Probabilistic Genotyping in Forensic DNA: Principles, Software, Validation and Limitations

Probabilistic genotyping (PG) has changed how forensic laboratories evaluate complex DNADNA, or Deoxyribonucleic Acid, is the genetic material found in cells, composed of a double helix structure. It serves as the genetic blueprint for all living organisms. Read Full Definition profiles. Instead of making every interpretation depend on a fixed allele-inclusion rule, PG estimates how probable the observed short tandem repeatA short tandem repeat is a microsatellite with repeat units that are 2 to 7 base pairs in length, with the number of repeats varying among individuals, making STRs effective for human identification purposes Read Full Definition (STR) results would be under competing propositions. Depending on the model, it can accommodate allelic dropout, drop-in, stutter, peak-height variation, allele sharing and unequal contributions from several individuals. The output is usually a likelihood ratio (LR) expressing the relative support provided by the genetic results.

PG is neither an automatic identification system nor a guarantee that every DNA mixture can be interpreted. Different software packages make different modelling choices, and the defensibility of an LR depends on the evidenceEvidence is any form of proof, such as objects, materials, or scientific findings, presented to establish or disprove a fact in a legal proceeding. It is used to reconstruct events and link or exclude individuals Read Full Definition, model, population assumptions, validationValidation, often referred to as method validation, is a crucial process in the laboratory when introducing a new machine, technology, or analytical technique. It involves a series of systematic steps and assessments to ensure that Read Full Definition, propositions and analystA designated person who examines and analyzes seized drugs or related materials, or directs such examinations to be done; independently has access to unsealed evidence in order to remove samples from the evidentiary material for Read Full Definition decisions. NIST’s 2024 scientific foundation review identifies probabilistic genotyping as a particularly valuable approach for DNA mixtures at the sub-source level while stressing the need for validation with casework-like samples.

This Simply Forensic guide explains how PG works, compares major software programs, provides three fully calculated hypothetical statistical examples, discusses validation and common failure modes, and explains how results should be reported. It builds on our guides to CPI, RMP and LR statistics, mixture interpretation and low-template DNA.

Why is probabilistic genotyping needed?

A clear, high-quality single-source STR profile often permits relatively direct genotype inference. Mixtures are different. Several contributors may share an allele, causing their signal to appear at the same electropherogram position. A minor contributor’s allele may overlap a stutter product or fall below detection. Degradation, inhibition, stochastic PCR variation and limited template can also distort observed peak heights. These factors leave several possible genotype combinations consistent with the same dataInformation in analog or digital form that can be transmitted or processed. Read Full Definition.

Traditional binary approaches classify peaks and apply allele-inclusion rules. Such procedures can be suitable for defined types of profiles but may lose information, particularly when some peaks are uncertain or quantitative differences are informative. PG evaluates a range of genotype explanations, with probabilities derived from a model of the evidence-generation process. This makes it possible to use information that would otherwise be lost or treated as all-or-nothing. For background, see our number-of-contributors guide and STR artefact guide.

The 2025 SWGDAM probabilistic-genotyping guidelines make clear that software assists interpretation rather than replacing the analyst. PG is especially useful where there are credible quantitative data and a validated model capable of handling the relevant mixture complexity. It is not suitable for every failed, contaminated or uninterpretable profile simply because a program accepts an input file.

The mathematical foundation: an LR compares two propositions

For a specified set of observed DNA data E and competing propositions H₁ and H₂, the evidential likelihood ratio is:

LR = P(E | H₁, model and assumptions) / P(E | H₂, model and assumptions)

For example, H₁ may state that the person of interest (POI) and one unrelated unknown individual contributed DNA. H₂ may state that two unrelated unknown individuals contributed instead. Each likelihood evaluates the same observed evidence, but assumes a different contributor proposition. Models sum or integrate over possible unknown genotypes and other parameters. Those possibilities are weighted by genotype frequencies and the probability of observing the peaks if the genotype combination were true.

When the LR exceeds 1, the observed DNA results favour H₁ over H₂ in the specified orientation. Below 1, they favour H₂. An LR of 1 is neutral for that comparison. **An LR is not P(H₁ | E)**. Computing a posterior probability for a proposition would require prior odds and Bayes’ theorem, along with the rest of the relevant evidence. At source level, the LR does not determine the activity, time or manner of DNA deposition. See our hierarchy-of-propositions guide.

