BoltzOmics: Rapid Assessment of Genetic Variant Effects on Drug Binding Using Boltz-2 Deep Learning Model

A mechanistic understanding of how genetic variants modulate drug–receptor binding is central to precision medicine, with direct implications for predicting drug response and accelerating drug development. However, experimental characterization of mutation-drug interactions is time-consuming and expensive, while existing computational methods face trade-offs between accuracy and computational efficiency. Deep learning approaches have improved protein structure prediction, with models such as AlphaFold enabling accurate prediction of protein folds from sequence. Boltz-2 extends these advances, producing accurate wild-type protein structures in minutes rather than the days or weeks required by traditional methods. The performance of Boltz-2 in modeling genetic variant effects on drug binding, a critical application for precision medicine, has not yet been evaluated. To address this issue, we have developed an interactive platform integrating the deep learning structure prediction model Boltz-2 to rapidly assess mutation effects on drug binding. Starting from amino acid sequences, we present an open-source workflow that queries public databases for potential genetic variants, generates both WT and mutant protein structures, and screens multiple drugs across these variants to predict changes in binding affinity. Our results show the efficacy of the approach across four diverse drug targets: hERG (cardiac ion channel), NaV1.5 (sodium channel), HER2 (oncogenic receptor), and CYP3A4 (metabolic enzyme), representing critical protein classes in drug discovery, therapy and safety. The Boltz-2–based pipeline yielded improved prediction accuracy relative to traditional approaches and enabled end-to-end prediction of mutation-dependent changes in drug binding with minimal setup. The efficiency gained through application of the Boltz-2–based pipeline is predicted to enable rapid evaluation of large numbers of variants and support high-throughput assessment of drug–variant interactions. These results highlight the potential of AI-based approaches to accelerate computational drug discovery. This repository contains the data files kept from drug docking and co-folding prediction runs, including structural and parameter files. Boltz2/- `*.json`: Prediction summary files, including input parameters and output results GALigandDock/- `<protein>/*.sc`: score files used for comparison- `inputs/<protein>/<drug>/*.pdb`: input complex structures for docking runs- `inputs/<protein>/<drug>/*.params`: ligand parameter files for docking runs- `inputs/<protein>/<drug>/protocol.xml`: docking protocol file- `inputs/<protein>/<drug>/GADock*.sh`: docking job scripts DeepDTA/- `predictions.json`: prediction file used for comparison- `inputs/drugs.txt`: ligand SMILES file- `inputs/*.txt`: protein sequence files GraphDTA/- `predictions_graphdta.json`: prediction file used for comparison- `inputs/drugs.txt`: ligand SMILES file- `inputs/*.txt`: protein sequence files RosettaFlexDock/- `flexdock_results.csv`: results file used for comparison- `inputs/<protein>/<drug>/*.pdb`: input complex structures for docking runs- `inputs/<protein>/<drug>/*.params`: ligand parameter files for docking runs- `inputs/<protein>/<drug>/protocol*.xml`: protocol files- `inputs/<protein>/<drug>/*Dock*.sh`: job scripts- `inputs/parse_flexdock_results.py`: results parsing script- `inputs/submit_all_flexdock.sh`: batch submission script- `inputs/deploy_flexdock.sh`: script that copies protocol and job files into drug folders AutoDockVina/- `results.csv`: results file used for comparison- `inputs/<protein>/receptors/*.pdb`: receptor structures for docking runs- `inputs/<protein>/ligands/*.sdf`: ligand structures for docking runs