COMA: efficient structure-constrained molecular generation using contractive and margin losses.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Structure-constrained molecular generation aims to create novel drug candidates with improved properties while retaining structural similarity to known molecules.
• Existing models struggle to balance high structural similarity with significant property enhancement.

Purpose of the Study:

• To develop a novel structure-constrained molecular generation model.
• To simultaneously achieve property improvement and high structural similarity in generated molecules.
• To address limitations in current molecular generation techniques for drug discovery.

Main Methods:

• A two-phase training approach for a molecular generation model.
• Phase 1: Metric learning with contractive and margin losses to learn molecular representations.
• Phase 2: Reinforcement learning for optimized molecular structure exploration and property improvement.

Main Results:

• The proposed model demonstrates superior performance compared to state-of-the-art baselines.
• Ablation studies confirm the effectiveness of the model's components.
• Successful application in generating sorafenib-like molecules for drug-resistant patients.

Conclusions:

• The developed model effectively balances structural similarity and property enhancement in molecular generation.
• This approach shows promise for accelerating drug discovery, particularly for complex cases like drug resistance.