Binary, semi-continuous and fully continuous models

The NIST scientific foundation review distinguishes discrete or semi-continuous models from continuous models partly by whether the actual quantitative heights influence the likelihood. This distinction should not be confused with whether an analyst looks at the peak-height information during preliminary data assessment; appropriate analytical review remains necessary in either approach.

What information is lost when peak heights are ignored?

Imagine a hypothetical locusA locus is the precise position of a gene on a chromosome. Different alleles of the same gene occupy the same locus. Read Full Definition with two credible allele positions: allele 10 at 900 RFU and allele 12 at 180 RFU. A semi-continuous model may represent the observation as “10 detected, 12 detected” and assign probabilities to missing or extraneous allelesAlleles are different forms of a gene resulting from mutations or variations in the DNA sequence or gene expression. They can be dominant (expressed with one copy) or recessive (expressed only with two copies). Read Full Definition. A continuous model may additionally evaluate whether those heights are plausible under a one-person heterozygote, a dominant component with a weaker contributor, a shared allele or another genotype configuration. The heights are informative but do not uniquely identify the contributors, and the example does not yield an LR without a fully specified statistical model.

What biological and analytical features can PG model?

Allelic dropout and drop-in

Dropout occurs when an allele actually present in the source fails to produce a qualifying detected signal. PG assigns probability to this possibility, although the exact parameterization varies. Some models use dropout probabilities as direct inputs; others represent PCR signal variability and detection limits so that missing signals arise probabilistically. Drop-in addresses unexpected allelic signals not explained by the nominated contributor genotypes, again under model-specific definitions. A suspected extra peak must first be evaluated as possible stutter, contaminationContamination - The unwanted transfer of material from another source to a piece of physical evidence. The inadvertent touching of a weapon, thereby adding fingerprints to it is an example of evidence contamination. Read Full Definition, another contributor or an analytical feature. For more detail, see our low-template DNA guide.

Stutter, allele sharing and peak-height balance

Repeat slippage during PCR generates stutter products, which can overlap genuine minor-contributor alleles. Fully continuous models can represent stutter peaks as dependent on parent-allele signals using empirically fitted behaviour. A high peak at a shared allele position may reflect molecules from more than one person; software must account for possible genotype assignments rather than treating the entire fluorescence as one individual’s signal. Peak-height balance and its stochastic variation are likewise modelled with methods supported by the system’s validation.

Degradation, contributor proportions and analytical limits

More degraded DNA tends to lose informative signal at longer amplicons. Models may represent locus- or size-dependent attenuation. Contributor proportions determine how much DNA from each donor is expected to contribute to the observed fluorescence, but they may not be constant across loci or precisely known. The detection threshold, saturation, missing loci and excluded artefact regions must be handled according to the specific model and laboratory procedures. Not every PG platform models every characteristic in the same way.

From data to LR: how the PG workflow operates

The software does not simply “read a suspect’s peaks and declare a match.” It evaluates the probability of the observed evidence under different genotype combinations and, for continuous models, quantitative signal expectations. Population frequencies govern how likely possible unknown genotypes are. The analyst selects and documents assumptions, sets propositions and determines whether the data and model are suitable. The resulting likelihood is conditional on those decisions.

What algorithms are used: MCMC, maximum likelihood and Bayesian networks?

Markov chain Monte Carlo (MCMC) can sample a complex probability distribution over genotype combinations and nuisance parameters. This is useful when exact summation is computationally difficult. MCMC-based systems require adequate mixing and convergence checks; repeated runs may be used to assess computational stability. A chain producing a number does not prove that it explored all relevant genotype configurations appropriately.

Maximum-likelihood estimation (MLE) identifies parameter values maximizing the likelihood under the model. Some PG programs use likelihood optimization as their main inference method and may have alternative Bayesian modes. Bayesian networks express conditional relationships in a graphical model and can support summation over many possible genotype configurations. Algorithms differ not only in computational efficiency but also in which parameter uncertainties are estimated, integrated or optimized.

