Source-linked AI summary
Analysis of Forensic DNA Mixtures with Artefacts
R. G. Cowell, T. Graversen, S. Lauritzen, J. Mortera
TL;DR
Interpreting forensic DNA mixtures is difficult when samples are low-quality or contain DNA from several people. This paper models electropherogram peak heights, enabling maximum-likelihood estimation with multiple unknown contributors; the authors report sensible, robust results on challenging cases.
Problem
Low-quality crime-scene DNA and mixtures from several people make it uncertain whether a given individual's DNA is present.
Source: §1 · pp. 1–2Method
A Bayesian-network representation of a peak-height model makes maximum-likelihood estimation computationally feasible with multiple unknown contributors.
Source: §1 · p. 2Results
The authors report sensible, robust results on real challenging cases, including difficult low-template DNA mixtures.
Source: §5 · p. 26Takeaways & Limitations
Further analyses such as likelihood-ratio calculation and mixture deconvolution follow directly from the model without additional approximations, model modifications, or ad hoc heuristics.
Source: §5 · p. 26Allele frequencies are treated as fixed and known, ignoring their inherent uncertainty.
Source: §5.3 · p. 30
Abstract
from arXiv · showhide
DNA is now routinely used in criminal investigations and court cases, although DNA samples taken at crime scenes are of varying quality and therefore present challenging problems for their interpretation. We present a statistical model for the quantitative peak information obtained from an electropherogram (EPG) of a forensic DNA sample and illustrate its potential use for the analysis of criminal cases. In contrast to most previously used methods, we directly model the peak height information and incorporates important artefacts associated with the production of the EPG. Our model has a number of unknown parameters, and we show that these can be estimated by the method of maximum likelihood in the presence of multiple unknown contributors, and their approximate standard errors calculated; the computations exploit a Bayesian network representation of the model. A case example from a UK trial, as reported in the literature, is used to illustrate the efficacy and use of the model, both in finding likelihood ratios to quantify the strength of evidence, and in the deconvolution of mixtures for the purpose of finding likely profiles of one or more unknown contributors to a DNA sample. Our model is readily extended to simultaneous analysis of more than one mixture as illustrated in a case example. We show that combination of evidence from several samples may give an evidential strength close to that of a single source trace and thus modelling of peak height information provides for a potentially very efficient mixture analysis.
1 Introduction
The section introduces a simplified statistical model for interpreting crime-scene DNA mixtures, using a Bayesian-network representation to support maximum-likelihood estimation from electropherogram peak heights while accounting for uncertainty and measurement artefacts. It also defines evidential-efficiency loss in terms of the maximum posterior probability from deconvolution.
1 Introduction
- Crime-scene DNA mixtures can contain tiny, degraded, or multiple contributors’ samples, making it uncertain whether a suspect’s DNA is present and raising measurement and statistical challenges.Mixtures may include people unrelated to the crime, including in traces from objects handled by groups.Source: §1 · pp. 1–2
- The proposed peak-height model simplifies earlier models and, with an efficient Bayesian-network representation, enables maximum-likelihood estimation when several unknown contributors are posited.The simplifications include eliminating discretizations of some continuous parameters; the network representation supports fast computation.Source: §1 · p. 2
1.1.1 Short Tandem Repeat (STR) markers
- STR profiles encode each locus as an unordered allele pair, and loci chosen for forensic identification are variable between individuals and may be treated as mutually independent.The article adopts Hardy–Weinberg equilibrium, under which an individual’s two alleles are assumed to be sampled independently from the same population.Source: §1.1.1 · p. 3
1.1.2 The PCR process
- PCR amplification and electrophoresis produce EPG peaks whose heights measure amplified allele amounts, but imperfect copying and artefacts such as dropout, stutter, and dropin complicate interpretation.The model captures amplification variability using gamma distributions; low-template DNA analyses are particularly prone to dropin and dropout.Source: §1.1.2 · pp. 5–6
1.2 Motivating example
- In the UK pub-crime example reported in Gill et al. (2008), two bloodstains were mixtures of at least three people, and the profiles were consistent with DNA from victim K1, known individual K2, and defendant K3.That interpretation would require the MC18 allele 22 peak at D2 to be stutter or dropin and K2’s TH01 allele 9 to have dropped out in MC15.Source: §1.2 · pp. 6–7
1.3 Objectives of analysis
- DNA-profile analysis can quantify evidence strength between competing hypotheses or deconvolve a mixture by estimating likely genotypes of unknown contributors.Deconvolution can produce candidate genotypes for a perpetrator database search, considering probable combinations of unknown contributors.Source: §1.3 · p. 7, Weight of evidence · pp. 8–9
Weight of evidence
