High-throughput electronic property prediction of cyclic molecules with 3D-enhanced machine learning

Area of Science:

• Computational Chemistry and Materials Science
• Machine Learning in Chemistry

Background:

• Existing molecular datasets lack structural diversity for complex organic molecules, hindering machine learning (ML) model generalizability.
• Complex organic molecules are crucial for bioactive compounds and functional materials.

Purpose of the Study:

• To introduce the Ring Vault dataset, a high-quality collection of diverse cyclic molecules.
• To develop and evaluate ML models for predicting key electronic properties of these molecules.
• To establish a framework for high-throughput screening and rational design of cyclic molecules.

Main Methods:

• Creation of the Ring Vault dataset containing 201,546 cyclic molecules across various structural types and elements.
• Quantum mechanical (QM) calculations on a subset of 36,000 molecules to generate property data.
• Training and evaluation of three ML models (Graph Attention Network, Chemprop, AIMNet2) for predicting electronic properties (HOMO-LUMO gap, IP, EA, redox potentials).

Main Results:

• The fine-tuned AIMNet2 model, utilizing 3D conformational information, achieved R² values > 0.95 and reduced mean absolute errors (MAEs) by over 30% compared to 2D models.
• Principal Component Analysis (PCA) of AIMNet2 embeddings revealed correlations between electronic properties and structural features.
• The Ring Vault dataset provides a robust foundation for molecular property prediction.

Conclusions:

• The developed methodology and Ring Vault dataset offer a robust framework for accelerating the discovery and design of functional cyclic molecules.
• This work facilitates high-throughput screening and rational design for applications in drug discovery, organic electronics, and energy materials.
• The study advances the understanding of complex structure-property relationships in organic chemistry.