MolXProt: A Cross-Attention Transformer-Based Graph Neural Network for Protein-Ligand Binding Affinity Prediction

Area of Science:

• Computational Chemistry
• Drug Discovery
• Bioinformatics

Background:

• Accurate prediction of drug-target binding affinities (DTAs) is crucial for efficient drug discovery.
• Existing methods like docking and empirical scoring often struggle with generalization to new cases.
• Developing scalable and accurate computational models for DTAs remains a significant challenge.

Purpose of the Study:

• To introduce MolXProt, a novel transformer-based graph neural network architecture for predicting DTAs.
• To evaluate the scalability and accuracy of MolXProt on large and diverse datasets.
• To investigate the model's ability to learn residue-atom interactions and provide insights into binding mechanisms.

Main Methods:

• Developed MolXProt, integrating graph ligand representations with protein language models via bidirectional multihead cross-attention.
• Utilized protein-token compression for computational efficiency.
• Performed analysis of token-residue interactions and latent space properties.
• Applied posthoc isotonic correction to address prediction bias.

Main Results:

• MolXProt scales to over 100,000 protein-ligand pairs, achieving 50% prediction accuracy within ±1.0 kcal/mol and 80% within ±2.0 kcal/mol.
• The model explicitly learns residue-atom interactions and identifies key binding pocket residues in benchmark complexes.
• Identified model biases, such as overpredicting strong binders and underpredicting weak binders, linked to data imbalances and noise.
• Demonstrated partial mitigation of bias using isotonic correction and successful learning of continuous binding affinity manifolds.

Conclusions:

• MolXProt presents a novel, scalable, and computationally efficient architecture for DTA prediction.
• The model provides valuable insights into binding mechanisms through transformer-based cross-attention.
• Further calibration and data handling strategies are needed to fully address prediction biases for improved drug discovery applications.