Improving ΔΔG Predictions with a Multitask Convolutional Siamese Network.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• Drug discovery lead optimization is costly and time-consuming.
• Relative binding free energy (RBFE) methods estimate binding changes from ligand modifications.
• Accurate RBFE prediction accelerates the identification of potent drug candidates.

Purpose of the Study:

• To propose and evaluate a Siamese convolutional neural network (CNN) for predicting RBFE.
• To compare the Siamese CNN's performance against a standard CNN and previous state-of-the-art methods.
• To investigate methods for improving RBFE prediction generalization.

Main Methods:

• A Siamese CNN architecture was designed for RBFE prediction.
• A multitask loss function was employed to enhance latent space regularization.
• Model performance was evaluated using Pearson's R correlation coefficient.
• Few-shot learning was explored to improve generalization on unseen protein families.

Main Results:

• The Siamese CNN achieved a higher prediction accuracy (Pearson's R = 0.553) than a standard CNN (Pearson's R = 0.5).
• The multitask loss improved upon previous Siamese network performance.
• RBFE prediction performance varied across different protein families (-0.44 to 0.97).
• Few-shot learning significantly improved generalization performance.

Conclusions:

• Siamese CNNs are well-suited for predicting RBFE in drug discovery.
• The proposed multitask loss enhances RBFE prediction accuracy.
• Few-shot learning is a viable strategy to improve model generalization for diverse protein targets.