Evaluating Pre-Trained Deep Learning Models for Integrin-Ligand Interaction Prediction Using AlphaFold 3 and Experimental Crystal Structures

Accurate computational prediction of small molecule binding to integrins is critical for drug discovery, yet deep learning models have not been systematically benchmarked against one another on these therapeutically important targets. This study evaluated three deep learning approaches: AlphaFold 3, DiffDock, and DynamicBind, for predicting integrin-ligand binding poses across two structurally distinct integrins: LFA-1 (containing an I-domain) and αIIbβ3 (lacking an I-domain). Twelve experimentally determined co-crystal structures (six per integrin) were used as benchmarks. For each receptor-ligand pair, 25 poses were generated per model and evaluated by heavy-atom root mean square deviation (RMSD) against the crystallographic reference conformation. Poses were subsequently rescored using AutoDock Vina in score-only mode to assess whether binding affinity estimates correlate with geometric accuracy. Results demonstrated target-dependent performance with no single model excelling uniformly. AlphaFold 3 produced superior poses on LFA-1 (mean RMSD 1.66 Å, 67.3% success rate at ≤2 Å), significantly outperforming DiffDock and DynamicBind. However, this advantage disappeared on αIIbβ3, where all three models converged to comparable performance (RMSD 1.73-1.90 Å). A significant Model × Integrin interaction confirmed that pose-prediction accuracy is target-dependent rather than a uniform model property. Critically, Vina rescoring failed to reliably rank poses by accuracy across all model-integrin combinations. These findings indicate that rigid-receptor scoring is incompatible with flexible-receptor model outputs and should not be used as a post-hoc filter. Target-specific benchmarking and dynamics-aware scoring functions are essential for reliable integrin-ligand predictions in drug discovery.