A new paradigm for applying deep learning to protein-ligand interaction prediction

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Protein-ligand interaction prediction is crucial for drug design.
• Current machine learning and deep learning models often lack accuracy and physical interpretability.
• Accurate prediction requires robust scoring systems for docking poses and binding affinity.

Purpose of the Study:

• Introduce IGModel, a novel deep learning framework for predicting protein-ligand interactions.
• Utilize geometric information of protein-ligand complexes for prediction.
• Ensure output scores have intuitive physical meaning.

Main Methods:

• Developed IGModel, a deep learning approach using geometric information of protein-ligand complexes.
• Integrated prediction of root mean square deviation (RMSD) and binding strength (pKd) into a single framework.
• Evaluated performance on CASF-2016, PDBbind-CrossDocked-Core, and DISCO benchmark datasets.

Main Results:

• Achieved state-of-the-art accuracies on standard docking power test sets.
• Demonstrated generalizability and robustness on unbiased and AlphaFold2-generated target structure datasets.
• Provided insights through latent space visualization and interpretability analysis.

Conclusions:

• IGModel offers a novel and effective deep learning framework for protein-ligand interaction prediction.
• The model provides physically meaningful predictions, advancing drug design.
• IGModel shows significant potential for future applications in computational drug discovery.