Do not confuse the posterior distribution of unknown genotypes under a model with the likelihood ratio for two evidential propositions. A program may use Bayesian techniques internally yet report a likelihood ratio rather than the probability that the POI contributed.

Major forensic DNA probabilistic-genotyping software

This is a technical comparison, not a product ranking. STRmix and TrueAllele are proprietary commercial systems. EuroForMix is distributed openly and its model has been described by Bleka, Storvik and Gill (2016). The DNAmixtures project explains its Hugin dependency and DNAmixturesLite alternative. The Netherlands Forensic Institute identifies LRmix Studio as a legacy application that has been superseded by newer NFI tools. The choice of tool depends on validated casework scope, not a blanket claim that one software package is universally best.

Inter-software comparisons can be informative but must be interpreted carefully. Cheng and colleagues (2021) found close agreement for several relatively unambiguous analyses but also differences for mixtures and rare-allele settings; differences in minimum allele frequencies and theta corrections could be consequential. A separate 2024 study by Greenspoon and colleagues compared two widely deployed continuous systems. Agreement is evidence of consistency under tested conditions, not independent proof that all model assumptions are correct.

Worked example 1: a basic genotype likelihood ratio

Start with a complete, high-quality hypothetical single-source genotype A/B. Assume allele frequencies p(A) = 0.20 and p(B) = 0.30, unrelated random donors, Hardy–Weinberg equilibrium, no dropout, no drop-in, and an exact reliable genotype call.

Let H₁ state that the person of interest, whose known genotype is A/B, contributed the DNA. Let H₂ state that one unrelated, randomly selected unknown individual contributed instead. The simplified likelihoods are:

Thus these data are about 8.33 times as probable under H₁ as under H₂. This is a pedagogical limiting case where the LR resembles the inverse genotype frequencyFrequency is a fundamental concept in physics and wave theory. It refers to the number of times a specific point on a wave, such as a crest or trough, passes a fixed reference point in Read Full Definition. It is not a PG software output and does not mean there is an 8.33-in-9.33 probability that H₁ is true. In low-template or mixed results, the simple perfect-detection assumptions may fail.

Worked example 2: how dropout assumptions change the LR

Suppose the actual observed evidence at a hypothetical single-source locus contains only allele A. The POI is a known A/B heterozygote. Assume p(A) = 0.30, with all non-A alleles together accounting for frequency 0.70. In a deliberately simple semi-continuous teaching model, each of the two allele copies is independently detected with probability d = 1 − q, and drops out with probability q. We assume no drop-in, no artefacts, independent copies, and that seeing at least one detected A copy creates the observed A position. This per-copy model is a mathematical teaching simplification, not a fitted PCR model.

Under H₁, the POI has A/B, so A must be detected and B must drop out: P(E | H₁) = (1 − q)q. Under H₂, a random unrelated person contributes. Their genotype can be A/A (probability 0.30² = 0.09) or A/non-A (probability 2 × 0.30 × 0.70 = 0.42). For A/A, at least one copy must be detected: probability 1 − q². For A/non-A, A must be detected and the other allele must drop out: probability (1 − q)q. Therefore:

P(E | H₂) = 0.09(1 − q²) + 0.42(1 − q)q.

For q = 0.20, the numerical calculation is LR = 0.1600 / 0.1536 ≈ 1.042. With q = 0.05, the same A-only observation instead favours H₂ (LR approximately 0.433). If the model assumes no dropout at all, an A/B POI cannot yield only A and the idealized H₁ likelihood is zero. The conclusion therefore changes when the permitted dropout process changes.

The lesson is not that laboratories should choose a dropout probability producing a preferred LR. The model and plausible parameters must be established through appropriate scientific data and validation. Fully continuous PG can model observed peak heights and stochastic variation in considerably greater detail than this example. For why dropout probabilities cannot be inferred from a single high mixture peak, see our threshold guide.

Worked example 3: how justified conditioning changes the question

Consider a separate toy example with two contributors at a single locus and exactly the observed allele set {A, B}. Assume p(A)=0.30, p(B)=0.50, and the remaining allele-frequency mass is 0.20. There is complete detection, no artefacts, independent unrelated unknown contributors, and no quantitative peak-height information. The POI’s genotype is B/B.

