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

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Genome-wide association studies (GWAS) identify noncoding variants linked to complex traits, but determining causality is challenging.
• Deep learning models, including CNNs and Transformers, are used for variant effect prediction, yet inconsistent benchmarks hinder comparisons.
• Standardized assessment is needed to compare leading models for predicting variant effects in enhancers and prioritizing causal single-nucleotide polymorphisms (SNPs).

Purpose of the Study:

• To establish a standardized benchmark for evaluating deep learning models in predicting variant effects in enhancers.
• To compare the performance of leading CNN and Transformer-based models for variant-effect prediction and causal SNP prioritization.
• To guide the selection of appropriate models for analyzing noncoding genetic variants.

Main Methods:

• Evaluated state-of-the-art deep learning models on nine datasets from MPRA, raQTL, and eQTL experiments.
• Assessed model performance on predicting regulatory impact of 54,859 SNPs in enhancers across four human cell lines.
• Compared models for predicting SNP regulatory effects and identifying causal SNPs within linkage disequilibrium (LD) blocks, including fine-tuning effects.

Main Results:

• CNN models (TREDNet, SEI) showed superior performance in predicting SNP regulatory impact in enhancers.
• Hybrid CNN-Transformer models (Borzoi) excelled at causal variant prioritization within LD blocks.
• Fine-tuning improved Transformer performance but did not fully close the gap with CNNs for enhancer effect prediction.

Conclusions:

• CNN architectures are most reliable for estimating enhancer regulatory effects of SNPs under a unified benchmark.
• Hybrid CNN-Transformer models are superior for causal SNP identification within linkage disequilibrium (LD) blocks.
• These findings provide guidance for selecting models in variant-effect prediction for noncoding regions.