- The weight of evidence is reported as WoE = log10 LR, comparing prosecution and defence hypotheses that differ in whether K3 or an unknown contributor is included.The unknown contributor is assumed drawn randomly from a reference population with known allele frequencies; the defence hypothesis is comparative, not an assertion by the defence.Source: Weight of evidence · pp. 7–8
- For the case example, K3’s match probability gives a single-source weight of evidence of 14.5 bans, while ignoring markers D18 and D19 reduces it to 12.0 bans; a mixed trace cannot exceed the single-source weight of evidence.The mixed-trace weight of evidence is bounded above by −log10 of the suspect’s match probability; the loss of evidential efficiency measures the weight of evidence lost by using a mixture.Source: Weight of evidence · p. 9
- Without a specified suspect, the evidence can be assessed against the posterior-most-likely unknown person K* by replacing U with K* under the prosecution hypothesis.The prior probability π* is the probability under the defence hypothesis that a random individual has genotype K*.Source: Weight of evidence · p. 9
- The generic loss of evidential efficiency equals −log10 of the maximum posterior probability from deconvolution, and is zero for a uniquely identifying single-source trace.Source: Weight of evidence · p. 9
2 A gamma model with artefacts
The section builds a gamma model of electropherogram peak heights and extends it to account for artefacts, estimate unknown parameters, and compute mixture likelihoods. Its Bayesian-network computation is linear in the number of allele types, although exponential in the number of unknown contributors.
2 A gamma model with artefacts
- The model uses absolute peak heights and a simpler stutter description, simplifying computation for mixtures with many unknown potential contributors.Contributor-specific peak contributions are gamma distributed, with mean peak height proportional to DNA amount and variance proportional to the mean.Source: §2 · p. 10, §2.1 · pp. 10–11
2.1 Basic model
- Contributor fractions are assumed constant across markers and ordered for identifiability; allele contributions are modeled with gamma distributions.In a single heterozygous diploid trace without artefacts, μ is the mean peak height and σ the coefficient of variation.Source: §2.1 · pp. 10–11
2.2 Incorporating artefacts
- Dropout is modeled by thresholding total peak height; at small DNA amounts, the gamma model tends to give lower dropout rates than the logistic model.The stutter proportion Xa differs from the conventional stutter-to-parent peak ratio because peaks can contain both proper-allele and stutter contributions.Source: §2.2 · p. 12
- Spontaneous dropin is represented by additional unknown contributors with very small DNA amounts, and silent alleles are represented as never producing an observed peak.Dropout is treated entirely as a threshold phenomenon, leaving apparatus failure as an unmodeled alternative explanation.Source: §2.2 · p. 13
2.3 Joint likelihood function
- For each hypothesis, the full likelihood sums over possible contributor genotypes weighted by their hypothesis probabilities; unknown parameters can be estimated by numerical maximum likelihood.The likelihood calculation conditions on genotypes, contributor proportions, and model parameters, with observed peak areas independent given these quantities.Source: §2.3 · pp. 14–15
- With estimated parameters, evidence against a suspect in a mixed trace cannot be stronger than evidence from a high-quality single-source trace.The network representation enables efficient marginalization over unknown profiles, avoiding brute-force summation.Source: §2.3 · p. 15, §2.4 · p. 15
2.4 Computation
- The genotype Markov representation makes computational complexity linear in the number of allele types, while effort remains exponential in the number of unknown contributors.Binary observation nodes incorporate peak-height information, allowing likelihood evaluation by a single network propagation.Source: §2.4 · p. 16
3 Related literature
The related literature spans methods based on allele presence or absence and methods that also use peak intensities. Earlier quantitative approaches vary in artefact handling, computation, and the mixture settings they report.
3.1 Methods based on qualitative information
- Qualitative methods model allele presence or absence, with approaches incorporating combinations of dropout, contamination, stutter, dropin, and other mixture complications.Curran et al. (2005) address multiple contributors and replicate runs; Balding and Buckleton (2009) classify peaks as present, absent, or masking.Source: §3.1 · p. 17, §3 · p. 17
3.2 Methods based on quantitative information
- The literature includes gamma, normal, and log-normal peak-height models, alongside methods that derive relative peak heights or preprocess potential stutter peaks.Table 2 is captioned as reporting separate maximum-likelihood estimates for MC15 and MC18.Source: §3.2 · p. 18, Table 2 · p. 19
- Quantitative approaches use peak intensities, but reported methods differ in artefact modeling, deconvolution, and the mixture settings they analyze.Cowell et al. (2011) include stutter, dropout, and silent alleles but report increased computational complexity that limited practical use to at most four contributors; Perlin et al. (2011) report performance for 16 two-person mixtures, not mixtures with more than two people.Source: §3.2 · p. 18
4 Case analysis
The case analyses use peak-height information to assess evidential strength, deconvolve contributors, interpret artefacts, and examine a single-source profile. Joint analysis improves deconvolution and yields strong evidential weight, while interpretation of apparent silent alleles has a marker-specific caveat.
