StratoMod: predicting sequencing and variant calling errors with interpretable machine learning

Despite the variety in sequencing platforms, mappers, and variant callers, no single pipeline is optimal across the entire human genome, so users make tradeoffs when designing pipelines for specific applications. Currently, assessing such tradeoffs relies on intuition about performance in given genomic contexts. We present StratoMod, an interpretable machine-learning classifier that predicts germline variant calling errors in a data-driven manner. StratoMod precisely predicts recall using HiFi or Illumina data and, through its interpretability, quantifies contributions from difficult-to-map and homopolymer regions to outcomes. We use StratoMod to assess effects of mismapping on predicted recall for linear vs graph-based references and identify hard-to-map regions where graph-based methods excel and by how much, leveraging a draft benchmark based on the Q100 HG002 assembly that includes previously inaccessible difficult regions. StratoMod also predicts clinically relevant variants likely to be missed, improving over pipelines that primarily filter likely false positives. This enables precise risk-reward analyses when designing variant calling pipelines.

Nathan Dwarshuis, Peter Tonner, Nathan D. Olson, Fritz J. Sedlazeck, Justin Wagner, Justin M. Zook