I have read the journal’s policy and have the following conflicts. Mark Perlin is a shareholder, officer and employee of Cybergenetics in Pittsburgh, PA, a company that develops genetic technology for computer interpretation of DNA evidence. Cybergenetics manufactures the patented TrueAllele® Casework system, and provides expert testimony about DNA case results. Kiersten Dormer and Jennifer Hornyak are current or former employees of Cybergenetics. Lisa Schiermeier-Wood and Dr. Susan Greenspoon are current employees of the Virginia Department of Forensic Science, a government laboratory that provides expert DNA testimony in criminal cases and is adopting the TrueAllele Casework system. This does not alter our adherence to all the PLOS ONE policies on sharing data and materials.
Conceived and designed the experiments: MWP LSW. Performed the experiments: KD JH. Analyzed the data: MWP SG. Contributed reagents/materials/analysis tools: MWP LSW. Wrote the paper: MWP KD SG.
Mixtures are a commonly encountered form of biological evidence that contain DNA from two or more contributors. Laboratory analysis of mixtures produces data signals that usually cannot be separated into distinct contributor genotypes. Computer modeling can resolve the genotypes up to probability, reflecting the uncertainty inherent in the data. Human analysts address the problem by simplifying the quantitative data in a threshold process that discards considerable identification information. Elevated stochastic threshold levels potentially discard more information. This study examines three different mixture interpretation methods. In 72 criminal cases, 111 genotype comparisons were made between 92 mixture items and relevant reference samples. TrueAllele computer modeling was done on all the evidence samples, and documented in DNA match reports that were provided as evidence for each case. Threshold-based Combined Probability of Inclusion (CPI) and stochastically modified CPI (mCPI) analyses were performed as well. TrueAllele’s identification information in 101 positive matches was used to assess the reliability of its modeling approach. Comparison was made with 81 CPI and 53 mCPI DNA match statistics that were manually derived from the same data. There were statistically significant differences between the DNA interpretation methods. TrueAllele gave an average match statistic of 113 billion, CPI averaged 6.68 million, and mCPI averaged 140. The computer was highly specific, with a false positive rate under 0.005%. The modeling approach was precise, having a factor of two within-group standard deviation. TrueAllele accuracy was indicated by having uniformly distributed match statistics over the data set. The computer could make genotype comparisons that were impossible or impractical using manual methods. TrueAllele computer interpretation of DNA mixture evidence is sensitive, specific, precise, accurate and more informative than manual interpretation alternatives. It can determine DNA match statistics when threshold-based methods cannot. Improved forensic science computation can affect criminal cases by providing reliable scientific evidence.
DNA analysis is the forensic gold standard in human identification
With increased societal expectations
Human analysts may simplify short tandem repeat (STR)
An “analytical” threshold helps human analysts distinguish between allelic data peaks and baseline instrument noise. The Combined Probability of Inclusion (CPI) mixture interpretation method first applies this analytical threshold to decide which peaks at a locus are sufficiently tall to be considered alleles. If a reference individual’s alleles are included in this set of mixture alleles, then CPI uses all the alleles in the mixture set to calculate a match statistic (the inclusion probability) as the square of the sum of the allele frequencies. (Allele determination can be viewed as a separate human interpretation step that precedes the CPI statistical calculation step. For clarity in this paper, we consider the entire data analysis procedure to comprise the CPI interpretation method). The method does not make assumptions about the number of contributors.
There is naturally occurring random variation in the polymerase chain reaction (PCR)
In 2010, the United States Scientific Working Group on DNA Analysis Methods (SWGDAM) published guidelines to help resolve such mixture genotyping issues
The Virginia Department of Forensic Science (DFS) implemented the new SWGDAM mixture interpretation guidelines, and reviewed their DNA evidence using stochastic thresholds. In 2011, DFS identified 375 criminal cases in which their stochastic threshold method had produced an inconclusive result or a less informative match statistic
Mathematical modeling can account for quantitative STR data patterns
DFS pursued a probabilistic genotyping approach for their DNA mixture evidence. They arranged for Cybergenetics (Pittsburgh, PA) to apply their validated TrueAllele Casework system to DNA mixture evidence in 144 cases. Cybergenetics produced DNA match reports on 92 evidence items in 72 cases. This is the largest data set on which case reports have been generated for probabilistic genotyping of DNA mixture evidence.
