acca 0 00 0 0 0 0 0 01 02 03 0 10 20 30 40 50 0123456789 11 13 15 17 19 0 10 20 30 40 50 60 70 ratio of medial axis length to area ratio of medial axis length to maximum animal width r x Maxiumum Curvature Along Medial Axis i s j h curvature average 1 pixel width Ai 5 pixel widths per pixel pixel widths per pixel 1 pixel width 0 06 35 0 05 30 4 0 04 4 25 J 20 4 0 03 15 4 0 02 4 40 0 01 4 05 4 0 00 T T T T T T 00 errr T 20 30 40 50 6
64. ionary clusters tend to be more persistent and worms appear to adhere to one another acquiring an overall linear shape that our image analysis almost always misidentifies as a single worm Future developments in image analysis might be able to detect and disambiguate worm clusters However at present the fastest and most reliable alternative is to identify and annotate clusters by visual inspection using the Worm Browser Note 4 section 4 2 This task can be performed easily and quickly Processing an experiment with three thousand animals takes less than an hour and is part of the data validation recommended for any novel data set to confirm that all steps of the image analysis pipeline are functioning as expected second worm enters cluster second death first worm enters cluster first death t ti to t time 2 Figure 8 4 Interval censoring of a terminal worm pair schematic The first worm becomes stationary at t A cluster is nucleated once a second worm joins A terminal pair is created if the latter becomes stationary in the cluster The last of the two animals dies at tz a valid individual death time The first worm to die unnoticed within the terminal pair is then assigned upon visual inspection of machine death calls a death time corresponding to the midpoint of the time interval between t and t2 To the Lifespan Machine the whole episode appears as a single worm becoming stationary at t and dying at t2 Alt
65. is metric shows the same trend as worm area providing further evidence that mor phological changes surrounding the death time are independent of our segmentation algorithm d We determined the average intensity of pixels inside worms over time This metric too shows the same trend as worm area suggesting that the animals both shrink and diffract less nearing death while getting larger and darker post mortem a b 25 C 20 C same legend as in a wildtype empty wildtype daf 2 RNAi daf 16 mu86 empty daf 16 mu86 daf 2 RNAi Worm area fraction of area at death Worm area fraction of area at death 0 9 2 1 0 12345678910 2 1 0 1 2 345678910 Time relative to death d Time relative to death d Figure 7 3 Generality of the volume dip To assess the generality of the late life characteristics described in Fig 7 1 we performed the same analysis on animals grown on a different bacterial food source HT115 a Area vs time relative to death on HT115 at 25 C b Area vs time relative to death on HT115 at 20 C Nature Methods doi 10 1038 nmeth 2475 35 Supplementary Note 8 Censoring 8 1 Right censoring and interval censoring 0 ee eee cece eee eee eee 36 8 2 Tracking population size and right censoring strategies 064 37 8 3 Worn clusters and interval censoring cc cece cee ces enews eecens 40 8 4 Censoring strategies for
66. ive to death a 1 254 25 C wildtype daf 16 m86 35 C wildtype 2 AE T aN g w w J anm on 61 154 daf 16 m86 os age 1 hx546 os age 1 hx546 So 144 eg ty Ea J ES J 6 1 054 6 1 054 gs s o 5108 Sa 2 44 Bo 5 0 95 0 95 OQ perp pt 0 91 l l j l j 2101412345678 910 0 1 0 0 1 02 03 04 05 Time relative to death d Time relative to death d e f 1 0 1 0 F 2 4 8 054 wildtype 8 0 55 wildtype 0 0 2 0 0 5 1 0 5 1 0 8 4 0 54 daf 16 mu86 o 05 daf 16 mu86 J oo g 0 0 g qo 3 1 0 a 2 054 age 1 hx546 2 05 age 1 hx546 O 0 0 0 0 i j 0 9 1 i 1 2 0 9 1 1 1 1 2 Worm area fraction of area at death Worm area fraction of area at death Time interval 75d 0 25 d Time 2h relative to death 0 25d 0 25 d relative to death Oh 1 75 d 2 25 d 3h Figure 7 1 The volume of a worm dips at or near death a We followed the total area in pixels of a single wildtype animal dying on OP50 at 25 C This area is plotted against time before and after the animal s death t 0 as determined by visual inspection Dots represent scores solid lines are cubic spline fits 3 18 The size of the animal decreases starting about a day before its apparent death after which the trend sharply reverses and size increases The volume of a worm dip
67. k in the Worm Browser Note 4 section 4 2 and Supplementary Video 6 The thresholds are chosen such as to minimize the de viation between computed and certified death times of the gold standard set Fig 5 1 Fig 5 1 also indicates that the death calls are fairly robust with regard to the so chosen parameters Using these parameters the automated analysis of posture changes returns the first t at which A last falls below threshold for the required duration of time The discrete time interval associated with the death event is t t 1 and corresponds to the wall clock time elapsed between onset of adulthood and scans t and 1 We examined several image transforms such as equalization or normalization prior to pairwise comparison of images that might amplify and clarify local motion Yet none yielded a distinctive advantage over the simple brightness adjustment equation 3 Nature Methods doi 10 1038 nmeth 2475 30 Supplementary Note 6 Statistics of worm movement The movement analysis described in sections 5 1 and 5 2 can be used to partition the population of worms into movement classes similar to those defined in 19 Stationarity or posture span refers to the time between cessation of translatory locomotion and death when the only motion consists in subtle changes of head and tail position without affecting overall location For wildtype and age 1 hx546 animals shown in Fig 3 of the main text we calculated the stationari
68. lates with thinner agar lawns as these situations yield more distinct images of worms Our posture analysis algorithm works best on animals that do not exhibit long pauses in late life move ment as our use of a pause threshold to remove noise works less well in these cases Occasional imperfect performance in all these components can be seen by scrutinizing Supplementary Videos 1 5 We expect that improvements will be made in many components of the Lifespan Machine For example movement time series analysis might use Hidden Markov models or worm detec tion strategies might integrate image data from consecutive frames The consequences of such local improvements however should be evaluated with regard to both the independent behavior of the component and its integrated effect on the overall quality of the survival curve Beyond transient limitations that are more reflective of the current state of development of the Lifespan Machine our approach is restricted in some more fundamental aspects e Worm density on a single plate Our software can analyze 100 or more worms on each plate In practice however worm be havior introduces complexities that limit the usefulness of our analysis of high density pop ulations At all densities we see animals often die closely as pairs juxtaposed head to head and tail to tail This circumstance makes it hard for the image analysis to automatically distinguish a single dead animal from a pair of dead animals
