Deep Learning Prediction of Molecular Conformational Energies from Off-Equilibrium Organic Molecular Structures

Abstract

Accurate prediction of molecular conformational energies is essential for understanding structural stability and accelerating computational workflows in chemistry, drug discovery, and molecular materials research. This study developed a deep learning framework to predict the relative conformational energies of off-equilibrium organic molecular structures using molecular descriptors and geometric features. A synthetic dataset comprising 1,440 off-equilibrium conformers generated from 110 chemically distinct organic molecules was employed. The dataset included physicochemical descriptors, structural distortion parameters, and conformational energy values, and was partitioned into independent training, validation, and testing subsets at the molecule level to ensure unbiased model evaluation. Prior to model development, data preprocessing eliminated target leakage, standardized numerical descriptors, and retained only independent structural variables. A fully connected deep neural network was trained using the Adam optimization algorithm and evaluated with mean absolute error, mean squared error, root mean square error, coefficient of determination, and Pearson correlation coefficient. The proposed framework achieved strong predictive performance across all evaluation datasets, with test results demonstrating an RMSE of 12.94 kcal/mol, an R² value of 0.894, and a Pearson correlation coefficient of 0.947. Feature importance analysis identified bond-length root mean square error, displacement root mean square deviation, and molecular size descriptors as the principal contributors to conformational-energy prediction. These findings demonstrate that deep learning can effectively capture nonlinear relationships between molecular geometry and conformational energy, providing a computationally efficient alternative for large-scale conformational analysis and molecular screening.