A structure-guided approach to non-coding variant evaluation for transcription factor binding using AlphaFold 3

Abstract

Non-coding single-nucleotide variants (SNVs) that alter transcription factor (TF) binding can affect gene expression and contribute to disease. Sequence-based methods can excel at predicting TF binding, but rely on training data and can exhibit TF-specific biases. Here we propose a structure-guided approach for non-coding SNVs, using AlphaFold 3 (AF3) to model TF-DNA complexes and FoldX for downstream physics-based assessment. Benchmarked against SNP-SELEX data for six TFs (SPIB, ELK3, ETV4, SF-1, PAX5 and MEIS2), the FoldX-based strategy showed good agreement with experimental allele preferences. Interestingly, differences in AF3s interface predicted template modelling (ipTM) score aligned even more closely with SNP-SELEX results, generally surpassing energy-based metrics. Application to known disease-associated variants recapitulated most reported effects for TFs including NKX2-5, GATA3 and USF2A-USF1. In these examples, considering both {Delta}ipTM and FoldX energies proved more reliable than either metric alone. While less accurate than state-of-the-art sequence-based methods, this work demonstrates that structural modelling can yield interpretable insights into how non-coding variants influence TF binding. By highlighting both the promise and limitations of AF3 in this context, our study provides a framework for complementary structural evaluation of regulatory variants.