69. le scanners allows observation times to be staggered thereby permitting a finer temporal resolution for capturing the time dependent hazard function Twenty or more scanners Construction of a scanner farm requires the dedicated use of several incubators or a tem perature controlled room At this scale any number of projects becomes possible including further development and testing of the apparatus itself in parallel with running experiments The tunable statistical resolution makes our method ideally suited for large rapid reverse genetic and chemical screens for effects on survival at intermediate or higher statistical res olution Acknowledgments We would like to thank Joy Alcedo for providing the hsf 1 and glp 1 mutant strains Becky Ward and Debora Marks for critical reading of our manuscript and Catalina Romero Debora Marks and members of the Fontana lab for helpful discussions and encouragement throughout this project We would like to thank Tom Kolokotrones Eric Smith and Lee Jen Wei for discussions and sta tistical advice and Mason Miranda our departmental IT specialist for patiently meeting our un bounded needs for data storage This work was funded by NIH grants Ro3 AGo32481 Ro3 AGO32481 S1 and Ro1 AGoO34994 01 References 1 Jones J 2008 GIMP User s Manual 2 Boser BE Guyon I Vapnik V 1992 A training algorithm for optimal margin classifiers In Proceedings of the Fifth Annual Workshop on Computational Lear
70. lent GPL licensed statistical software environment In R we used the surv rms flexsurv and muhaz packages Worm Browser See section 4 2 Image acquisition server software Custom software controls experiments by communicat ing with the scanners via USB Scanning schedules are stored in the MySQL database and take roughly the form At time run a scan on device D between positions 1 y1 and z2 y2 Typically four plates are imaged in one scan representing a column of plates on the scanner The server monitors the schedule database and executes scans as they become due Various telemetry is collected including the start and stop times of each scan various gross image properties such as average image intensity and a description of delays or errors that oc curred Since a Lifespan Machine may run for three weeks or more it is likely that errors will crop up in some system component such as a scanner the network connection a power supply etc The image server software has several layers of redundancy and error handling to mitigate such problems including the ability to send emails or text messages reporting new problems Image processing software The image processing pipeline and associated job scheduling software is written in C The Worm Browser is written in C using the library FLTK GNU GPL V2 to provide cross platform GUI functionality 5 Availability All software is currently 2013 accessible at www lifespanmac
71. lette for worm recognition sssssssessssrssecrsseossereseesseereseeresees 9 2 3 Worm and non worm objets i isin scisacscsrrivakircoaiedasehsaasaieansesoareoesasseuranes 10 2 4 Support Vector Machine for worm recognition ssssrsssesssssssesssssesesrsseee 10 2 5 Image segmentation examples ssesesseeesssssescensssseseoseseoersaresesersseeeese 11 3 1 Scanner modification diagram sesressssseresressssesseessseosseressesseresrerssee 18 3 2 Photographs of chassis modifications sssssssssssussssrsssesssreseeossesssereess 19 3 3 Photographs of mounted scanner fan sssssssssssssrssssresressereseeosseesseesess 20 a Fan wrm dari a eA gaa ea NS 21 3 5 Inside the scanner chassis osesessssesesensnseseseeseseseerosesessesesesrereseseo 23 3 6 Reference images demonstrating sufficient focus cssccescesssecseeeeseeseees 24 4 1 Worm Browser Screen shot sssesessssosesesssseseseesesesreceseseserseseseeresesso 26 5 1 Parameter robustness of posture analysis cc cece cee e eee e eee eee eeeeeeeeeeeenees 30 6 1 Population statistics of movement Classes ccccccccccccccccccceeceeceeceeeenees 31 6 2 Stationarity posture span ss sssssesessesesessresensessesessecesercsceseseesseeeese 32 7 1 The volume of a worm dips at or near death 2 0 2 cacceccacscccascaceascaeedsndswasnasnaens 33 7 2 Coincidence of volume dip and cessation of MOTION cccc
72. ls a and b illustrate the sample size dependence of the Log rank test In 800 replicates we drew two sub population of the specified size abscissa at random from a wildtype and daf 16 mutant population observed at high temperature 34 5 C by the Lifespan Machine At this temperature the daf 16 mutation appears to produce a 10 reduction in average lifespan Fig 5a For each pair of random subsets we used the Log rank test to decide whether or not to reject the null hypothesis that the samples are statistically indistinguishable Panel a shows the scatter of p values found in each replicate at each sample size To allow fast rendering of the figure only a random subset of all 800 replicates at each size are shown Note that the ordinate of panel a is logarithmic The red line indicates 99 confidence i e 0 01 significance Caption continued on next page Nature Methods doi 10 1038 nmeth 2475 43 Figure 9 1 Cont d Each dot above the red line is a test with non significant outcome incor rectly failing to reject the null hypothesis The green line marks the median of the p value distribu tions Panel b shows the frequency with which the wrong null hypothesis is correctly rejected as a function of sample size This is known as the power of the test The complement of this fre quency is the type II error rate or the rate of false negatives It can be readily seen how the power increases with sample size The bottom panels
73. lt from small differences in these conditions It is therefore important to maintain a constant environment across the different plates in an experiment This is technically difficult using the manual assay and the problem is compounded when using a large distributed apparatus like the Lifespan Machine Factors such as temperature Fig 1 3b or the composi tion of the bacterial lawn may vary slightly between plates and scanners creating distinct micro environments that can impose extrinsic lifespan variation on the intrinsic variation of interest Scanners for example can differ in temperature because they produce and dissipate their own heat within a large confined space incubator whose airflow they obstruct We see temperature differ ences of up to 1 C Figs 1 3b and 1 3c between scanners located inside the same incubator and identify statistically significant differences between the lifespans of animals housed on different scanners Fig 2c of the main text and Fig 1 2 Since the populations on each scanner are drawn from the same homogeneous population the differences between survival curves most likely reflect temperature differences rather than biological variation We deploy fans Fig 1 1 and optimize scanning frequencies Figs 1 3b to minimize temperature differences between scanners Ideally these modifications would reduce environmental variation to the point where its effects fall below the detection limits of our
