Multiple Instance Fine-mapping method uses deep sequence models to predict causal regulatory variants

Alexander Rakowski and Christoph Lippert, publishing in PLOS Genetics, have described Multiple Instance Fine-mapping (MIFM) — a computational method designed to improve the identification of causal genetic variants in regulatory regions identified through genome-wide association studies (GWAS).

Fine-mapping — the process of pinpointing which variant within a GWAS signal is functionally responsible for an association — is complicated by linkage disequilibrium (LD), which means that many variants in a genomic region are statistically correlated with one another. MIFM addresses this by applying a multiple instance learning (MIL) objective: rather than requiring a single variant to be labelled as causal, it groups putatively causal variants from the same LD block and trains a deep sequence model across the group. This sidesteps the lack of large ground-truth datasets that hampers end-to-end training of conventional models.

The authors demonstrate that MIFM improves prediction of causal regulatory variants compared with baseline approaches, using gene expression data without requiring individual-level genotype data. The method is relevant to statistical genetics, functional genomics, and researchers working to interpret non-coding GWAS hits — a persistent bottleneck in translating association signals into mechanistic understanding.