Vladimir Chupakhin

From Static to Dynamic Structures: Improving Binding Affinity Prediction with Graph-Based Deep Learning @Jianyang_Zeng 1. Breaking new ground in drug discovery: The Dynaformer model integrates molecular dynamics (MD) simulations and graph-based deep learning, achieving state-of-the-art accuracy in predicting protein-ligand binding affinities. This approach goes beyond static structures, considering dynamic protein-ligand interactions over time. 2. Why it matters: Traditional models focus on static X-ray structures, which limit prediction accuracy. Dynaformer, trained on MD trajectories of 3,218 protein-ligand complexes, captures the full thermodynamic ensemble, significantly improving affinity predictions and ranking. 3. Real-world impact: In a virtual screen of HSP90, Dynaformer identified 20 promising drug candidates, with 12 showing measurable binding affinities, including novel scaffolds. This demonstrates the model's potential to accelerate hit discovery in drug development. 4. Performance: Dynaformer outperformed all baseline models on the CASF-2016 benchmark, with a Pearson r of 0.858 and low prediction bias, highlighting its robust prediction capabilities. 5. In-depth analysis: Case studies revealed Dynaformer’s strength in modeling enthalpy and entropy changes from MD data, leading to more accurate predictions for complex binding scenarios. It’s particularly effective in distinguishing ligands with subtle structural differences. 6. Future potential: Incorporating more high-quality MD data and refining the model could further enhance prediction accuracy, making Dynaformer a valuable tool in early drug discovery. 💻Code: 1drv.ms/f/s!Ah9r82oejj… 📜Paper: onlinelibrary.wiley.com/doi/10.1002/ad…