BiAtt-GVAE: Molecular Design for Specific Target via Graph Variational Autoencoder Based on Bi-Channel Interactive...

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Designing bioactive molecules with specific properties for therapeutic targets remains a significant challenge in drug discovery.
• Existing models often struggle to effectively constrain and enhance generated molecules for desired properties.

Purpose of the Study:

• To introduce BiAtt-GVAE, a novel deep learning model for enhanced bioactive molecule generation.
• To improve the design of molecules with desired properties for specific biological targets.

Main Methods:

• Developed BiAtt-GVAE, a generative model incorporating conditional constraints.
• Designed a bi-channel interactive attention network to capture protein-ligand interactions and ligand structure-property relationships.
• Utilized a multi-head cross-attention block for learning ligand structure-properties.

Main Results:

• BiAtt-GVAE demonstrated effectiveness on EGFR and CDK2 targets.
• A case study on SARS-CoV-2 Mpro inhibitors showed the model's capability to generate novel compounds.
• Subsequent analysis confirmed high novelty and affinity ratings for generated Mpro inhibitors.

Conclusions:

• BiAtt-GVAE successfully generates novel bioactive molecules with high affinity for specific targets.
• The model's architecture effectively captures crucial interaction and structure-property information.
• BiAtt-GVAE shows significant promise for accelerating drug design and discovery.