ChemFixer: Correcting Invalid Molecules to Unlock Previously Unseen Chemical Space

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery

Background:

• Deep learning models generate potential drug candidates but often produce chemically invalid molecules.
• Invalid molecules limit the utility of generated chemical spaces and hinder practical drug discovery applications.

Purpose of the Study:

• To introduce ChemFixer, a novel framework for correcting chemically invalid molecules into valid ones.
• To enhance the practical applicability of deep learning-based molecular generation.

Main Methods:

• ChemFixer utilizes a transformer architecture.
• The model is pre-trained with masking techniques and fine-tuned on a custom dataset of valid/invalid molecular pairs.

Main Results:

• ChemFixer significantly improved molecular validity across various generative models.
• The framework preserved the chemical and biological properties of generated molecules, expanding the diversity of drug candidates.
• Application to drug-target interaction prediction improved ligand validity and identified promising ligand-protein pairs, even with limited data.

Conclusions:

• ChemFixer effectively addresses the challenge of invalid molecules in deep learning-based drug discovery.
• The framework enhances molecular validity, expands accessible chemical space, and shows promise for data-limited scenarios and downstream tasks.