This study describes the results of computer-based probabilistic genotyping mixture interpretation on 101 reported matches, out of 111 genotype comparisons. (A DNA match is defined here operationally as a comparison between an evidence and reference genotype, relative to a population, that gives a reproducible positive match statistic). The 10 comparisons that did not produce a match are also characterized. The study compares the computer’s information yield with two methods of manual interpretation on the same evidence items. Previous TrueAllele Casework validation studies have been published on samples of known composition
The present study compares three interpretation methods for analyzing DNA mixtures from actual casework, specifically, the automated TrueAllele Casework computer system, with the traditional CPI and updated mCPI manual threshold-based methods
TrueAllele | CPI | mCPI | ||
quantitative | qualitative | qualitative | ||
continuous | binary | ternary | ||
used | ||||
used | ||||
analytical | analytical and stochastic | |||
probability model | data above analyticalthreshold | data above analyticalthreshold | ||
allele pairs | alleles | alleles | ||
automated | manual | manual | ||
statistical | alleles | alleles | ||
assumed | ||||
with genotype | with alleles | with alleles | ||
all | inclusion | stochastic inclusion | ||
likelihood ratio | inclusion probability | inclusion probability | ||
include, exclude orinconclusive | include | include | ||
information | inclusion | inclusion |
The paper begins by describing three methods of DNA mixture interpretation, one automated and two done manually. We then present the case materials used and the interpretation procedures employed. We examine the TrueAllele automation method’s reliability, using its inferred match statistics to assess how sensitive, specific, precise and accurate it is. (In Forensic Science, “sensitive” and “specific” describe the reliability of analytical instrumentation. With DNA interpretation methods, they can similarly describe the respective degree of positive or negative identification). We compare how well TrueAllele, CPI and mCPI preserve identification information relative to one another, as measured by their match statistics. We conclude by observing that computer-based DNA mixture interpretation can provide an improvement over current manual forensic processes.
The DNA samples used in this study were lawfully obtained by DFS in accordance with Virginia code Section 9.1–1101. All personal identifiers were removed from DNA data prior to computer interpretation. The submitted scientific manuscript contains only summary statistics, and discloses no personal or case information.
The DNA mixture interpretation process begins with electronic data signals. These signals are examined to form genotypes. Comparison of an evidence genotype with a reference genotype, relative to a population, can then produce a DNA match statistic.
A STR locus is a length polymorphism, where alleles have different numbers of short DNA units (typically four or five base pairs) that are repeated in tandem
Following DNA extraction and quantification, STR analysis proceeds in two steps. First, PCR amplification with a set of fluorescently labeled primers creates millions of allele copies from the DNA template. Random variation in a 31 cycle PCR process
Penta E is one of 15 STR loci in the Promega PowerPlex 16 multiplex kit
Quantitative DNA mixture data are shown at the Penta E STR locus. The x-axis measures allele fragment size (bp), and the y-axis measures DNA quantity (RFU); a boxed peak number denotes allele length. The two contributor mixture is formed from a 7,14 major genotype and a 10,12 minor genotype. The result is a pattern of peak heights that reflect the underlying genotypes.
STR mixture data can be interpreted in different ways, giving rise to different DNA match statistics.
TrueAllele infers genotypes with a probability model that uses a computer to automatically propose peak patterns, and assess how well they explain the quantitative data (
TrueAllele’s inferred (probabilistic) genotypes can be entered into standard formulae to calculate a likelihood ratio (LR) (
Many variables are considered in genotype modeling, such as the genotypes of each contributor at every locus, the mixture weights (that sum to 1) of the contributors, the DNA template mass, PCR stutter, relative amplification, DNA degradation and the uncertainties of all these variables. A likelihood function assesses how well particular values of these variables explain the observed quantitative STR data peaks, determining the probability of the (fixed) data conditioned on the (changing) variable values.
With DNA mixture data vector
We can visually understand this likelihood function as constructing a pattern of allele heights that can be compared with the peak height data.
