AlphaFlex: Accuracy modeling of protein multiple conformations via predicted flexible residues

Understanding protein conformational dynamics is critical for elucidating disease mechanisms and facilitating structure-based drug design. While AlphaFold2-based methods have advanced static structure prediction, modeling multiple conformation states with both accuracy and efficiency remains challenging. We present AlphaFlex, an innovative framework that integrates flexible residue prediction with a directed masking strategy for multiple sequence alignments. By selectively relaxing co-evolutionary constraints in dynamic regions, AlphaFlex enables accurate prediction of biologically relevant states while preserving structural fidelity. Benchmarking on 69 apo-holo pairs from the CoDNaS dataset, including enzymes, binding proteins, and chaperones, reveals AlphaFlex’s superior performance in both flexible residue identification and multiple conformation prediction. Comparative analysis against AF-Cluster, AF2-conformations, and AlphaFlow demonstrates AlphaFlex’s significant advantage, achieving the highest success rate (42%) in accurately predicting apo/holo structures (TM-score >0.95 for both states) while consistently outperforming these methods in reproducing both apo and holo conformations. Notably, AlphaFlex maintains robust performance for membrane proteins, resolving inward- and outward-facing conformations with equal fidelity. Case studies further reveal its unique ability to sample physically plausible intermediate states along transition pathways. These results establish AlphaFlex as a pivotal computational tool for mapping protein conformational landscapes, with promising potential to accelerate structure-based therapeutic development.