Probabilistic Genotyping and the Software Behind It

Probabilistic genotyping is a class of software that takes a DNA profile no analyst could resolve by hand and returns a number. It does not decide who contributed. It calculates how well the observed data are explained by one stated proposition compared with another, and reports the comparison as a ratio. Everything contestable sits either in the model or in what the analyst told the model to assume.

What the program computes

The starting point is the trace and a set of candidate explanations. For a mixture with an assumed number of contributors, there are many combinations of genotypes that could have produced the observed peaks, and the program enumerates them and assigns each a weight. The weight reflects how well that combination accounts for what the instrument recorded, given the variation expected at the amounts present.

Two propositions are then evaluated against those weights. The output is a ratio: how much more probable the observed data are under the first proposition than under the second. A large ratio favors the first, a very small one favors the second, and a ratio near unity means the data separate the propositions hardly at all. Many laboratories report the size of the ratio in verbal bands rather than reading digits aloud.

The ratio inherits everything upstream of it. The interpretation problems that make a mixture hard in the first place, set out in mixture interpretation and its limits, do not disappear because a program is doing the arithmetic; they become model assumptions instead of manual judgments.

How the two model families differ

Semi-continuous models use the presence or absence of variants. They treat a locus as a list of variants seen, estimate a probability that a variant dropped out or dropped in, and weigh genotype combinations accordingly. They make fewer assumptions about the instrument and are less sensitive to how the laboratory set its thresholds. They also discard the information carried in how tall each peak is.

Fully continuous models use the heights. They model expected peak height as a function of the amount of template contributed, model stutter as a proportion of the parent peak, and model the variability around both. Because they use more of the data, they extract more information from a difficult sample and typically produce ratios further from neutrality. The price is a larger set of parameters estimated from the laboratory's own data, and a stronger dependence on the instrument and chemistry those parameters were fitted to.

Neither family is a substitute for adequate data. Where a component is represented by a handful of low peaks, both approaches return a modest figure, and the fully continuous model returns it with more machinery. A laboratory that runs both applies written criteria to decide which model a sample goes to.

Propositions and the analyst's inputs

The ratio has no meaning apart from the two propositions compared. A common pair asks whether the sample came from the reference person and one unknown, or from two unknowns. Changing the alternative changes the answer: comparing against a relative rather than an unrelated person moves the figure, sometimes substantially. The propositions are written by a person, and should appear in the report in the words used to run the calculation.

Other inputs matter as much. The assumed number of contributors constrains the entire enumeration. An assumed contributor, where the calculation is conditioned on a known person such as the owner of a garment, removes that person's variants from the pool of things to be explained and can raise the figure for everyone else. Which references were compared, and whether the analyst knew them before fixing the interpretation, is recorded in the case file rather than in the report.

The two validation studies

Developmental validation is performed by the developer of the system. It tests the model across mixture types, contributor numbers, ratios and template amounts, and establishes that the software does what its designers intended under the conditions examined. It describes the program rather than any particular laboratory that later runs it.

Internal validation is performed by the laboratory that will use the system, on its own instruments, chemistry and staff. It runs mixtures of known composition through the whole workflow and compares the figures produced against the truth known in advance. A complete study records the mixture types tested, the contributor numbers and ratios covered, the template amounts, the parameters fitted for the laboratory, the spread seen on repeated runs, and the range of conditions the laboratory concluded it could interpret. That last item is the operative one, because it is the boundary a case sample either sits inside or outside.

The NIST scientific foundation review examines probabilistic genotyping within the likelihood-ratio framework and treats the empirical record as the measure of what the tools support. Under Rule 702 the proponent must show that reliable principles were reliably applied to the facts, and the internal validation study is where the second half of that showing usually lives. The broader shape of that demonstration is taken up in the reliability showing for a forensic method.

Access to the code and the record

Two requests are often conflated. The narrower one seeks case-specific material: the electronic data, the settings used, the run output, the diagnostics and the interpretation procedure. Those are laboratory records of a test performed in the case, and Rule 16 reaches results of scientific tests along with a statement of the bases for an expert's opinions. Objections tend to concern format and burden rather than entitlement.

The broader request seeks the source code. Developers resist on the ground that the code is a trade secret whose disclosure would destroy its commercial value, and they point to published validation studies as the appropriate measure of whether a tool works. The opposing position is that a figure produced by a program cannot be examined by testing outputs alone, since an implementation error is invisible in validation runs that never encounter it, and that a protective order can address commercial concerns. Courts have divided, and orders granting access commonly set conditions on who may inspect the code and under what secrecy obligations.

Practical review runs on separate tracks. A reviewer reads the internal validation study against the case sample, asks whether the settings and propositions match what the report describes, checks whether reruns fell inside the laboratory's tolerance, and asks what the figure becomes under a different contributor count. Funds for that review are available in federal court under 18 U.S.C. § 3006A(e) on an ex parte showing of necessity, described in retaining an independent forensic expert. What the figure claims, and what it does not, is the subject of how a DNA statistic is expressed.