AffinityVAE: A multi-objective model for protein-ligand affinity prediction and drug design

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Traditional protein-ligand affinity prediction methods are computationally intensive and struggle with structural changes.
• Data-driven deep learning approaches offer promise but often lack interpretability.
• Existing methods may require additional domain knowledge or structural data, limiting their broad applicability.

Purpose of the Study:

• To develop an interpretable and efficient deep learning model for protein-ligand affinity prediction.
• To address the limitations of current methods in terms of computational cost and interpretability.
• To enhance drug discovery pipelines through improved affinity prediction and data generation.

Main Methods:

• Proposed Affinity Variational Autoencoder (AffinityVAE), a multi-objective model integrating interaction feature mapping and variational autoencoders.
• Introduced protein-ligand interaction feature maps to enhance model interpretability.
• Designed an adaptive autoencoder for chemical properties to generate diverse, novel ligand data, expanding the training set.

Main Results:

• AffinityVAE demonstrated high prediction performance, outperforming recent methods in efficiency and accuracy.
• The protein-ligand interaction feature map provided a novel approach to interpretability in affinity prediction.
• The adaptive autoencoder successfully increased the diversity and quantity of protein-ligand binding data.

Conclusions:

• AffinityVAE offers a powerful, interpretable, and efficient solution for protein-ligand affinity prediction.
• The model has significant potential to accelerate drug development by enhancing data availability and prediction accuracy.
• This work advances the integration of deep learning with interpretable methods in computational drug discovery.