Comparative Analysis of Deep Learning Models for Predicting Causative Regulatory Variants

Background/Objective: Genome-wide association studies (GWAS) have linked many noncoding variants to complex traits and diseases, but distinguishing as-sociation from causation remains difficult. Deep learning models—particularly CNN- and Transformer-based architectures—are widely used for this task, yet comparisons are hindered by inconsistent benchmarks and evaluation practices. We aimed to establish a standardized assessment of leading models for predicting variant effects in enhancers and for prioritizing putative causal SNPs. Methods: We evaluated state-of-the-art deep learning models under consistent training and evaluation conditions on nine datasets derived from MPRA, raQTL, and eQTL ex-periments. These datasets profile the regulatory impact of 54,859 single-nucleotide polymorphisms (SNPs) across four human cell lines. Performance was compared for two related tasks: predicting the direction and magnitude of regulatory impact in enhancers and identifying likely causal SNPs within linkage disequilibrium (LD) blocks. We addi-tionally assessed the effect of fine-tuning on Transformer-based models and the impact of certainty in experimental results. Results: CNN models such as TREDNet and SEI performed best for predicting the reg-ulatory impact of SNPs in enhancers. Hybrid CNN–Transformer models (e.g., Borzoi) performed best for causal variant prioritization within LD blocks. Fine-tuning benefits Transformers but remains insufficient to close the performance gap. Conclusions: Under a unified benchmark, CNN architectures are most reliable for esti-mating enhancer regulatory effects of SNPs, while hybrid CNN–Transformer models are superior for causal SNP identification within LD. These comparisons help guide model selection for variant-effect prediction in noncoding regions.