74. main text misses the juxtaposition of the survival curves acquired by machine with those obtained by visual inspec tion with the Worm Browser for the mutants shown here in panels a and b Visual and automated curves were indistinguishable by Log rank p gt 0 04 glp 1 and hsf 1 strains were grown at 15 C before being shifted to 25 C Cumulative probability 1 0 0 8 35 C 0 6 age 1 hx546 Wild type 0 4 daf 16 mu86 0 2 0 I 8 6 4 2 0 2 4 6 8 Error in machine death time hours Figure 13 8 Error of LM operation at high temperature To complement the visual validation of the LM at 35 C we plot the LM error as in Fig 2f of the main text associated with Fig 4c of the main text Nature Methods doi 10 1038 nmeth 2475 57 Nature Methods doi 10 1038 nmeth 2475 Supplementary Note 14 Limitations As a system the Lifespan Machine is best evaluated in terms of the quality of the survival statistics it produces Yet components such as imaging hardware worm classifiers and movement detec tion algorithms can also be evaluated in isolation For example adoption of imaging devices other than flatbed scanners e g cell phone cameras may decrease both the physical bulk of the Lifes pan Machine and the potential for temperature variation between plates Furthermore we find that our SVM classifiers produce the best results on younger worms strains with larger bodies and p
75. me indicates the number of animals that have joined the ranks of the unaccounted for Each down step indicates the number of those that have become observable again The orange curve is the censoring curve Each up step indicates the number of animals lost from further observation and that must therefore be right censored at that time inter val a In this scenario the orange curve is constructed right to left matching the green curve at first Each green down step of size 1 is matched against the closest up step of size 1 and the orange curve is extended to the green level at that last match b In this scenario each green down step is matched at random with a previous green up step When all down steps have been paired the orange curve is built left to right tracking the number of unpaired green up steps Although the green curve m t is the same in cases a and b the two strategies result in quite different censoring curves When m t decreases some worms that went missing before t are again accounted for Yet we have no information to deduce when these recaptured animals went missing In other words we cannot determine unambiguously how many of the worms that went missing at a given time were actually lost Consequently we cannot know exactly how many worms must be right censored at what time The many possibilities are bounded by two extreme scenarios In one extreme scenario we hypothesize that the recaptured animals are those th
76. n data There may be mutants whose behaviors makes automated analysis impractical for example mutants that show no movement late in life or mutants that crawl off the plates at a high fre quency or mutants that bag cryptically late in life The possibility of these cases underscores the importance of quality control with the Worm Browser to detect them Nature Methods doi 10 1038 nmeth 2475 59 Natura Matt Nature Mett Supplementary Note 15 Resolution and scalability The apparent smoothness of a survival curve depends not only on population size but also on sam pling scanning frequency A large population that is observed infrequently will yield a survival curve with large steps as many deaths accrue between successive measurements An example is the wildtype manual control presented in Fig 2b of the main text where a population of 500 animals is observed at a frequency of once a day producing large steps near the mean lifespan At the other extreme frequent observation of small populations also yields survival curves with pronounced steps as in that case each individual death accounts for a sizeable fraction of the pop ulation Evidently some advantages of large populations are only fully realized when paired with frequent measurements Therein lies one advantage of an automated method like the Lifespan Machine Another advantage is scalability When the overall lifespan of a population is short as it is for example at 35 C
77. ng to equation 5 which minimizes the time a worm is assumed to have been missing Curve 6 cyan As for curve 5 but the procedure pairs missing events and recapture events at random see section 8 2 Nature Methods doi 10 1038 nmeth 2475 41 8 5 Determining total population size Sometimes the total number of worms N at t 0 may be unknown for example when animals are loaded onto plates without being counted In this situation we estimate the total population size by determining and comparing two quantities First we take the maximum of a t D t fort 1 T see section 8 2 and Fig 8 3 This is the largest number of worm objects ever identified during the course of the experiment Let this number be No The visual identification of final clusters with the Worm Browser provides another estimation of the total number of worms call it Ne N is the number of worms found in clusters at the end of the experiment plus the number of singleton death events D t as determined by the machine We take N to be the maximum of No and Ne Nature Methods doi 10 1038 nmeth 2475 42 Supplementary Note 9 Statistical power The effect of a perturbation on lifespan is established through comparison of the perturbed popu lation against an unperturbed control using a statistical test In the case where the perturbation does have an effect we would like the test to reject the null hypothesis that the two populations are identical not only with
78. ning Theory ACM Press pp 144 152 3 Sall J Lehman A Stephens M 2012 JMP Start Statistics a Guide to Statistics and Data Analysis Using Jmp Fifth Edition SAS Institute 4 R Development Core Team 2009 R A Language and Environment for Statistical Computing R Foundation for Statistical Computing Vienna Austria URL http www R project org ISBN 3 900051 07 0 62 Nature Methods doi 10 1038 nmeth 2475 5 Kuhn H 1955 The Hungarian method for the assignment problem Naval Research Logistics Quarterly 53 6 Apfeld J O Connor G McDonagh T DiStefano PS R C 2004 The AMP activated protein kinase AAK 2 links energy levels and insulin like signals to lifespan in C elegans Genes Dev 18 3004 3009 7 Kaplan EL Meier P 1958 Nonparametric estimation from incomplete observations Journal of the American Statistical Association 457 481 8 Turnbull BW 1976 The empirical distribution function with arbitrarily grouped censored and truncated data Journal of the Royal Statistical Society Series B 38 290 295 9 Stata Corporation 2007 Stata Base Reference Manual Release 10 College Station TX Stata Press 10 Buckley J James I 1979 Linear regression with censored data Biometrika 66 429 436 11 Ward A Liu J Feng Z Xu XZS 2008 Light sensitive neurons and channels mediate photo taxis in C elegans Nature Neuroscience 11 916 922 12 Mathew M Mathew N Ebert P 2012 WormScan A
79. notype These animals show abnormal development of the nervous system and exhibit both slow and uncoordinated movement 15 Using the posture analysis parameters optimized for wildtype animals our image analysis consistently identifies death times of unc 119 as being on average only 1 5 days shorter than wildtype The death times of unc 119 were slightly under estimated by 0 3 days in mean lifespan compared to the visual inspection This difference is not statistically significant p 09 Movement parameters can be optimized for specific behavioral mutants if necessary Nature Methods doi 10 1038 nmeth 2475 55 a 40 D 45 25 C o gt 0 87 o gt 0 87 J lt Z 0 65 2 0 64 2 4 no J S 2 0 4 amp 0 4 O iS Loa visual inspection E oo wild type unc 50 e306 a 0 0 HR Rete z 0 0 0 2 4 6 8 T 10 T Tor 12 14 16 18 Age days of adulthood 25 C visual inspection wild type rol 6 su1006 unc 4 e120 Ps ree eh ie eae et eae 0 2 4 6 8 10 12 14 16 18 Age days of adulthood Figure 13 5 Survival curves of unc 50 e306 and unc 4 e120 rol 6 su1006 at 25 C Panel a Hermaphrodites possessing the unc 4 e120 allele have been previously characterized 16 as hav ing a significantly shorter lifespan than animals with the wildtype allele Pairwise comparisons be tween automated analysis colored curves and visual inspection of the image record
