In this project, 1) a machine learning pipeline for analyzing unique molecular identifier (UMI) barcoded reads was created; 2) on low template data, single source allele calling was improved by 13 percent (PR AUC); and 3) allele calling for a 100 pg balanced mixture was improved by 10 percent.
PCR artifacts are an ever-present challenge in sequencing applications. These artifacts can seriously limit the analysis and interpretation of low-template samples and mixtures, especially with respect to a minor contributor. In medicine, molecular barcoding techniques have been employed to decrease the impact of PCR error and to allow the examination of low-abundance somatic variation. In principle, it should be possible to apply the same techniques to the forensic analysis of mixtures. To that end, several short tandem repeat loci were selected for targeted sequencing, and a bioinformatic pipeline for analyzing the sequence data was developed. The pipeline notes the relevant unique molecular identifiers (UMIs) attached to each read and, using machine learning, filters the noise products out of the set of potential alleles. To evaluate this pipeline, DNA from pairs of individuals were mixed at different ratios (1−1, 1−9) and sequenced with different starting amounts of DNA (10, 1 and 0.1 ng). Naïvely using the information in the molecular barcodes led to increased performance, with the machine learning resulting in an additional benefit. In concrete terms, using the UMI data results in less noise for a given amount of drop out. For instance, if thresholds are selected that filter out a quarter of the true alleles, using read counts accepts 2381 noise alleles and using raw UMI counts accepts 1726 noise alleles, while the machine learning approach only accepts 307. (Published abstract provided)