On the Frustration to Predict Binding Affinities from Protein-Ligand Structures with Deep Neural Networks

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Accurate prediction of protein-ligand binding affinities is crucial for efficient drug discovery.
• Current deep learning models face challenges in generalizing and avoiding memorization.

Purpose of the Study:

• To investigate whether explicit descriptions of protein-ligand noncovalent interactions improve binding affinity prediction.
• To assess the extent of memorization versus true learning in deep neural networks for this task.

Main Methods:

• Utilized modular message passing graph neural networks to model proteins and ligands in various states.
• Compared models with and without explicit noncovalent interaction descriptors.
• Evaluated performance using simple memorization-based models as baselines.

Main Results:

• Explicitly describing protein-ligand noncovalent interactions offered no significant advantage over simpler descriptors.
• Memorization-based models achieved comparable performance, indicating a dominance of data recall over genuine learning.
• Current deep learning approaches may be overfitting to the training data.

Conclusions:

• Rethinking the necessity of detailed atomic environments for noncovalent interactions in binding affinity prediction.
• Emphasizing the need for more comprehensive protein-ligand structural data and community-driven efforts to mitigate biases.
• Suggests focusing on noncovalent interactions while potentially simplifying atomic environment descriptions.