PURE: policy-guided unbiased REpresentations for structure-constrained molecular generation.

Area of Science:

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

Background:

• Deep learning for structure-constrained molecular generation (SCMG) faces challenges including predisposition to existing data, discrete-continuous space mismatch, and metric leakage.
• Existing methods often struggle with the inherent discrete nature of molecules and can be biased by task-specific evaluation metrics during training.

Purpose of the Study:

• To introduce Policy-guided Unbiased REpresentations (PURE), a novel framework for SCMG that addresses current deep learning limitations.
• To develop a method that learns high-quality, unbiased molecular representations by simulating molecular transformations.
• To improve the exploration of discrete molecular search spaces for drug synthesis applications.

Main Methods:

• PURE employs a combination of self-supervised learning and a policy-based reinforcement learning (RL) framework.
• The approach simulates molecular transformations within a drug synthesis context, avoiding reliance on external molecular metrics.
• Template-based molecular simulations and a semi-supervised training design are utilized to navigate the discrete molecular search space.

Main Results:

• PURE achieves competitive or superior performance compared to state-of-the-art methods across multiple benchmarks, despite lacking metric biases.
• The framework learns high-quality representations with an inherent notion of task-specific similarity.
• The study demonstrates the successful application of PURE in identifying sorafenib-like compounds to combat drug resistance.

Conclusions:

• Current deep learning approaches for SCMG require reevaluation, highlighting the need for methods that naturally align with molecular generation problems.
• PURE provides a robust and unbiased framework for SCMG, demonstrating its potential in drug discovery and development.
• The methodology shows promise for identifying novel drug candidates, exemplified by its application in combating drug resistance.