80. ns available to the experimenter such as the use self sterile mutants see Fig 13 6 If the experimenter can tolerate some loss of precision during the first few days of death animals may be passaged by hand as young adults and then loaded onto scanners after they cease laying eggs during mid life as in the bagging case described above If the research community as a whole decides to move away from FUdR new meth ods will need to be devised to prevent live progeny Because of our adherence to standard culturing techniques these methods are likely to be applicable in the context of the Lifespan Machine as well Another is the use of temperature sensitive sterile strains However these solutions may not always be appropriate and in certain cases the reproductive span of in dividuals may persist into late ages precluding the simple solution of transferring animals onto scanner only after they have ceased reproducing The Lifespan Machine would not be very useful in such instances In the main text we demonstrate the use of the Lifespan Machine as an assay for survival under exposure to the oxidant t BuOOH Fig 4 of main Quite generally our method can be employed to assess the impact of chemicals on lifespan The inaccessibility of plates how ever requires that chemicals must persist at defined initial concentrations throughout an experiment as they cannot be replenished Traits that may affect the acquisition and interpretation of lifespa
81. nt to detect 10 lifespan differences between two genotypes at 99 confidence and 99 power Therefore a single scanner enables standardized high resolution survival assays appropriate for detecting subtle genetic effects A single scanner installation involves three components an incubator for temperature control a scanner for image capture and a computer to control the scanner and store data The incubator is standard equipment in most labs working with C elegans and small installations may simply use one shelf of a pre exisiting incubator The use of single scanner installations constrains the number of different populations that can be simultaneously monitored and the statistical power at which those populations can be characterized limiting the number of possible applications For example the characteriza tion of mortality rate deceleration may not be feasible using a single scanner since hazard deceleration becomes prominent only after the median lifespan at which point the dwindling population size attainable with single scanners may provide insufficient statistical power Moreover we find several instances where the proportional hazards assumption of certain statistical tests e g Fig 3b are not appropriate necessitating the use of less powerful sta tistical approaches requiring larger populations to be monitored All of these limitations are compounded by the loss of worms Note 8 section 8 1 from plates as well as the l
82. o mated approach uses a slightly different proxy the permanent cessation of all spontaneous motion as determined by retrospective analysis of long term observations We observe these two proxies correlate surprisingly well Fig 2b and Fig 12 1 Yet absent a principled theory of death it seems conceptually impossible to control for one proxy by comparing it to another Specifically we can not expect the comparison of a manually performed lifespan experiment to act as a strict control for an automated experiment as the two measure somewhat different phenotypes This is why we established quality and consistency of our method in two ways First by evaluating the Lifespan Machine s ability to recapitulate in a self consistent fashion known effects on lifespan resulting from a wide variety of interventions such as mutations RNAi and exposure to stressors of physi cal heat and chemical t BuOOH nature Figs 3 and 4 of the main text Second by developing software facilitating the human ex post inspection and validation of the visual record associated with an experiment a b 1 0 o gt 0 8 machine scanner lt by hand scanner Z Z by hand box a a 0 6 wn wn Cc Cc Q Q 0 4 5 machine scanner S by hand scanner V by hand box by hand scanner Lr 02 ee r b h d b 24 5 24 6 24 7 24 8 24 9 ae 25 1 25 2 y nan OX Temperature C 0 0 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4
83. o confirm that the light source in the TPU functions properly 2 Disconnect the scanner and remove the TPU 3 The top surface of the scanner chassis is secured by several screws located under small plastic covers in each corner Pop out the plastic caps and release the screws Pop off the top surface 16 Nature Methods doi 10 1038 nmeth 2475 Natura Matt Nature Mett of the scanner and place it somewhere safe Be careful not to get finger prints on the bottom of the scanner glass 4 Using your hand gently push the scanner bar 3 4 of the way toward the front of the scanner to move it out of harm s way 5 Place a piece of tape over the long slit opening of the scanner bar to prevent plastic chips from entering the bar 6 Using a spiral saw or your favorite cutting device cut holes Hi H2 H3 and H4 Cuts H1 and H3 run somewhat close to internal scanner wiring The cuts shouldn t come extremely close but be careful and fold any nearby cables out of harm s way Any time you use a spiral saw wear Safety goggles and stay aware The spiral saw tends to throw little chips of plastic 7 Take an 8cm fan and use it to confirm that a fan will fit snug into each hole Enlarge the holes where necessary However don t remove more plastic than required as doing so will make it harder to hot glue the fans in place 8 Using compressed air either from a can or from a hose hooked up to house air blow away all the plastic chip
84. on and dimensions for each cut are illus trated for both the scanner chassis and the Transparency Unit TPU Nature Methods doi 10 1038 nmeth 2475 18 Figure 3 2 Photographs of chassis modifications a The scanner chassis with its lid removed and four cuts Hi H4 made into its walls b e Close up views of each cut f h 8cm fans are glued into holes Hi H3 Nature Methods doi 10 1038 nmeth 2475 19 Figure 3 3 Photographs of mounted scanner fan a The Transparency Unit TPU is shown with its top and bottom separated b c Three cuts H5 H7 are made into the bottom of the TPU d Four holes are drilled into the TPU lid and e filled with machine screws f The holes are shown from the top of the lid Nature Methods doi 10 1038 nmeth 2475 20 12V DC 0O2A 1A 16A Figure 3 4 Fan wiring diagram The wiring that supplies the fans is shown MX indicates a standard Molex connector The connection between fans F4 F8 can be made by clipping off the Molex connectors and splicing the wires together using solder or by using a butt clamp or a wire nut Be careful to add appropriate insulation Be careful to plug the Molex connectors in the correct order otherwise excess current may be passed through each connector Be careful to use an appropriate of gauge wire for the indicated current The eight fans attached to each scanner draw about 1 6 amps at 12V DC To prevent heat build up
