Affinity prediction using deep learning based on SMILES input for D3R grand challenge 4.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in cheminformatics

Background:

• Molecular docking predicts ligand pose and binding affinity, crucial for drug design.
• Current methods require extensive feature engineering and struggle with ligand position uncertainty.
• A robust model for sequential pose and affinity prediction is needed.

Purpose of the Study:

• To develop an efficient and accurate computational pipeline for molecular docking.
• To address the limitations of existing methods in feature engineering and pose prediction accuracy.
• To improve the prediction of both ligand pose and binding affinity in drug discovery.

Main Methods:

• A pipeline integrating a bipartite graph neural network (GNN) with transfer learning was developed.
• The model was trained on a re-docking dataset to learn generalizable features.
• Evaluation was performed on the Drug Design Data Resource Grand Challenge 4 (D3R GC4) dataset.

Main Results:

• The proposed model demonstrated superior performance on the BACE target protein, exceeding the best participant's results by 9% in stage 2.
• Competitive performance was achieved on the CatS target protein, indicating broad applicability.
• The model effectively handles the probabilistic deviation of ligand positions in docking predictions.

Conclusions:

• The developed pipeline offers a robust and accurate approach to molecular docking, combining pose and affinity prediction.
• This method reduces the need for extensive feature engineering, streamlining the drug discovery process.
• The model's success in the D3R GC4 challenge highlights its potential for advancing computational drug design.