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

Non-coding single-nucleotide variants (SNVs) that alter transcription factor (TF) binding can profoundly affect gene expression and cellular function. Although sequence-based computational methods can excel at predicting TF binding sites, they often exhibit TF-specific performance biases and rely heavily on high-quality training data. Here, we propose a structure-guided approach using AlphaFold 3 (AF3) to model TF-DNA complexes for downstream biophysical evaluation with FoldX. We benchmark this method on a SNP-SELEX dataset in which reference and alternative DNA alleles are tested for differential TF binding in vitro . For each allele pair, we generate AF3 models and assess binding differences using FoldX-derived binding energy changes and AF3’s interface predicted template modelling (ipTM) score. Across three TFs, SPIB, SF-1 (steroidogenic factor 1, encoded by NR5A1), and PAX5, differences in ipTM (ΔipTM) often align with SNP-SELEX results, slightly surpassing energy-based metrics, although performance varies by TF, with notably poorer performance for PAX5. We further apply this approach to clinically relevant non-coding SNVs from published studies, accurately recapitulating 5 of 6 reported effects. These findings highlight the promise of highly interpretable structural modelling for non-coding variant interpretation, which could even encompass complex homomeric or heteromeric TF assemblies. AF3 provides robust TF-DNA complex structures that can capture subtle differences in binding induced by single-nucleotide changes. Augmented by biophysical binding evaluation, such an integrative approach can elucidate how non-coding variants affect TF binding and regulatory function.