85. on and object recognition using the SVM illustrat ing the difficulties of disentangling worm clusters Refer to sections 8 2 and 8 3 as well as Figs 8 3 and 8 5 for more detail on the accounting of multi worm clusters Supplementary Videos 1 5 provide time lapse movies of plates with wildtype age 1 hx546 and unc 64 e246 mutants for the entire duration of an experiment both with and without metadata overlay unprocessed segmented classified Image image objects Figure 2 5 Image segmentation examples Caption on next page ae 1 Nature Methods doi 10 1038 nmeth 2475 a Figure 2 5 Cont d a A scanner raw image of an agar plate with worms on patches of bacterial lawn E coli OP50 Scale bar on the left corresponds to 5mm the scale bar on the right to 1mm b The same image after background subtraction using a median filter with foreground objects resulting from image segmentation highlighted in color Objects classified as worms by the SVM are colored red non worms are colored blue c Enlarged images of worms indicated in panel b as they appear in consecutive image processing steps The scale bar marks 1mm The first third and fourth rows are examples of correctly processed animals The second row illustrates the difficulties of parsing multiple worm clusters Note how several of the worms are incorrectly identified Such errors when they occur in the context of dead worms can be rapidly identified in
86. oss of entire plates due to fogging or desiccation which often lowers the population sizes captured on a single scanner from their theoretical maximum by about 10 Five or Ten Scanners The limitations of single scanner installations are easily alleviated by using multiple scan ners The lifespan machine is designed to allow multiple scanners to be deployed with little additional difficulty or complexity compared to setting up a single scanner If a laboratory acquires a dedicated computer and incubator for automated lifespan analysis it is not very laborious to fill the incubator with scanners A large incubator with 5 shelves will hold 10 scanners forming a natural medium sized installation Scaling up from a single scanner to ten scanners is merely a matter of purchasing modifying and plugging in the additional scanners our image server software is designed to accept new scanners even while experi ments are running A ten scanner installation affords population sizes that provide greater statistical power not only for standard analyses such as the Log rank test or Cox regression but also for the ap plication of somewhat more exotic analyses such as non proportional hazards approaches or techniques that directly estimate the time dependent hazard function Our intended pur pose of providing high statistical power is not to simply permit detection of increasingly small lifespan effects as there will be a point when statistical signifi
87. our scanning schedule these intervals are typically either short 15 minutes or long 1 hour and 45 minutes but they can be longer if the system fails to detect a worm in scans immediately preceding its death Quite generally when events are only known to have occurred within an interval a number of approaches exist to handle what is generally referred to as interval censoring One interval censoring technique is the Turnbull estimator 8 provided by statistical packages such as JMP 3 or STATA 9 When scan intervals are small compared to even the shortest lifespan in the population a reasonable and computationally much faster alternative consists in simply choosing the interval midpoint as the time of death for all events known to have occurred in that interval and then apply standard time to event KM Nature Methods doi 10 1038 nmeth 2475 36 We compared the Turnbull estimator with the simple midpoint estimation Fig 8 1 Using Turn bull estimation the median of survival fell between 12 days 3 53 hours and 12 days 3 60 hours Using the interval midpoint the median of survival occurred at 12 days 3 hours This 36 minute difference is overwhelmed by several sources of experimental error that can produce much larger differences for example temperature In view of this negligible difference the substantially faster speed and the more useful analytics provided by many statistical packages when using time to event KM we chose th
88. played by this rate in a non reproducing population like a plate of sterile worms is analogous to that of a rate con stant in a first order chemical reaction that thins out an initial population of molecules by degra dation However the hazard rate h t is typically not a constant but a time dependent quantity that may and often does increase with time reflecting the functional decline associated with ag ing ds t dt h t s t where s t is the survival function h t can be observed directly by counting the frequency of death events relative to the survivors the population at risk at the be ginning t of intervals of fixed duration At period hazard h t s t s t At Ats t To do this accurately however requires population sizes that are larger and more frequently observed than is typically the case in manual lifespan experiments The Lifespan Machine makes this fairly straightforward and the machine can be scaled up if desired to test many conditions We determined three kinds of parametric hazard functions for our data Gompertz Weibull and Log logistic Gompertz and Weibull hazards represent exponential and polynomial increases in the risk of failure respectively and are widely used to describe mortality data in biological and technological systems The Log logistic is a Weibull distribution with a deceleration term We used it here as a reference to compare the deceleration of mortality rates observed in most c
89. pment and improvement Figure 1 1 Lifespan Machine hardware a A Thermo Forma incubator provides temperature control for ten flatbed scanners b A scanner unit with open lid and plates loaded showing fans mounted on the lid to create a flow of air between lid and scanner surface when the lid is closed c A scanner unit with closed lid as during operation shown from rear to highlight fans mounted into the chassis d The typical plate configuration consists of 16 plates inserted face down through holes punched into a rubber mat The mat seals the plates to the glass surface of the scanner stabilizing their position and preventing dessication Note the light source located in the lid trans illumination mode Nature Methods doi 10 1038 nmeth 2475 4 1 2 Temperature control during image acquisition An important aspect of our method is the control of temperature within a tight range reducing the large excursions that are sometimes unavoidable in the manual method where plates are moved in and out of an incubator for scoring The notorious temperature sensitivity of C elegans is high lighted by its variation between scanners Fig 1 2 Mean worm lifespan da 248 25 252 254 256 Mean scanner temperature C Figure 1 2 The effect of scanner temperature on lifespan The lifespans of wildtype animals were acquired as described for Fig 2 of the main text and loaded onto 160 plates distributed across 10 scanners The av
