Mixtures

When the peaks stop belonging to one person, interpretation becomes a statistical argument rather than a reading.

Reading One Profile Is Simple; Reading Two Is Not

A single-contributor profile from a good-quality sample is close to unambiguous. At each genetic site — each locus — a person has two alleles, so the instrument returns one or two peaks. You record the numbers, you compare them, you report. The logic is direct. The moment a second person's DNA is present in the sample, that clarity evaporates. Two contributors can produce up to four peaks per locus. Three contributors can produce six. Once you are above that threshold, the peaks begin to overlap, share positions, and mask one another, and the question shifts from "what does this say?" to "what could this plausibly say, and how much more plausible is that than the alternatives?"

Mixture interpretation has been a live problem since the earliest casework. Alec Jeffreys's original multilocus probes, developed at the University of Leicester in the mid-1980s, produced banding patterns where two contributors' bands could be distinguished visually when the ratio of their contributions was not too extreme. As the field moved to PCR-based short tandem repeat (STR) analysis and capillary electrophoresis, the sensitivity rose dramatically — but so did the complexity of the profiles the method could recover. More DNA could be read from a single touch; that DNA was far more likely to belong to more than one person.

What the Instrument Gives You

A capillary electrophoresis run produces an electropherogram: a series of peaks plotted against size, each peak representing an allele at a given locus, its height proportional to the quantity of DNA in the reaction. In a clean two-person mixture where both contributors are present at roughly equal amounts, the electropherogram may be interpretable by eye — two peaks per locus for each contributor, sitting neatly apart. In practice, the picture is rarely this accommodating.

Several artefacts complicate the read. Stutter is a PCR byproduct: replication slippage generates a small spurious peak one repeat unit shorter than the true allele, typically below fifteen percent of the true peak's height but sometimes higher in degraded or low-template samples. Drop-out occurs when an allele amplifies so weakly that its peak falls below the analytical threshold and disappears from the record entirely; drop-in introduces a single extraneous allele from contamination. A shared allele — where two contributors happen to carry the same allele at the same locus — produces a single peak that may look like a single-contributor result but actually represents two people's contributions stacked on top of each other. Each of these effects can cause an analyst to overcount contributors, undercount them, or misread which peaks belong together.

CROSS-REF

The question of how many contributors are present is itself not always answerable with certainty. A standard heuristic counts the maximum number of alleles seen at any locus and works back from there, but shared alleles can hide contributors and stutter can mimic them. The 2016 President's Council of Advisors on Science and Technology review ↗ of forensic feature-comparison methods found that complex mixture interpretation — particularly mixtures of three or more contributors — lacked the scientific foundation then claimed for it, a finding that put the field's existing practices under serious pressure.

From Reading to Calculation

The response to these complications was a shift in how mixture evidence is reported. Rather than an analyst asserting "this person is included" or "this person is excluded" — a binary that the data rarely justified — laboratories moved toward probabilistic frameworks that express the evidence as a likelihood ratio: the probability of seeing this electropherogram if the person of interest contributed DNA, divided by the probability of seeing the same electropherogram if they did not.

Probabilistic genotyping software — STRmix, TrueAllele, ArmedXpert and others — performs this calculation by modelling the full range of genotype combinations that could have produced the observed peaks, weighting each combination by its prior probability and by parameters learned from laboratory-specific data about stutter rates, peak height variability and other instrument behaviour. The output is a single number: a likelihood ratio that may run from modestly supportive of inclusion to many billions to one. What the software is doing, at its core, is turning an interpretive judgment into a quantified statistical argument — but the quality of that argument depends entirely on whether the underlying model is a good description of the actual data.

That dependency is where the current scientific argument lives. The National Institute of Standards and Technology has done substantial work on mixture studies, including collaborative exercises where the same profiles were sent to many laboratories and the range of results recorded; the spread in reported likelihood ratios for complex mixtures was wide enough to concern the field. SWGDAM publishes interpretation guidelines that laboratories are expected to work to, and those guidelines have been revised multiple times as the understanding of mixture behaviour has deepened. In England and Wales, the Forensic Science Regulator's codes of practice set requirements for validation, and the expectation that software be transparently documented and independently tested.

The Limits That Remain

None of this machinery removes the underlying difficulty: a mixture is fundamentally an ambiguous object, and below certain mixture ratios or above a certain number of contributors, even the best software is constructing a plausible story from limited evidence rather than reading an unambiguous signal. A minor contributor present at less than ten percent of the total DNA in a sample may produce peaks indistinguishable from stutter. A four-person mixture may yield likelihood ratios that vary by orders of magnitude depending on which analytical threshold is applied, which stutter model is used, or how the analyst has characterised the number of contributors before running the software.

The Innocence Project ↗ has documented cases in which mixture evidence contributed to wrongful convictions subsequently overturned. The National Academy of Sciences, in its 2009 report on forensic science in the United States, stressed the need for rigorous validation of forensic methods generally, while treating DNA analysis as the most scientifically grounded of them. That validation work continues, and the methods have genuinely improved — probabilistic genotyping is a real advance over binary inclusion — but the evidence a mixture produces is not the same kind of evidence a single-contributor profile produces, and presenting it as though it were is where the courtroom danger lives. A likelihood ratio is a number, and numbers carry authority; understanding what the number rests on, and what assumptions were made to produce it, is what determines whether that authority is earned.

The moment a second person's DNA is present in the sample, that clarity evaporates.