Linear combinations of genotype allele pairs can explain the observed quantitative mixture data. Here, a major 7,14 contributor (blue bars) having twice the DNA as a minor 10,12 contributor (green bars) explains the data well, with a high likelihood value. Alternative genotype choices or combinations would not explain the data as well, and thus have lower likelihood.
The posterior genotype probability is proportional to the likelihood value times the prior population probability
Bayes theorem
The modeling approach is objective in the sense that only evidence data is used to infer genotypes, without any knowledge of a reference comparison genotype. Proceeding
Allele Pair | TrueAllele | CPI | mCPI | ||||||||
prior | likelihood | posterior | LR | likelihood | posterior | LR | likelihood | posterior | LR | ||
7 | 7 | 4.3% | 1 | 17% | 1 | 67% | |||||
7 | 10 | 3.3% | 2 | 1 | 13% | ||||||
7 | 12 | 7.1% | 2 | 1% | 1 | 28% | |||||
7 | 14 | 1.9% | 1 | 8% | 1 | 30% | |||||
10 | 10 | 0.6% | 1 | 1 | 2% | ||||||
10 | 14 | 0.7% | 1 | 3% | |||||||
12 | 12 | 2.9% | 8 | 1% | 1 | 11% | |||||
12 | 14 | 1.6% | 1 | 1 | 6% | ||||||
14 | 14 | 0.2% | 1 | 1% | 1 | 3% | |||||
Total | 100% | 100% | 100% |
Inclusion methods of DNA mixture interpretation begin by applying an analytical threshold to the quantitative STR peak data. The Virginia DFS analytical thresholds are specific to each fluorescent dye channel: 73 RFU (blue dye), 84 RFU (green), 75 RFU (yellow) and 52 RFU (red). Peaks above the threshold are designated as “allele” events, while those below are not used (
The purpose of this threshold is to distinguish allelic signal from background noise. Applying the threshold (red line) reduces the quantitative peaks to all-or-none putative allele events (blue bars). The analytical threshold operation eliminates individual peak heights, as well as their collective pattern.
The inclusion likelihood function assigns 1 to all allele pairs included in the allele list, and 0 otherwise. This CPI likelihood also assumes that all alleles from each contributor are present. With four allele events, for example, there are ten possible allele pairs (
The inclusion approach disperses probability over (in this example) ten genotype values. Many of these allele pairs (e.g., 7,7) are not compatible with a minor contributor genotype, based on the peak height data shown. Since the total probability is 1, diverting genotype probability away to infeasible solutions reduces the probability at more likely solutions, and thereby lowers match strength. Starting from highly informative STR data, CPI may reduce considerably the reported identification information, or even eliminate it entirely by dismissing an evidence item as “inconclusive”. Inclusion protocols are susceptible to examination bias, since a reference genotype could be considered (e.g., to assess potential allelic dropout) when determining whether to use a locus in a CPI statistical calculation
Replicate STR experiments exhibit natural variation in peak height, as described by probability model data variance parameters
The higher mCPI stochastic threshold can make less use of the STR data. In our Penta E mixture example, the 5s injection peak heights of alleles 10 and 12 now fall below the stochastic threshold of 320 RFU (
A higher threshold level (red line) is used in manual review to address random peak variation by differentiating more certain (blue bars) from less certain peaks. The stochastic threshold removes more STR loci from statistical consideration, which makes less use of the available data.
The likelihood ratio is a standard DNA match statistic
There are several ways to calculate a LR match result, all of which produce the same number
When a genotype likelihood function accounts for observed quantitative evidence data, a reproducibly inferred LR number can accurately summarize the extent of match between that evidence and a reference. A positive log(LR) provides a weight of evidence supporting a match, a negative log(LR) does not favor a match, while a log(LR) near zero is inconclusive. The LR value is always scientifically meaningful. Scientists sometimes verbally describe a LR using an arbitrary subjective scale
The
With
Using a
A method’s posterior genotype probability is the product of its likelihood and the prior, normalized to sum to unity (
The LR is shown (
The CPI allele inclusion method uses all data peaks above a predetermined analytical threshold to form allele pairs
mCPI uses a stochastic threshold to produce a DNA match statistic. mCPI only uses those loci at which all of the peaks are above the stochastic threshold. In our Penta E locus example, the data peaks corresponding to the known 10,12 individual are both under the threshold, setting the mCPI likelihood to zero. Therefore the mCPI posterior probability and LR (from the calculation 0/2.7) of the locus would both be zero, as well. This locus was not used in the mCPI calculation.