90. rmed using the flexsurv package of R 4 A quantile quantile Q Q plot is a diagnostic for assessing the extent to which survival data con form with the distributions whose parameters we determined In a Q Q plot we graph the time it takes a functional form Gompertz Weibull or Log logistic to reach a certain quantile against the time it takes the empirical survival function to reach the same value In our figures we place the former on the abscissa and the latter on the ordinate The values plotting positions are de termined by the steps in the Kaplan Meier survival curve obtained from our data Note 8 section 8 1 The 95 confidence intervals in the Q Q plot derive from the confidence intervals generated Nature Methods doi 10 1038 nmeth 2475 47 in the survival curve assembly from data Greendwood s formula using the survfit function of R s survival package The extent to which empirical data and MLE fitted functions coincide can be assessed by how closely the points hug the identity line It is worth pointing out that while we determine parameters based on the data up to the median lifespan we extend the Q Q plot to in clude the entire range of data This makes the onset of deceleration neatly visible as an upward departure from the diagonal for Gompertz and Weibull hazard functions 11 2 Hazard of wildtype The next two figures evidence that high resolution lifespan data for wildtype conform with a Weibull distribution better
91. rstanding that in the proportional case the t denote the logarithm of lifespan The model can be written as tig Q Pi eg ifiA s 6 s 1 tsj a Bit sj 7 i 1 In these equations a is a landmark against which the device dependent effects are determined A possible choice is the grand mean 1 of all N lifespans or their log measured in the same incubator or the average lifespan of a population on a particular scanner The grand mean is a good choice when the environment varies evenly between scanners and worms are distributed evenly across the scanners When a is the average lifespan on a specific scanner with temperature T we effectively correct the lifespans on all other scanners as if they were at temperature T The 8 i 1 s 1 are the deviations of the mean lifespan u gt gt jtij N of the N individuals on scanner 7 from a and e is the deviation of t from y Since we must have 5 _ 6 0 8 is determined by B Be The relation with linear regression is made explicit by introducing a categorical variable X j whose vector value encodes the scanner hosting individual j tj a Xj6 6 8 with the vector of scanner effects to be determined Estimates of the parameters of these types of equations are handled routinely with statistical packages such as R 4 or JMP 3 In our case we also must account for the fact that we place animals on scanners several days after onset of adulthood The tim
92. rved below Amax for longer than a threshold value 7 which exceeds the maximum resting time of live animals To obtain these two threshold values Ajax and 7 we assemble a gold standard time series for several hundred worms whose re Methods doi 10 1038 nmeth 2475 29 Q z S 1 25 4 s 5 14 g O 5 0 75 5 7 ne S a 2 g c 057 g 5 025 4 i Z 0 T T T T 0 1 1 x x 0 0005 0 001 0 003 0 006 0 01 0 02 0 04 0 06 0 1 02 0304 06 0 0005 0 001 0 003 0 006 0 01 0 02 0 04 0 06 0 1 02 0304 06 Movement threshold Movement threshold Figure 5 1 Parameter robustness of posture analysis To determine an optimal combination of the parameters Ajax movement threshold and 7 hold time image strips of several hundred animals were inspected using the Worm Browser and their death times were annotated Many combinations of threshold and hold time values were systematically tested Each combination was used to estimate a death time for each animal using our posture analysis algorithm resulting in an estimation error when compared to the by hand annotation The panels show the mean squared error a and the standard deviation of the squared errors b The algorithm produces accurate and robust results for threshold values between 0 01 and 0 02 and for hold time values between 6 and 14 hours The chosen parameter combination is shown as a green dot arrow death times are annotated by visual inspection of the image stac
93. s at or near death Figure on previous page b A single wildtype animal dying at 35 C exhibits a consistent increase in size after its death Caption continued on next page Nature Methods doi 10 1038 nmeth 2475 33 Figure 7 1 Cont d Data prior to death is scant because the stationary movement phase at this temperature is very short c To evaluate whether death related contraction and expansion of worm area is typical we determined individual time series of areas for 44 wildtype 90 daf 16 mu86 and 143 age 1 hx546 animals whose death time on OP50 at 25 C was determined by visual inspection of the image stack using the Worm Browser A cubic spline was fit to minimize the mean squared error across all individuals at each measurement This reveals a conspicuous contraction before death and expansion thereafter d We proceeded as in c for 189 wildtype 304 daf 16 mu86 and 194 age 1 hx546 animals on OP50 at 35 C Animals dying from acute heat stress were not observed to show a decrease in size immediately before death This either repre sents a true difference in late life morphological changes compared to animals dying at 25 C or the shrinking close to death occurs while animals are still crawling and can therefore not be measured with this technique e The average behavior of worms shown in c and d may mask individual variation To estimate the effect of such variation we grouped all measurements made within three h
94. s produced by the cuts If ignored these plastic chips have the nasty habit of working their way into moving parts in particular the scanner s power button If this button ever stops working there s probably a chip in it somewhere 9 Place a 8cm fan into hole H Make sure the power cable is facing towards the front of the scanner preferably from the upper right hand corner of the fan Fan F1 should face outwards the airflow should be heading into of the scanner Glue it in place along all four sides You may need to hold the fan in place with one hand and glue with another In the same way glue fans F2 and F3 into holes H2 and H3 Fan F2 and F3 should both be facing outwards 10 Remove the protective tape added earlier to the scanner bar 3 3 3 Modifying the transparency unit 1 Put the scanner chassis aside and consider the scanner TPU The bottom of the TPU is se cured to its the cop cover by eight screws Release them and pop off the bottom cover and place it somewhere safe 2 Make cuts H5 H7 Note that the TPU connector is located at the center of H7 Remove this cable from its notch and make cut H7 You might consider cutting a new notch somewhere next to cut H7 so that this cable can be again secured 3 Using a power drill or the spiral saw drill holes S1 through S5 Make sure to use a bit appro priate for the machine screws you are using as you ll want the screws to fit snugly 4 Using compressed air either
