PLAIG: Protein-Ligand Binding Affinity Prediction Using a Novel Interaction-Based Graph Neural Network Framework.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Accurate prediction of protein-ligand binding affinity is crucial for efficient drug discovery.
• Existing machine learning models struggle with generalization due to incomplete feature representations.
• Overfitting remains a challenge in developing robust predictive models.

Purpose of the Study:

• To develop a generalized machine learning framework for predicting protein-ligand binding affinity.
• To address the limitations of existing models in terms of generalization and overfitting.
• To create a novel approach integrating structural and interaction features for enhanced prediction accuracy.

Main Methods:

• Developed PLAIG, a Graph Neural Network (GNN) framework representing binding complexes as graphs.
• Integrated protein-ligand interactions and molecular topology to capture unique features.
• Employed principal component analysis (PCA) and ensemble learning (stacking regressor) to mitigate overfitting.

Main Results:

• PLAIG achieved a Pearson Correlation Coefficient (PCC) of 0.78 on the PDBbind v.2019 refined set and 0.82 on the v.2016 core set.
• External validation on DUDE-Z showed an average AUC of 0.69, distinguishing active ligands from decoys.
• Hybrid models integrating PLAIG with other approaches reached an average PCC of 0.88, with a maximum of 0.98.

Conclusions:

• PLAIG offers a robust and generalized approach for protein-ligand binding affinity prediction.
• The GNN-based framework effectively integrates diverse molecular features, outperforming existing models.
• Future work will focus on incorporating docking methodologies and evaluating performance on de novo ligands.