Considering all loci in this mixture, TrueAllele’s log(LR) was 16.32; the weight of evidence was 7.04 ban for CPI, and 6.00 ban for mCPI. This example illustrates how genotype modeling makes more use of the data to preserve DNA match information, while an already diminished CPI match statistic can be further reduced by the mCPI stochastic threshold. Our study examines this phenomenon on a larger set of Virginia DFS case matches, comparing the three mixture interpretation methods TrueAllele, CPI and mCPI.
The Virginia DFS identified DNA mixture cases where computer interpretation could potentially make more use of the STR data than manual review. The selection criteria included having a probative DNA item, possible use of that item as evidence in a criminal trial, an included person of interest, and a need for accurate DNA match information. Items that were easy to interpret manually were not chosen.
The 72 cases spanned a full range of biological evidence, including touch, epithelial cells, blood, saliva and semen (
Sample type | Count |
blood | 10 |
epithelial/skin | 30 |
fingernails | 2 |
hair | 1 |
saliva | 4 |
semen | 3 |
stain | 1 |
touch | 41 |
Contributors | Items | |
2 | 40 | |
3 | 65 | |
4 | 8 | |
2 or 3 | 16 | |
3 or 4 | 3 | |
2, 3 or 4 | 1 |
The mixture weights as calculated by TrueAllele were evenly distributed between 10% and 90% (
Mixture Weight | Count |
0.05 | 3 |
0.15 | 13 |
0.25 | 5 |
0.35 | 12 |
0.45 | 18 |
0.55 | 12 |
0.65 | 11 |
0.75 | 12 |
0.85 | 12 |
0.95 | 4 |
Virginia DFS generated STR data using the Promega PowerPlex 16 kit (Madison, WI), analyzed on an Applied Biosystems 3130×l Genetic Analyzer (Foster City, CA). The DFS case materials included electronic.fsa data files from the sequencer, their own case reports and case context descriptions. DFS sent these electronic materials via secure file transfer protocol (sFTP) to Cybergenetics during the latter half of 2011. The data files were organized in batches for computer processing.
Cybergenetics TrueAllele Casework is a computer system for resolving DNA mixtures into their component genotypes
TrueAllele divides DNA identification into two phases
For each received batch of cases, Cybergenetics processed the.fsa files in the TrueAllele Analyze module to assess data quality. For computing efficiency, EPG peaks below the baseline noise level of 10 RFU were not used (since they do not affect the results). The quality-checked quantitative peak data were then uploaded to a TrueAllele database.
A trained first TrueAllele operator processed a case by downloading from the database the electronic data for all evidence items. The operator examined the EPG signals, and estimated the number of contributors for each evidence item based on the number of peaks observed at each locus. If relevant and available, an assumed reference could be used. (For example, with an intimate sample from a sexual assault, assuming the victim’s genotype as a known contributor to the mixture is forensically meaningful). Appropriate DNA interpretation case questions were uploaded as “requests” from the VUIer to the TrueAllele database for processing.
Following this initial processing, an experienced second TrueAllele operator then reviewed the computer results, and determined whether further analysis would be required. Such additional TrueAllele analyses could entail assuming a different number of contributors, considering DNA degradation, or repeating the question using more computer processing time. When the number of contributors was ambiguous, multiple contributor assumptions were tested; the assumed number of contributors (when there are enough) does not have a major effect on the inferred genotypes or match statistics. Reportable DNA results were replicated in two or more independent computer runs.
