To ML-Predict or Not to ML-Predict: The Impact of Machine Learning-Predicted Protein Structures on FEP Accuracy and Data Augmentation

Abstract The rapid advancement of machine learning (ML)-based protein structure prediction, exemplified by AlphaFold2 and extended by newer models such as AlphaFold3 and Boltz-2, has generated significant optimism for structure-guided drug discovery. In particular, ligand–protein cofolding approaches offer the potential to overcome limitations in generating starting structures for physics-based free energy perturbation (FEP) calculations. However, the practical readiness of ML-predicted structures for FEP applications remains insufficiently evaluated. Here, we systematically assess experimentally determined crystal structures, a homology model, and ML-predicted protein structures as inputs for FEP using a well-characterized congeneric series targeting the tyrosine kinase cSrc. A data set of 133 compounds was evaluated through more than 1400 FEP calculations under minimal optimization to approximate “out-of-the-box” performance. By maintaining consistent preparation protocols, we isolate the impact of structural origin on predictive accuracy. Variable performance was observed across both experimental and ML-predicted structures, highlighting that even under this idealized benchmark scenario, significant challenges remain in reliably generating and refining predictive protein–ligand complexes. This study demonstrates that predictive variation in micro and macro conformational states─rather than the structural source─governs predictive reliability, underscoring the need for careful validation when integrating ML-derived structures into FEP workflows.