First, without conditioning on any other known person, compare H₁ = POI B/B + one unrelated unknown, with H₂ = two unrelated unknowns. Under H₁, the unknown genotype must be A/A or A/B to provide A without introducing an outside allele. Therefore P(E | H₁) = 0.30² + 2(0.30)(0.50) = 0.39. Under H₂, all four allele draws must be A or B, with at least one A and one B:

P(E | H₂) = (0.80)⁴ − (0.30)⁴ − (0.50)⁴ = 0.339.
Unconditioned LR = 0.39 / 0.339 ≈ 1.150.

Now suppose separate, independent case information justifies treating a known individual with genotype A/A as an assumed contributor under both propositions. Compare H₁ = assumed A/A contributor + POI B/B, with H₂ = assumed A/A contributor + one unrelated unknown. Under H₁, the exact allele set {A,B} is assured by the two known genotypes: P(E | H₁) = 1. Under H₂, the unknown must have A/B or B/B: P(E | H₂) = 2(0.30)(0.50) + (0.50)² = 0.55. Thus the conditioned LR = 1 / 0.55 ≈ 1.818.

Conditioning changed both propositions, not merely the denominator. These figures are not competing answers to precisely the same question. The presence of an assumed contributor must be independently justified, and an LR increase is not the purpose of conditioning. This exercise demonstrates why the propositions and assumptions must accompany any result.

Figure 3: sensitivity is model-dependent

Real continuous models estimate or integrate more parameters, including signal variability and possible contributors. As complexity increases, sensitivity analysis may include alternative NOC assumptions, population-frequency settings, genotype relationships and stochastic parameters. A single neat diagram cannot establish the performance of a full software system.

How should a forensic laboratory validate probabilistic genotyping?

Developmental validation concerns the scientific model and software as developed and tested by its authors. Internal validation concerns whether the specific laboratory can use a particular version, STR kit, instruments, settings, data-processing workflow and casework range reliably. The Haned and colleagues (2016) paper clarifies the distinction between the statistical model and its implementation and describes essential validation concepts. A vendor’s published paper cannot replace tests using the laboratory’s own materials and workflows.

Validation should evaluate known-contributor LRs, non-contributor LR distributions, discriminatory ability and any calibrationThe act of checking or adjusting (by comparison with a standard) the accuracy of a measuring instrument. Operation that, under specified conditions, in a first step, establishes a relation between the quantity values with measurement Read Full Definition or error-rate summaries supported by the design. It is misleading to claim a universal “false inclusion rate” for all DNA mixtures, because such behaviour depends on contributor number, comparison population, template, mixture ratio, assumptions and the particular decision rule.

Where MCMC is used, laboratories should understand how to assess convergence and stability in accordance with the system’s validated diagnostics. Where an optimizer is used, they should evaluate whether the solution is stable and whether alternative starting values or fit diagnostics reveal problems. No single diagnostic applies universally to all PG programs.

Important limitations and common failure modes

Uncertain number of contributors

A model assuming two contributors may not represent a true three-person mixture adequately, particularly when allele sharing hides a low-level donor. Increasing the number of contributors is not automatically a conservative fix: it changes the hypothesis space, parameter identifiability and LR. A validated approach should explain how candidate NOC assumptions are assigned, tested or represented, without choosing the setting solely because the POI appears to fit.

Incorrect conditioning and inappropriate propositions

An analyst should not assume that a known person’s genotype was present merely because it improves deconvolution. The presence of an assumed contributor must have an independent scientific or case-context basis. A proposition must also address the relevant comparison—unrelated unknown, close relative, known partner or another realistic alternative. A non-contributor comparison with an implausible H₂ may create a misleading evidential statement.

Population data, rare alleles and relatedness

Allele frequencies, population substructure correction (θ/F_ST), sampling uncertainty and minimum-frequency procedures can change LR values. A sibling alternative differs from a random unrelated donor. The Cheng et al. comparison documented particularly notable sensitivity to population treatments of rare alleles. Such choices need documented justification rather than copying numbers from another population without analysis.