A reporting scientist examined all the computer results in a case. After careful review of the replicated genotypes, together with the data and mixture weights, a concordant genotype subset was identified. Concordant genotypes had similar probability distributions, mixture weights and Kullback-Leibler (KL) statistics
In the VUIer program, the TrueAllele scientist indicated the three genotypes (evidence, reference and population) needed to calculate a LR. Virginia’s databases of Black, Caucasian and Hispanic populations were used, and the co-ancestry coefficient was set at 1%. All three LRs were reported; for comparison purposes in this study, we conservatively took the smallest of the three match statistics.
Cybergenetics processed the data, and prepared DNA match reports for 72 cases requested by DFS. These cases encompassed 92 items of evidence and 111 comparisons to reference individuals. TrueAllele LR values were collated from these reports. DFS had independently conducted manual mixture calculations on most of the reported TrueAllele matches. These CPI and mCPI match statistics were collected and recorded in a LR format.
A DFS forensic examiner assessed DNA evidence to determine whether a person of interest could be eliminated from the data. This assessment considered the number of contributors, sample type, DNA quantity, potential drop out, and other factors. When the data were inconclusive or the person had been eliminated, no match statistic was calculated.
We assessed the reliability of DNA mixture interpretation methods through information metrics based on log(LR). The data set comprised 111 computer-inferred evidence genotypes and match statistics; our focus is on the 101 reported matches. We consider in turn how specifically, precisely and sensitively the TrueAllele system performs. Information comparisons with CPI and mCPI methods are possible because formally these manual methods are LRs
Recall that the identification hypothesis H is that a particular individual contributed their DNA to biological evidence. The alternative hypothesis ∼H is that they did not, i.e., that the DNA was left by someone else. Forensic science standardly approximates ∼H with a random man hypothesis that the DNA contributor is an unrelated person selected at random from a genotype population
Specificity measures the extent to which a mixture interpretation method does not misidentify the wrong person. Since identification information is expressed through the log(LR), let X be a real-valued random variable of log(LR) values. We want to understand the TrueAllele distribution of Pr{X = x | ∼H}, the information X conditioned on randomly selected genotypes (that are not contributors to the mixture). The specificity statistic Pr{X>0 | ∼H} then tells us how frequently a positive log(LR) occurs by chance.
Toward this end, we generated ten thousand random genotypes from each of the three Virginia ethnic populations. This generation was done by randomly selecting alleles in proportion to their prevalence in the population database. We compared the 101 matching TrueAllele-inferred evidence genotypes to these random reference genotypes, relative to the appropriate population, to calculate log(LR) values; the co-ancestry coefficient was set to 1%. These values provided a representative log(LR) sampling of over a million nonmatching comparisons for each population.
The resulting empirical Pr{X = x | ∼H} distribution is shown in
A histogram shows empirical log(LR) distributions for 101 evidence genotype comparisons relative to 10,000 randomly generated references. There are 1,010,000 data points for each of the three ethnic populations. Note that the negative values are located far to the left of zero.
n = 3,030,000 | Black | Caucasian | Hispanic |
Minimum | −30.000 | −30.000 | −30.000 |
Mean | −19.467 | −19.217 | −19.547 |
Maximum | 2.381 | 2.726 | 3.782 |
Standard deviation | 6.543 | 6.723 | 6.637 |
0 | 39 | 32 | 29 |
1 | 8 | 11 | 9 |
2 | 2 | 1 | 1 |
3 | 0 | 0 | 1 |
log(LR) >0 | 49 | 44 | 40 |
The specificity value Pr{X>0 | ∼H} was estimated by counting the fraction of positive log(LR) outcomes. For all three ethnic populations, TrueAllele’s false positive rate was less than one in twenty thousand (
TrueAllele’s genotype model has hundreds of variables. Therefore the (largely continuous) probability model cannot be solved directly by brute force integration or enumeration. Instead, MCMC computing is used to statistically sample from the joint posterior probability distribution, a standard numerical solution for high-dimensional hierarchical models. Such methods exhibit sampling variation between independent computer runs.
Precision describes a method’s reproducibility on the same data. To measure precision, we examined the identification information obtained in duplicate computer runs of the 101 matching genotypes. The observed log(LR) pairs are shown in
The scatterplot shows log(LR) values for 101 duplicate computer runs on the same evidence. Each point gives the first (x) and second (y) values. The data lie close to the y = x diagonal, which represents exactly replicated results.
