Multiple instance fine-mapping: predicting causal regulatory variants with a deep sequence model

Abstract

Identifying causal genetic variants in a computational manner remains an open problem. Training end-to-end prediction models is not possible without large ground-truth datasets, while results of genome-wide association studies (GWAS) are entangled by linkage disequilibrium (LD), and gene expression datasets do not contain genetic variation at individual-level. Here, we propose Multiple Instance Fine-mapping (MIFM) - a multiple instance learning (MIL) objective to overcome the lack of strong labels by grouping putatively causal variants together based on their LD scores. Using MIFM, we trained a deep classifier on a dataset aggregating over 13, 000 GWAS to predict causal variants based on their underlying DNA sequences. We validated variants prioritized by MIFM by constructing polygenic risk scores which transferred better to different target ancestries. Furthermore, we demonstrated how MIFM can be used to disentangle effect sizes of highly-correlated variants to better fine-map GWAS results. 1 Author summaryGenome-wide association studies have identified tens of thousands genetic variants associated with traits or diseases. However, the majority of identified variants is only spuriously correlated with the phenotype of interest, having no causal effect on it. Instead, these variants are often inherited together with nearby biologically causal variants, thus creating the spurious associations. Fine-mapping, i.e., predicting which variants are causal, is crucial for downstream tasks, such as uncovering the biological mechanisms affecting the phenotype or robustly identifying individuals with high genetic risk of a disease. While most fine-mapping methods are based on the available association statistics or functional annotations of genetic regions, it should be possible to identify causal variants based on their neighboring DNA sequences. However, training a standard machine learning classifier for that task is obstructed by the scarcity of strong, ground-truth labels. Here, we proposed a method to train sequence models predicting variant causality using weakly-labeled data. We trained a model on a large set of associated variants, and demonstrated its utility by improving cross-ancestry predictions of genetic risk, or disentangling the effect sizes of highly correlated variants.