RIGR: Resonance-Invariant Graph Representation for Molecular Property Prediction.

Area of Science:

• Computational chemistry
• Machine learning for drug discovery
• Molecular graph representations

Background:

• Machine learning models for molecular property prediction often use SMILES strings, which struggle with molecules exhibiting resonance.
• Resonance variance leads to inconsistent predictions in frameworks like Chemprop, hindering accurate molecular property analysis.
• Existing methods lack a unified approach to handle the diverse resonance forms of a single molecule.

Purpose of the Study:

• To develop a novel molecular representation that is invariant to resonance structures.
• To improve the accuracy and consistency of machine learning-based molecular property predictions.
• To provide a general graph featurization scheme for diverse chemical tasks.

Main Methods:

• Introduction of the resonance-invariant graph representation (RIGR) for molecules.
• Implementation of RIGR within a directed message-passing neural network (D-MPNN) architecture.
• Evaluation on a large dataset including radicals and closed-shell molecules, comparing RIGR against the Chemprop featurizer.

Main Results:

• RIGR ensures all resonance structures map to a single representation, eliminating prediction variance.
• The RIGR featurizer achieved comparable or superior prediction performance using 60% fewer features than Chemprop.
• Alternative methods like data augmentation with resonance forms showed limitations.

Conclusions:

• RIGR effectively addresses the challenge of resonance variance in molecular property prediction.
• RIGR offers a more efficient and robust featurization scheme for machine learning models.
• RIGR is available as an open-source option in Chemprop, demonstrating broad applicability in chemical research.