What question did this study set out to answer?

To develop a machine-learning-based framework to discover new HER2 inhibitors among existing FDA-approved drugs.

April 8, 2026Open Access

Machine learning-driven drug repurposing for HER2-positive breast cancer

Key Points

To develop a machine-learning-based framework to discover new HER2 inhibitors among existing FDA-approved drugs.
Created a dataset of 13,944 HER2 bioactivity records from BindingDB.
Built predictive models using optimized descriptors and feature selection.
Employed Random Forest model with 75 descriptors for performance assessment.
Conducted docking studies against both wild-type and mutant HER2 proteins.
Performed molecular dynamics simulations and ADMET analysis on top candidates.
The Random Forest model achieved R² = 0.81 with RMSE = 0.86 and MAE = 0.6483.
FDA0870 showed strong interactions with HER2, outperforming the reference ligand.
Docking scores for wild-type HER2 ranged from -10.5 to -7.5 kcal/mol, while mutant scores were -11.6 to -8.0 kcal/mol.
Molecular dynamics simulations indicated stable complexes with average RMSDs of 3.4 Å for wild-type and 3.9 Å for mutants.
ADMET analysis indicated FDA0870 had good drug-like properties, suggesting potential for further development.

Abstract

Breast cancer, particularly the Human Epidermal Growth Factor Receptor 2 (HER2)-positive subtype, remains a significant clinical challenge due to its aggressive nature and the frequent development of resistance to targeted therapy. In this study, we created a comprehensive machine-learning (ML)-based drug repurposing framework to identify new HER2 inhibitors among FDA-approved drugs. A carefully curated dataset of 13,944 HER2 bioactivity records from BindingDB was used to build high-accuracy predictive models. This involved optimizing descriptors, reducing dimensionality, and selecting key features. The Random Forest model trained on 75 optimized descriptors demonstrated the best performance (R² = 0.81, RMSE = 0.86, MAE = 0.6483). This allowed us to understand structure-activity relationships and perform reliable virtual screening of 4,099 FDA-approved drugs. Docking studies against HER2 wild-type (7PCD) revealed strong interactions, with scores from -10.5 to -7.5 kcal/mol. For the exon-20 insertion mutant (8U8X), the affinities were even higher (-11.6 to -8.0 kcal/mol), exceeding the reference co-crystal ligand in both cases. FDA0870 (Timolol maleate) consistently ranked as a top dual inhibitor, forming stable hydrogen bonds and hydrophobic contacts with key catalytic residues of both wild-type and mutant HER2. Molecular dynamics simulations supported its stability, with wild-type HER2-FDA0870 complexes showing an average protein RMSD of 3.4 Å and a ligand RMSD of 2.9 Å, whereas mutant complexes had an average protein RMSD of 3.9 Å and a ligand RMSD of 2.1 Å over extended trajectories. Persistent interactions with residues such as ASP863/ASP867, THR862/THR866, and LYS753 were observed. ADMET analysis indicated that the drug had good drug-like properties, moderate permeability, a high unbound fraction, and acceptable safety features, suggesting its potential for further development. This integrated ML, docking, MD, and ADMET approach demonstrates that FDA0870 is a promising repurposed dual inhibitor for HER2-positive and HER2-mutant breast cancer, highlighting the value of computational pipelines in speeding up the search for targeted cancer therapies.

AI से पूछें

Bookmark

View Full Paper