Methyl Cation Affinity and Methyl Anion Affinity Prediction Using Uni-Mol-Based Models

Predicting nucleophilicity and electrophilicity at atomic sites in organic compounds is crucial for the design of polar reactions. Methyl cation affinity (MCA) and methyl anion affinity (MAA), calculated using quantum-chemistry-based simulations, are good indicators of nucleophilicity and electrophilicity, respectively. Machine learning surrogate models have been developed to accelerate inference for MCA and MAA. However, their prediction accuracy, even for state-of-the-art models, remains inadequate for practical use. We present accurate machine-learning surrogate models for MCA and MAA, achieving root-mean-square errors of 8.90 [kJ/mol] for MCA and 10.02 [kJ/mol] for MAA. The model architecture comprises a pretrained Uni-Mol encoder block and a feed-forward neural network. Without pretraining on massive conformers, models with the proposed architecture still perform comparably, suggesting its architectural superiority to a molecular graph-based neural network model. The conformation used to calculate MCA and MAA as the model input is found to slightly improve prediction accuracy. Furthermore, the proposed models are robust to underlying MCA/MAA calculation protocols. The inference time of the proposed MCA and MAA surrogate models is less than 0.1 s per compound on a single GPU, and data scarcity arising from expensive MCA/MAA calculation protocol can be overcome by a simple fine-tuning approach, enabling fast and accurate reactivity prediction for synthetic chemistry.