A machine learning-based platform for predicting the potential energy of small molecules

Accurate estimation of molecular potential energy after energy minimization is vital for the rational design of targeted therapeutics. In this study, we present a machine learning platform that leverages sophisticated neural network architectures to predict the potential energy of small molecules. Accurate prediction of molecular potential energy can facilitate the computational characterization of drug- and gene-delivery-related molecules and support molecular modeling workflows. Our methodology involves designing and evaluating deep learning models, including convolutional neural network–bidirectional long short-term memory architectures, with and without multihead attention mechanisms. These models are trained on a comprehensive dataset comprising small molecules structures. By utilizing atom-wise sequential representations that preserve key geometric and chemical information, our approach effectively estimates potential energies and surpasses traditional descriptor-based machine learning methods. The integration of attention mechanisms enhances model generalization by capturing complex structural relationships, thereby streamlining virtual screening processes. Comparative analyses against existing benchmarks demonstrate the efficacy of sequence-based neural networks in molecular energy prediction, paving the way for more efficient drug discovery and polymer informatics. This platform offers a robust, computationally efficient tool for virtual screening and energy estimation of candidate molecules, supporting the advancement of targeted therapeutics through rapid molecule prioritization based on predicted potential energies.