From Static to Dynamic Structures: Improving Binding Affinity Prediction with Graph-Based Deep Learning

Area of Science:

• Computational chemistry and cheminformatics
• Structural biology and drug design
• Artificial intelligence in molecular modeling

Background:

• Accurate prediction of protein-ligand binding affinities is crucial for structure-based drug design.
• Current data-driven methods are limited by reliance on static protein structures, neglecting dynamic binding ensembles.
• Molecular dynamics (MD) simulations offer a way to approximate these dynamic ensembles.

Purpose of the Study:

• To develop a deep learning model that leverages MD simulations for enhanced protein-ligand binding affinity prediction.
• To evaluate the model's performance on benchmark datasets and in virtual screening applications.
• To accelerate the early stages of drug discovery through improved computational prediction.

Main Methods:

• Curated an MD dataset of 3,218 protein-ligand complexes.
• Developed Dynaformer, a graph-based deep learning model, to learn from MD trajectories.
• Applied Dynaformer to CASF-2016 benchmark dataset for scoring and ranking evaluation.
• Conducted virtual screening on heat shock protein 90 (HSP90) and experimentally validated hit compounds.

Main Results:

• Dynaformer demonstrated state-of-the-art scoring and ranking power on the CASF-2016 benchmark.
• The model outperformed previously reported methods in binding affinity prediction.
• Virtual screening identified 20 candidate compounds for HSP90, with 12 validated as hits, including novel scaffolds.

Conclusions:

• Dynaformer effectively predicts protein-ligand binding affinities by integrating MD simulations and deep learning.
• The model shows significant promise for accelerating virtual drug screening and the early drug discovery process.
• This approach offers a powerful computational tool for identifying novel drug candidates.