Multi-objective drug design with a scaffold-aware variational autoencoder.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Medicinal chemistry

Background:

• Designing molecules with multiple desired properties is a significant challenge in drug development.
• Existing fragment-based methods have limitations in chemical space accessibility and validity.

Purpose of the Study:

• To develop an innovative scaffold-aware variational autoencoder (ScafVAE) for in silico graph-based generation of multi-objective drug candidates.
• To expand the accessible chemical space beyond conventional fragment-based approaches while ensuring high chemical validity.

Main Methods:

• ScafVAE integrates bond scaffold-based generation with perplexity-inspired fragmentation.
• The model was pre-trained on a large molecular dataset and enhanced with contrastive learning and molecular fingerprint reconstruction.
• ScafVAE was adapted to generate dual-target drug candidates against cancer drug resistance mechanisms, with optional properties like drug-likeness and toxicity.

Main Results:

• ScafVAE demonstrated high accuracy in predicting molecular properties.
• Generated dual-target drug candidates showed strong binding affinity to target proteins (verified by docking and experimental measurements).
• Molecular dynamics simulations confirmed stable binding interactions, and generated molecules maintained optimized extra properties.

Conclusions:

• ScafVAE effectively expands chemical space and generates valid multi-objective drug candidates.
• The approach is adaptable to new desired properties and shows potential for developing targeted cancer therapies.
• ScafVAE presents a promising alternative to traditional drug generation methods.