Data-Based Prediction of Redox Potentials via Introducing Chemical Features into the Transformer Architecture

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Molecular Property Prediction

Background:

• Accurate prediction of molecular physicochemical parameters is crucial for designing novel reactions and materials.
• Predicting redox potentials of organic molecules has been a significant challenge in computational chemistry.
• Existing methods often lack the accuracy and speed required for high-throughput molecular design.

Purpose of the Study:

• To develop a deep learning method, TransChem, for rapid and accurate prediction of redox potentials in organic molecules.
• To integrate spatial and electronic molecular features with advanced language modeling techniques.
• To demonstrate the model's generalizability across diverse organic datasets and its utility in accelerating chemical discovery.

Main Methods:

• Developed TransChem, a chemical language model incorporating deep learning.
• Employed effective molecular characterization combining spatial and electronic features.
• Utilized a nonlinear molecular messaging approach (Mol-Attention) and perturbation learning.

Main Results:

• TransChem achieved high accuracy (R² > 0.97, MAE < 0.09 V) in predicting redox potentials for over 100,000 organic radicals.
• Demonstrated excellent generalization on smaller datasets, including 2,1,3-benzothiadiazole (MAE < 0.07 V) and electron affinity (MAE < 0.18 eV).
• Successfully predicted oxidation potentials for a large dataset of disubstituted phenols using high-throughput and active learning strategies.

Conclusions:

• TransChem represents a state-of-the-art approach for redox potential prediction, outperforming existing benchmarks.
• The model's ability to integrate chemical knowledge into language modeling advances molecular design.
• This work facilitates accelerated screening of reagents for reactions like selective cross-coupling of phenol derivatives.