410 likes | 1.14k Views
The technical artifacts of forensic STR testing: stutter, pull-up, spikes, blobs, and degradation. Jason Gilder Forensic Bioinformatics August 21, 2004. Stutter. By-product of PCR amplification process.
E N D
The technical artifacts of forensic STR testing: stutter, pull-up, spikes, blobs, and degradation Jason Gilder Forensic Bioinformatics August 21, 2004
Stutter • By-product of PCR amplification process
“The general guideline for stutter identification of one repeat unit less than the corresponding allele and less than 15% of the allele’s peak…….is used to mark suspected stutter products.” John M. Butler Academic press, 2001.
AND FURTHERMORE…… “Stutter products that are larger in size by one repeat unit than the corresponding alleles have not been observed in commonly used tetranucleotide repeat STR loci.” John M. Butler Academic press, 2001.
Study statistics • Evaluated 224 different single source unsaturated reference samples from 52 separate cases from routine forensic casework. • Samples from Profiler Plus® and Cofiler® test kits • Used only single injections – last injection for multiples • More than 15 different crime laboratories. • 37 observed instances of n+4 stutter.
The “prevalence” of n+4 stutter • 12.5% of all reference samples evaluated exhibit n+4 stutter (28/224). • 42.3% of all cases exhibit at least one instance of n+4 sutter(23/52). • 938 RFUs = Minimum primary peak height giving rise to an n+4 peak (≥ 50 RFUs). • 2,185 RFUs = Average height of primary peak = (±888).
Pull-up (software differences) Classic Advanced
Non-degraded/inhibited sample • DNA donor contributes an approximately equivalent amount of DNA across all loci • Resulting peaks are roughly equivalent in height
Degradation SMALL LARGE • When biological samples are exposed to adverse environmental conditions, they can become degraded • Warm, moist, sunlight, time • Degradation breaks the DNA at random sites • Larger amplified regions are affected first • Classic ‘ski-slope’ electropherogram • Degradation is unusual. • No standards for detection
Data set • Positive Control is ideal non-degraded sample • 165 - 9947A positive controls • Profiler Plus loci • Saturated samples (>4,500 RFUs) discarded • Samples with peaks < 200 RFUs discarded
Linear regression • Calculate slope of best-fit line through linear regression
Slope distributions • All three dye distributions appear to be normal • Can set cutoffs at α = 0.05 and α = 0.01
Normalized sum • Normalize each slope to be on a scale between 0 and 1 • mnorm = (m – mmin)/(mmax – mmin) • mmin is minimum slope in dye • mmax is maximum slope in dye • Add each dye’s slope to get normalized sum
Normalized sum distribution • Normalized sum has a normal distribution • Using a threshold of significance of α = 0.05: 0.0777*Bslope + 0.0997*Gslope + 0.1109*Yslope + 2.392 < 0.8
Results • Blue and Yellow have similar thresholds • Degradation/inhibition threshold in green lower
Summary • Objective method for identifying degradation/inhibition • No longer subject to expertise • Labs can perform own study or use ours
Objective identification of artifacts in forensic STR-PCR DNA profiles using a back-propagation neural network
Identifying testing artifacts in DNA data • Urea crystals, air bubbles, dirt, dye blobs, contamination, etc. can cause bad results • Classify peaks as “good” or “bad” • Peak morphology
Spikes and blobs • 5 Cases, 36 Samples, 780 Peaks (572 Good, 208 Bad) Peak Area Peak Height Blob: Peak Area / Peak Height > 10 + Spike: Peak Area / Peak Height < 4.5 -
Input Signals Synaptic Weights x0 = + 1 wk0 = bk Activation Function x1 wk1 nk f(*) Output yk x2 wk2 S . . . . . . xm wkm Neural Network • Data features are inputs (peak height, area, etc.) • Each feature is multiplied by a weight • Weights are adjusted • Neurons connected to form a network
Peak features • Back propagation • Peak height, area, width, PH / PA ratio, • Peak height at ¼ peak width, • Slope to ¼ peak width, • Slope to ½ peak width
Initial results • 6,045 peaks (2,900 “good” and 3145 “bad”) • 99% training, 99% testing • Removed features until only Peak Height left • 98.5% training, 98.8% testing • Effective peak height threshold of 248 RFUs
Expanding the dataset and choosing different features • Limit features to those not involving peak height • peak height, peak area, and peak height at ¼ peak width removed • Expand dataset to 16,212 peaks • 8,532 “good” peaks • 7,680 “bad” peaks • More “good” peaks at lower levels
Final neural network • Features: PA/PH, width, slope to ¼ width, slope to ½ width • 8022 / 8106 (99%) for training • 7997 / 8106 (98.7%) for testing • Removed peak height bias
9947A positive control • Many homozygotes • Are peaks still additive?
Additivity study • y = 1.865x + 208.5 => Hom = 1.9 * Het • R2 = 0.845
Track pants ? Bib Adventitious Match Condom from other case
Leskie inquest • Bib and track pants declared degraded • Condom declared non-degraded • Possible contamination from condom • Is condom degraded?
Leskie condom Average Pos Control Condom
Condom and pos control • Pos control tested at the same time as the condom • Condom well outside degradation thresholds (past 99.5% cutoff)
Data points • X point is size (in bp) • Y point is peak height (in RFUs) • Heterozygotes: 1 data point each • Homozygotes: divide height by half to get 2 points • Assumption: peaks are additive • 6 data points per dye