Off-scale peaks, unmodelled artefacts and bad inputs

A powerful model cannot recover information lost through invalid analytical assumptions. Incorrectly labelled or retained artefacts, inappropriate thresholds, contaminated evidence and off-scale detector responses can yield model outputs that appear precise"Precise" refers to the degree of closeness or consistency between multiple measurements or values taken under the same conditions. It indicates how well these measurements agree with each other, regardless of whether they are accurate Read Full Definition but do not correspond to a reliable biological interpretation. The laboratory must assess analytical quality before input and monitor model suitability after analysis.

Weak information and misleading precision

Complex or low-template profiles may contain too little information for a meaningful comparison. PG can still produce a numerical LR, but its evidential value may be near 1 or sensitive to model assumptions. Presenting many decimal places does not make the result highly certain. LRs for some non-contributors can exceed 1; validation should characterize how often and under what circumstances this occurs rather than claiming a positive LR uniquely identifies a source.

Transparency, software differences and courtroom scrutiny

Forensic laboratories should be able to explain the model’s essential assumptions, input processing, contributor assignments, version history, validation results, limitations and LR computation in a manner suitable for independent review. This does not require every analyst to derive a high-dimensional likelihood by hand, but it does require more than an assertion that a computer produced a number.

Proprietary software raises additional questions about access to code, documentation, test data and reproducibility, but closed-source status alone does not determine scientific validity. Open-source availability permits inspection but does not ensure a model has been appropriately validated or applied. NIST and the UK Forensic Science Regulator emphasize the continuing role of scientific judgment and context.

The FBI’s Quality Assurance Standards for Forensic DNA Testing Laboratories, effective July 1, 2025, require appropriate validation of software used for DNA interpretation and statistical calculations for laboratories within their scope. SWGDAM’s 2025 PG guidelines specifically distinguish a PG system from an “expert system” as defined by the FBI QAS; those regulatory categories should not be conflated. Standards and guidance do not necessarily have the same legal status outside their jurisdictions.

How should probabilistic-genotyping results be reported?

An appropriate report identifies the competing propositions and the calculated LR, indicating which proposition the findings support and the statistical model applied. SWGDAM’s 2025 Guidelines for Reporting Likelihood Ratios recommend presenting the numerical estimate, with appropriate language and relevant limitations. A verbal-strength scale, when used, must be defined rather than treated as a substitute for the LR.

For example, a qualified statement might read: “The DNA results were evaluated under the propositions that the person of interest and one unrelated unknown individual contributed, versus two unrelated unknown individuals contributing. The likelihood ratio was calculated using the laboratory’s validated probabilistic-genotyping procedure.” Actual reports should state the numerical LR and relevant population and analytical assumptions as required by the laboratory’s reporting practice. This is educational wording, not a universal template.

An LR describes the weight of the genetic observations for the specified source-level propositions. It does not by itself establish whether DNA was transferred directly, whether a particular activity took place or whether the POI committed an offence. Additional propositions and evidence would be needed to assess those questions.

Practical analyst checklist before approving a PG result

1. Confirm the EPG data: controls, sizing, artefacts, peak-height quality, saturation and detection rules are acceptable.
2. Document the software configuration: version, kit, model parameters, population database and validated settings.
3. Set propositions appropriately: specify the contributors, alternatives, POI status and independent basis for any assumed donor.
4. Assess contributor-number assumptions: record the evidence supporting the selected model and evaluate alternatives where justified.
5. Review diagnostics: check computational stability, convergence or optimizer diagnostics and model-fit information supported by the program.
6. Inspect uncertainty: consider dropout, drop-in, stutter, relatedness, frequency corrections and sensitivity to plausible choices.
7. Review the LR and report: confirm the ratio’s orientation, propositions, reproducibility, scope and limitations; retain data and decision history.

Frequently asked questions

Scientific references and further reading

Related Simply Forensic guides

• Forensic DNA Mixture Statistics: CPI, RMP and Likelihood Ratios
• Forensic DNA Mixture Interpretation: Major–Minor and Unresolved Mixtures
• Number of Contributors in Forensic DNA Mixtures
• Allelic Dropout, Peak Height Balance and Low-Template DNA
• Analytical and Stochastic Thresholds in Forensic DNA
• STR Artefacts and Troubleshooting