95. s refer to an animal that is temporarily unaccounted for as missing to distinguish it from being lost i e permanently missing For the purpose of adju dicating right censoring events a worm inside the Lifespan Machine can therefore be in three states missing lost or dead A significant number of animals may be missing at times In the manual procedure observers can search for missing animals and physically separate clumps but the Lifespan Machine has no such options This difference requires additional processing to han dle censoring First we must track the total number of animals on a plate from which we can infer missing animals Second we must determine how many among the missing animals were lost i e never returned from the missing state and at what times these losses occurred Because young worms move large distances between consecutive observations we cannot directly determine when a particular animal goes missing nor in case of its return the time when it becomes observable again Instead we apply a tunable heuristic explained below to estimate the timing of both events To estimate the statistical effects of this heuristic on our survival curves we compare two extreme cases that delimit whatever unknowable histories have occurred in an experiment The curves derived from these extreme censoring protocols are compared to each other in Fig 8 5 Let t index scans Let d t denote the number of deaths identified at tim
96. th pi P po P 2 pn E PO ty lt ta lt lt tn 1 The time ordered series of stationary positions p P are meant to correspond to the same spatial location across images X The time points need not necessarily be consecutive scans i e the track may contain gaps Such gaps can arise for several reasons such as the temporary misidentification of the object at that position a second worm might transiently cluster or cross creating a structure not recognized as a worm or the behavior of the tracking algorithm below These time tracks are built up sequentially proceeding backwards from the last scan in the experi ment toward earlier scans Stationary tracks of worms that died early may not remain identifiable throughout the experiment since a dead animal may dissolve over time and become unrecogniz able by the SVM Thus despite proceeding backwards from a state in which all worms are dead we might nonetheless have to initiate new stationary tracks as we encounter their ending Outlining the procedure requires a bit more notation Since a worm s time track is built over time we decorate w defined above with a scan time t iew pip2 Pnli to indicate the most up to date time track as of time t lt t Remember we proceed backward in time at the next addition of a time point the current p becomes pz and the current t becomes t2 To denote the set rather than the sequence of positions belonging to worm
97. the Worm Browser Supplementary Video 6 Nature Methods doi 10 1038 nmeth 2475 12 Supplementary Note 3 Assembly 3 1 PERROTT io ised ease elated asi canatphpa cdl re ape Sin due 6684 anes A hho w eee aoe oe aie 13 3 2 Complete parts Dst ercsi eer asdendedseevans cued ereak E vee ee eeewen E 13 3 3 Recommended scanner modifications 00 ecw cece eee eee neces 16 3 3 1 Overview of the fan mounting procedure ccesscceceeserees 16 3 3 2 Modifying the scanner chassis 6 ico sis booed aly oe eee oo Hae 16 3 3 3 Modifying the transparency unit 4 60s000saseteeus sun abowwde awn 17 3 3 4 Accessing and preparing the fixed scanner lens 20000 21 3 4 Instructions for focusing a scanner 2 66sec esas ee seeds caus sev eee eeus 22 3 1 Introduction The lifespan machine identifies worm death times through the combination of four interacting components i An image acquisition system based on modified flatbed document scanners ii An image acquisition control server running under Linux on a PC iii An image analysis server that can run in the background of any x86 Windows Apple or Linux computer iv A data annotation and aggregation client program that allows rapid inspection and annota tion of automated results This document contains specifications for a reference implementation of the image acquisition system An up to date detailed specification of the software is provided as part of
98. the source code documentation These instructions are lengthy representing the authors attempt at assembling a comprehensive self sufficient document The modifications themselves are actually quite straightforward A re search assistant RA with a life science undergraduate degree and no engineering background learned how to perform all modifications over the course of several hour long training sessions The modifications required to complete two scanners capable of monitoring over a thousand in dividuals should take a novice two or three half days All modifications are easily performed on a standard lab bench The authors do not currently possess good estimates for the operating lifespan of a modified scan ner Existing devices appear to operate correctly after nearly two years of continuous use Several have been in operation for nearly five years Across a fifty scanner installation the authors have retired only three scanners as a result of malfunction usually after only a few weeks of operation Presumably scanner fluorescent bulbs will eventually require replacement 3 2 Complete parts list The following list details the equipment required for a reference 10 scanner installation capable of monitoring 5600 nematodes simultaneously Smaller or larger clusters can be easily assembled at proportional cost All prices are from circa 2012 and approximate 10x Epson v700 Perfection V700 Photo Scanners Available from Amazon com or
99. tive to death d G d G 1 254 7 1 254 8 wildtype E wildtype ps 127 daf 16 m86 ga ti daf 16 m86 22 1154 age 1 hx546 3 5 1 154 age 1 hx546 2 amp 3 22 F 2 1 14 2 f 1 14 25 1 054 EE 1054 o J o ge 1 Ss 14 ee J oe S 0 954 ZS 0 954 en mr 0 94 10 2 1012345678910 Time relative to death d Figure 7 2 Coincidence of volume dip and cessation of motion Caption on next page Nature Methods doi 10 1038 nmeth 2475 34 Figure 7 2 Cont d Coincidence of volume dip and cessation of motion a This is the same panel as in Fig 7 1c b To corroborate that these volumetric changes occur contemporaneously with the cessation of movement we plotted spline fits as in Fig 7 1 of the corresponding move ment score time series Death time identification by movement score is seen to agree with visual determination of death On average animals indeed cease to contract and begin to expand approx imately at the time of their last movement c On visual inspection animals appeared to darken slightly with size Pixel intensity of animals and animal area might be related since area is calcu lated from a segmented image whose segmentation depends on pixel intensity To demonstrate that worms exhibited a net decrease in pixel intensity independently of segmentation we mea sured the total pixel intensity within a box of constant size bounding the animal across its image time series Th
100. ts which you can safely ignore You should see a little piece of gum melted onto the lens Remove it with a screwdriver or a razor blade This should take about a minute You ll need to work around the edges but usually chunks can be pried off in big pieces Make sure to clean around the edges of the lens 10 You have freed the lens Replace scanner bar s bottoms cover and secure all the screws Make sure the ribbon cable running into the bar enters through the correct slot 11 Slide the support bar back into the scanner bar and fit everything back into its original place 12 Rotate the support bar in situ such the friction joint holds it secure 3 4 Instructions for focusing a scanner 1 Obtain a few reference plates with which to calibrate the focal plane The authors prefer to optimize focus using live worms on agar plates but can imagine a more stable reference might be ideal If live worms are used let the plates dry enough so that they will absorb some liquid Seed with 125 250 microliters bacteria and let the liquid absorb providing the animals with a comfortable dry lawn 2 Pick 20 40 animals onto each plate 3 Very Important Let the animals calm down for an hour or two at room temperature after plating to ensure that they have stopped moving quickly If calibration is attempted using startled worms all images will appear blurry regardless of focus due to animal motion It is easy to waste an hour attempting in
