Bridging the phenotype-target gap for molecular generation via multi-objective reinforcement learning.

Area of Science:

• Artificial intelligence in drug design
• Computational chemistry
• Molecular generation

Background:

• AI-driven drug design faces challenges with current phenotype-based and target-based strategies.
• Phenotype-based methods incur high costs, while target-based methods overlook cellular responses.
• A unified approach is needed to bridge these limitations in de novo molecular generation.

Purpose of the Study:

• To introduce XMolRL, a novel generative framework for de novo molecular generation.
• To synergistically integrate phenotypic and target-specific cues for improved drug candidate generation.
• To address the limitations of existing AI-driven drug design strategies.

Main Methods:

• Pretraining a phenotype-guided generator on drug-induced transcriptional profiles.
• Fine-tuning using multi-objective reinforcement learning (RL) with a fused reward function.
• Incorporating docking affinity, drug-likeness, ranking loss, regularization, and entropy maximization.

Main Results:

• XMolRL demonstrates superior performance compared to state-of-the-art models.
• Generated molecules exhibit favorable drug-like properties, high target affinity, and potent inhibition (IC50) against cancer cells.
• The framework successfully steers towards chemotypes that are potent, diverse, and phenotypically aligned.

Conclusions:

• XMolRL offers a more effective solution for de novo drug discovery by combining phenotype-guided and target-aware strategies.
• The unified framework highlights the synergistic potential of integrating diverse data sources.
• This approach advances AI-driven drug design by generating high-quality candidate molecules.