Comparative Analysis of Deep Learning Models for Predicting Causative Regulatory Variants.

Motivation: Genome-wide association studies (GWAS) have identified numerous noncoding variants associated with complex human diseases, disorders, and traits. However, resolving the uncertainty between GWAS association and causality remains a significant challenge. The small subset of noncoding GWAS variants with causative effects on gene regulatory elements can only be detected through accurate methods that assess the impact of DNA sequence variation on gene regulatory activity. Deep learning models, such as those based on Convolutional Neural Networks (CNNs) and transformers, have gained prominence in predicting the regulatory effects of genetic variants, particularly in enhancers, by learning patterns from genomic and epigenomic data. Despite their potential, selecting the most suitable model is hindered by the lack of standardized benchmarks, consistent training conditions, and performance evaluation criteria in existing reviews. Results: This study evaluates state-of-the-art deep learning models for predicting the effects of genetic variants on enhancer activity using nine datasets stemming from MPRA, raQTL, and eQTL experiments, profiling the regulatory impact of 54,859 SNPs across four human cell lines. The results reveal that CNN models, such as TREDNet and SEI, consistently outperform other architectures in predicting the regulatory impact of single-nucleotide polymorphisms (SNPs). However, hybrid CNN-transformer models, such as Borzoi, display superior performance in identifying causal SNPs within a linkage disequilibrium block. While fine-tuning enhances the performance of transformer-based models, it remains insufficient to surpass CNN and hybrid models when evaluated under optimized conditions.