In short: Probably not.

Probabilistic genotyping software (for example, DNAStatistX) utilises peak heights of allele calls for analysis. In the regular workflow, allele numbers and peak heights can be easily exported from the analysis software (GeneMapper or GeneMarker), but in some cases, raw data might not be available, only the graphical electropherograms, for example, old cold cases or data from literature.

One of the promises of AI (Artificial Intelligence) and LLMs (Large Language Models) is that they can easily automate manual, repetitive tasks. To test the ability of LLMs, electropherogram images were generated from experimental test runs, in which only the allele numbers were denoted, not the peak heights. The models were prompted to analyse the uploaded images and output the results as tables.

While at first sight the results were impressive, comparison with the data from the fsa files revealed that, for each profile, the LLMs made multiple mistakes: missing alleles, determining the correct peak heights but switching the alleles, or completely hallucinating an arbitrary number.

This small test reinforces the notion that LLM output must always be checked and verified. In this case, if the expert must already manually read the data, the use of AI may be unnecessary.

Author

• Viktor Poór (University of Pécs, Medical School, Department of Forensic Medicine, Hungary)