Preprint: underestimating causal variants may explain why sequence-to-function models underperform

A preprint posted to bioRxiv from Cold Spring Harbor Laboratory (not yet peer-reviewed) proposes that a long-standing puzzle in regulatory genomics — why deep-learning sequence-to-function (S2F) models classify putatively causal eQTL single-nucleotide variants (SNVs) reasonably well but dramatically underperform simple linear baselines when asked to rank individuals' gene expression levels from their whole-genome sequences — has a methodological rather than architectural cause.

The authors argue that causal variant underestimation is the dominant overlooked driver: benchmark sets used to assess S2F models systematically exclude many true causal variants, so the models are evaluated against an incomplete reference. When the ground-truth variant set is sparse or biased, models that correctly weight a broader set of variants appear to perform poorly against a metric tuned to a narrow target.

The finding has implications for how the field evaluates and iterates on S2F models, which are widely used in functional fine-mapping of GWAS loci — the process of identifying which specific variant within an associated genomic region is most likely to be biologically causal. If the authors' analysis is confirmed, improvements to variant curation and benchmark construction may yield larger gains than further architectural innovation.

As a preprint, this work has not been peer-reviewed and findings should be interpreted with appropriate caution.