101. ty span and find that wildtype animals spent 1 27 0 17 days whereas age 1 hx546 mutants spent 3 0 0 15 days in this state see Fig 6 2 1 z a b 0 8 0 6 0 4 wildtype empty 3500 7 3000 n 2500 2 g 20004 Z 5 6 a 1500 fe Ee Cc E 10004 2 O 5004 04 f E a a T TTT TT T a a EE T a a i T TTTrr T i an or oe S 1 0 5 10 15 20 25 30 Time d A 0 4 daf 16 mu86 daf 2 RNAi 0 2 1 total observed 2 fast moving 0 3 posture change 4 cumulative deaths 0 5 10 15 20 25 30 Time d Figure 6 1 Population statistics of movement classes a The red curve curve 1 reports the total number of wildtype worms observed by the machine at any time in this experiment This number changes for reasons detailed in Note 8 section 8 2 and Fig 8 3 The green curve 2 is the number of individuals that crawl across the agar i e are not stationary in the sense of Note 5 section 5 1 The orange curve 3 shows the number of worms that are stationary but still change posture in the sense of Note 5 section 5 2 The blue curve 4 reports the number of animals that have died The overall plot exhibits a population dynamics of movement categories reminiscent of an irreversible chemical reaction chain in which A gt B gt C b As in panel a but for wildtype and daf 16 backgrounds with daf 2 RNAi and empty vectors Nature Methods doi 10 1038 nmeth 2475 31
102. ulations to differ in mean lifespan from wildtype by about 10 at 34 5 C In the simulation we observed sample size to have a substantial effect on power the probability of rejecting the wrong null hypothesis at a given level of confidence Fig 9 1b With a population of 110 animals every 5th test will get it wrong a power of 80 or 20 false negatives At around 200 animals power has increased to 99 In summary our arguments in this section indicate that sample sizes of sufficient power to detect with high confidence a difference in survival prospects of 15 or more are comfortably within reach of the manual method Detecting 10 differences at high confidence and power becomes more te dious on the order of 200 animals and quickly becomes a bottleneck for determining the mortality statistics associated with a larger number of environmental or genetic conditions The detection of small effect sizes also places more stringent requirements on the reproducibility of environmental conditions hardly achievable with the routine manual approach Finally biological insights often depend on showing the absence of an effect in which case very small effect sizes need to be deter mined with high confidence requiring population sizes of several hundred animals These sizes are routinely achievable with the Lifespan Machine at high throughput Nature Methods doi 10 1038 nmeth 2475 44 Supplementary Note 10 Aggregation of survival curves 10 1
103. vain to focus fast moving worms 4 If the scanner chassis lid is currently removed place it back on Do not tighten any screws because it will be necessary to enter the chassis in each step of the focusing process Do make sure that the lid is seated flush Sometimes the lid can catch on one side which will raise it slightly and affect the focus 5 Ifthe TPU is currently unattached put it on and plug it in 6 Ifthe scanner isn t plugged in to a computer and a power outlet do so 7 If the chassis and TPU fans aren t running plug them in and start them running Scanning produces heat and without operating fans plates will begin to fog over during calibration 8 Place the calibration reference plates into the scanner If plates are to be placed on glass sheets during experiments as is recommended remember to include this glass sheet into the calibration process 9 On the computer run a preview scan that captures the entire scanner surface 10 Select a small location of one plate on which to focus Smaller areas will be captured more rapidly speeding the calibration process If worms are being used try to select a small area that contains several worms The calibration process is fastest considering a single plate but this is somewhat risky as plates may vary slightly in height Ideally the focal plane should be adjusted to a consensus height that works well for most plates 11 On the computer capture an image 1
104. vironmental effects on lifespan 2 2 10 2 The device effect regression model 1 20 0 2 eee eee eee ena Supplementary Note 11 Hazard 0 00 cecciccsccscdateasdarcaseaventeaieasaenca tdaacardecoeseeraes 47 11 1 Hazard estimation and rendering 002 e eee nnenneae 47 n2 Hazard Oe kh ses sic AE ae DEA HS OR DHE ew a OG 48 i3 Hazard and ae Sie ooo oo oe ek rh A ee ee AR EME A we RE HS 51 Supplementary Note 12 Control for scanner effect ccccsecesseccecceceesesesaveceeees 52 Supplementary Note 13 Additional survival curves controls ceeeeeeeeeeeeeeees 54 Supplementary Note 14 Limitations 6 66 ccaciescessesnaseseesenen eodanensanenboneunpoenoens 58 Supplementary Note 15 Resolution and scalability ccccc cess eee eeeeeeeeeeeeeeeees 60 Supplementary Note 16 Use cases sseneseseseressesesessresessesressoseseesesssereseese 61 APNG WIE INGE esios oir ans nenene E oE E E RE AE ENEKE REAREN ANETKA 62 aane E E E E E E E E E T 62 List of Figures 1 1 Lifespan Machine hardware ccccccceceeceecescescesceeceseeseeeseesetsetseeseees 4 1 2 The effect of scanner temperature on lifespan cece cece ee cee eee eeeeeeeeeeeeees 5 1 3 Consistency of scanner surface temperature ccc cece eee eeeeeeeeeeeeeeseeseeseees 5 2 1 Registration of consecutive scanner IMAES ccccccccccccccccccccecceccecceceeees 7 2 2 Feature pa
105. x1 8 8 32 available from McMaster Carr a 100 unit box for 4 25 50x Machine Screw Nuts For example Stainless Steel Machine Screw and Hex Nuts 11 32x1 8 8 32 available from McMaster Carr a 100 unit box for 4 25 80x Nylon twine or string cut into 16 5 lengths Used for mounting fans Simpson s Individual Stringettes available from McMaster Carr 1800 feet for 8 00 1x Reasonably modern circa 2012 x86 PC or laptop capable of running Linux More details provided under Software Components section 1 3 1 of the scientific Supple mentary Information 650 10x 9x11 5 inch glass sheets Standard window pane glass available from most local hardware stores 5 per sheet 10x 8 7x11 3 inch rubber mats Cut from rubber stock for example High Strength EPDM Rubber Sheet 40A Black 36 x36 x1 16 available from McMaster Carr at 20 per foot 40x Plastic spacers to support the scanner lid 1 segments cut from 8 Polyethylene U Shaped Channels Leg Length 3 4 Base Length 5 4 Available for McMaster Carr 18 8 for eight feet 2x Gaffer s tape This is similar to duct tape but better in every way Available from McMaster Carr 10 a roll Black looks sleek pink looks totally rad 10x Wire to power fans 20 AWG gauge or lower Nature Methods doi 10 1038 nmeth 2475 4 10x Rubber mats Plates should be sealed to prevent desiccation Many methods for doing this are possible we use a design cut into

supplement - Walter Fontana Laboratory

Contents

Download Pdf Manuals

Related Search

Related Contents