4 Case analysis
- The case analysis uses a threshold of 50 for both traces, and its likelihood computations take a few minutes per trace on a standard laptop.The computational effort grows exponentially with the number of unknown contributors and linearly with the number of alleles and markers.Source: §4 · p. 19
4.1 Weight of evidence
- 14.1 bans for the combined traces is close to the 14.5-ban perfect-match single-source value; peak heights give 12.7 bans versus 10.0 from peak presence or absence.Under the combined analysis, the estimated weight of evidence is similar to a perfect single-source match; the peak-height comparison uses equal contributor fractions and matched frequency adjustments.Source: §4.1 · pp. 21–22
4.2 Mixture deconvolution
- Combining MC15 and MC18 places the defendant’s profile first among the five most probable unknown-contributor profiles, with 73% probability that the true profile is among them.For single-trace deconvolutions, that five-profile probability falls to 1% for MC15 and 39% for MC18.Source: §4.2 · pp. 22–23
4.3 Interpreting artefacts
- Under the joint defence analysis, the D2 allele 22 peak has 93% probability of being stutter, while TH01 allele 9 has dropped out.TH01 allele 6 has a 33% probability of being present despite no observed peak; its reference-population frequency is 23%.Source: §4.3 · p. 23
4.4 Single source analysis
- In the single-source profile, CSF1PO and FGA peaks are about 60% and 56% of D8’s height, respectively, and have higher posterior probabilities of a silent allele.The analysis also finds almost zero silent-allele probability at D8, where the peak is relatively high.Source: §4.4 · p. 25
- The silent-allele interpretation for FGA and CSF1PO needs caution because these end-panel markers often have lower peak heights for longer fragments.The authors state that this may be crucial for identifying silent alleles, even if evidence calculations and deconvolutions may be less sensitive to it.Source: §4.4 · pp. 25–26
5 Discussion
The authors present the model as a versatile tool for mixture analysis, with computations exact apart from deterministic optimization and no subjective allele-calling preprocessing. Their discussion highlights contributor-count sensitivity, possible model extensions, and validation and uncertainty issues that remain.
5 Discussion
- The authors report sensible, robust results on challenging real cases, including difficult LTDNA mixtures, and consider the methodology sufficiently developed for casework.They also describe the computations as exact apart from deterministic numerical optimization and note that the method avoids subjective allele calling.Source: §5 · p. 26
5.1 Number of contributors
- For eight contributors, the WoE falls only to 14 bans from 14.1 for three; even allowing eight for the defence and three for the prosecution gives 13.7 bans.The likelihood increases with contributor count, but more for the defence, so the likelihood ratio decreases.Source: §5.1 · p. 27, Figure 8 · p. 28
- The five smallest unknown contributors supplied less than 3.7% of the DNA and mainly explain minor peak imbalances, seeming irrelevant to mixture interpretation.Their alleles were effectively unobserved, apart from alleles shared with major contributors or in stutter positions.Source: §5.1 · pp. 27–28
5.2 Extensions and modifications
- The simplest model assumes marker-invariant peak parameters; proposed extensions include marker dependence, fragment-length adjustment for degradation, and alternative dropout modelling.The authors note that marker-specific parameters could greatly increase parameter count and that adding parameters increases estimation complexity.Source: §5.2 · p. 29
5.3 Other issues
- Model validation matters even when comparing competing hypotheses; the conditional probability transform for MC15 and MC18 indicates an excellent gamma-model fit.The authors describe graphical validation using internal predictions and residual analysis as promising and under development.Source: §5.3 · pp. 29–30, Figure 9 · p. 30
- Treating allele frequencies as fixed ignores their sampling uncertainty, and ignoring dependence from IBD or population heterogeneity may overstate WoE.The authors also have not incorporated uncertainty in other estimated parameters, although they report approximate standard errors based on the log-likelihood Hessian.Source: §5.3 · p. 30