The log(LR) variation between computer runs was generally greater at medium LR values having logarithms between 5 and 10 (
Sensitivity measures the extent to which a mixture interpretation method identifies the correct person. We therefore examine Pr{X = x | H}, the log(LR) distribution conditioned on the identification hypothesis H. In this observational case study, we want reassurance that H is true, so that the reference genotype actually contributed to the mixture evidence.
The preceding specificity results demonstrated that the false positive rate Pr{X>0 | ∼H} of TrueAllele’s mixture interpretation under the noncontributor hypothesis was less than 0.005%. (This is ten times smaller than the highly reliable 0.05% error rate for dual manual review of easily interpreted single-source reference samples
The TrueAllele log(LR) distribution of the 101 reported matches is shown in
Three histograms show the empirical log(LR) distribution for different mixture interpretation methods on the case data. Frequency distribution (
TrueAllele | CPI | mCPI | |
Minimum | 1.255 | 0.778 | 0.301 |
Median | 10.550 | 6.681 | 1.857 |
Mean | 11.054 | 6.825 | 2.145 |
Maximum | 22.962 | 16.724 | 6.447 |
Standard deviation | 5.421 | 2.217 | 1.675 |
N = | 111 | 81 | 70 |
Inclusion (≥0) | 101 | 81 | 53 |
Persuasive (≥6) | 82 | 54 | 2 |
Inconclusive | 17 |
More accurate genotype modeling employs a likelihood function that better explains the data, and so tends to produce a higher LR (relative to less accurate modeling) when there is a true match. However, the actual LR value depends on the genotype model, thus some other measure of accuracy is needed. Over a large ensemble of DNA mixtures having randomly distributed mixture weights (
A uniform probability density function (PDF) is a constant horizontal line. TrueAllele’s empirical PDF appears relatively constant across its range of observed log(LR) values (
Cumulative empirical log(LR) distributions are shown for uniform probability (black), and for each of the three mixture interpretation methods TrueAllele (blue), CPI (green) and mCPI (red). TrueAllele tracks a uniform distribution over a wide information range, whereas CPI and mCPI do not.
The two threshold-based manual methods produced less informative DNA statistics that were distributed differently than the computer’s 101 genotype modeling positive log(LR) results. On 81 comparisons, the CPI manual method yielded matches with a mean log(CPI) value of 6.825 ban (
We can again use the Kolmogorov-Smirnov statistic to test the accuracy of these two manual methods. The empirical CDFs of inferred log(LR) values for both CPI (
The numerical differences in average log(LR) between the three interpretation methods were statistically significant (
N = 81 | TA | CPI | TA – CPI | test | p-value |
11.623 | 6.825 | 4.798 | t = 8.396 | 1.350×10^{–12} | |
10.816 | 6.681 | 4.135 | W = 3047 | 6.664×10^{–11} | |
r = 0.2999 | |||||
r^{2} = 0.0900 |
N = 53 | TA | mCPI | TA – mCPI | test | p-value |
12.883 | 2.145 | 10.738 | t = 15.147 | 1.040×10^{–20} | |
12.537 | 1.857 | 10.679 | W = 1431 | 2.386×10^{–10} | |
r = 0.2945 | |||||
r^{2} = 0.0867 |
N = 52 | CPI | mCPI | CPI – mCPI | test | p-value |
7.069 | 2.180 | 4.889 | t = 17.417 | 4.082×10^{–23} | |
6.720 | 2.024 | 4.696 | W = 1378 | 3.497×10^{–10} | |
r = 0.5188 | |||||
r^{2} = 0.2692 |
The TrueAllele genotype modeling method showed a significant improvement over the older CPI allele inclusion method (
TrueAllele also showed a significant improvement over the newer mCPI allele inclusion approach (
Switching from allele inclusion to stochastic thresholds significantly reduced the match statistic (
These concordant multi-method match results increase confidence in the sensitivity experimental design, where reference genotypes were considered to be present in their respective DNA mixture items. Each of the three interpretation methods works differently, is accepted by courts as reliable criminal evidence, and was calculated independently in the study. In each pairwise comparison, the methods independently agreed on all matches (N>50) and gave positive identification information. These pairwise consensus results were obtained on highly reliable data subsets of the more readily interpretable mixtures.
