#9 · Genomics

Reference benchmarks and Bayesian models quantify variant call and effect reliability

The dominant thread is that variant-level conclusions depend on upstream choices in read processing, filtering, and measurement modeling, and these choices are now evaluated directly. A reference-standard DNA mixture benchmark reports that DRAGEN with recurrent false-positive allele filtering and VAF thresholds maximizes sensitivity and specificity in targeted and whole-exome sequencing. NanoSimFormer applies simulation to evaluation, reproducing human variant calling and bacterial assembly from nanopore signal via basecaller-guided training. What is newer is explicit uncertainty and evidence handling: Lilace provides Bayesian variant effect estimates with uncertainty quantification from FACS-based deep mutational scanning data, and Y-Trim gates Adaptase tail trimming in ssWGBS reads on evidence while separating trim admission from cut placement.