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

Area of Science:

• Genetics
• Computational Biology
• Bioinformatics

Background:

• Identifying causal genetic variants computationally is challenging due to limitations in existing datasets and methods.
• Genome-wide association studies (GWAS) results are often confounded by linkage disequilibrium (LD).
• Current gene expression datasets lack individual-level genetic variation data.

Purpose of the Study:

• To develop a novel computational method for identifying causal genetic variants.
• To overcome the limitations of lacking strong ground-truth labels in genetic datasets.
• To improve the fine-mapping of GWAS results by disentangling correlated variants.

Main Methods:

• Proposed Multiple Instance Fine-mapping (MIFM), a multiple instance learning (MIL) objective.
• Grouped putatively causal variants based on their LD scores to address the lack of strong labels.
• Trained a deep classifier on over 13,000 GWAS datasets using DNA sequences to predict causal variants.

Main Results:

• Validated prioritized variants by constructing polygenic risk scores that showed improved transferability to different ancestries.
• Demonstrated MIFM's capability to disentangle effect sizes of highly-correlated variants.
• Successfully improved fine-mapping of GWAS results.

Conclusions:

• MIFM offers a robust computational approach for identifying causal genetic variants.
• The method enhances the utility of GWAS by improving fine-mapping and polygenic risk score accuracy across ancestries.
• MIFM represents a significant advancement in computational genetics for variant prioritization.