Out of 111 TrueAllele genotype comparisons, 10 gave a negative log(LR) value, and so did not produce a positive match result (
Interpretation Method | Data Observations | |||||||
TrueAllele | CPI | mCPI | allele dropout | allele overlap | low peaks | peak imbalance | infeasible mixture | infeasible pattern |
−10.64 | 3 | 4 | 1 | 1 | ||||
−6.52 | 4 | 3 | 1 | 1 | ||||
−5.05 | 4 | 3 | 1 | 1 | 1 | |||
−4.87 | 3 | 1 | 1 | 1 | ||||
−4.86 | 3.48 | 4 | 1 | 1 | ||||
−3.22 | 6.04 | 6.34 | 2 | 1 | 1 | |||
−2.99 | 4.23 | 2 | 1 | 1 | 1 | |||
−2.18 | 2 | 1 | 1 | |||||
−1.41 | 4.08 | 1 | 1 | 1 | ||||
−0.67 | 2.95 | 0.60 | 1 | 2 | 1 | |||
1.26 | 3.96 | 1 | 4 | 1 | ||||
1.76 | 1 | 1 | 1 | |||||
2.01 | 2 | 8 | 1 | 1 | ||||
2.71 | 2 | 1 | ||||||
2.94 | 8 | 1 |
There were 5 genotype comparisons where CPI indicated a match, but the computer found no statistical support (
Five genotype comparisons gave a small positive TrueAllele log(LR) value of under 3 ban (
Modern criminal justice requires rapid and reliable processing of DNA evidence. Reliability is the basis of admissible evidence, and entails sensitivity, specificity and precision. However, when confronted with complex mixtures or touch DNA, manual review can become a challenging task. Since such mixtures may constitute the bulk of biological evidence found in serious crimes such as sexual assault or homicide, effective interpretation of these data is needed.
In this casework study, the newly adopted mCPI stochastic threshold method produced results in 53 of 70 DNA match comparisons, finding an average match statistic of 140 (
TrueAllele mixture interpretation does not always increase a DNA match statistic. In this study, the computer’s statistics were lower than the corresponding human CPI values in 15 reported matches
In addition to increased average sensitivity, TrueAllele also maintains excellent specificity. The computer’s LR can quantify negative match information, unlike manual interpretation methods that are restricted to positive (logarithmic) values. We examined several million genotype comparisons between computer-inferred evidence and randomly generated references, and found a false positive rate of under 0.005%. The negative match information in this simulation experiment had a log(LR) averaging around −19.5 ban.
TrueAllele calculates DNA match statistics with precision. Replicate computer runs on the same evidence data showed a within-group log(LR) standard deviation of 0.305 ban. Thus, independent runs on the same evidence item gave statistically similar DNA match statistics that were usually within an order of magnitude.
To assess accuracy, (logarithmic) match statistic distributions were examined. Across an entire ensemble of matches, the random sampling inherent in casework should produce a uniform log(LR) distribution. TrueAllele’s match distribution was statistically uniform, which lends support to the overall accuracy of its LR values. However, the manual methods each clustered around their average match value, and so did not exhibit a uniform distribution that would support their accuracy.
DNA, whether single source or complex mixture, can provide evidence that implicates criminals and exonerates the innocent. Current manual review of DNA mixture data applies thresholds that can discard valuable data and understate the evidential import of the identification information. As demonstrated in this casework comparison study, TrueAllele computer interpretation more effectively preserves DNA evidence and match information, relative to CPI and mCPI methods that use thresholds. Both prosecutors and defense attorneys may benefit from use of this validated computer technology to review complex DNA mixture evidence.
The authors would like to thank William Allan, Meredith Clarke and Matthew Legler of Cybergenetics for data analysis that was instrumental in conducting this study. They also appreciate the insightful comments from Farrel Buchinsky, Mark Fichman, Saul Shiffman and anonymous reviewers who helped improve the quality of the manuscript.