Enhanced prediction of molecular properties using transfer learning on sigma profiles

Abstract The performance of machine learning techniques for the prediction of a wide range of molecular properties has seen rapid improvements in recent years due to developments in both molecular representations and deep learning modeling techniques. Sigma profiles, which are computational descriptors representing the surface charge distribution of molecules, have shown promise as a molecular representation to support robust property prediction. Meanwhile, large-scale pretrained deep learning models based directly on molecular structure inputs, such as Uni-Mol, have demonstrated strong performance as general-purpose molecular representation learners. In this study, we seek to enhance the prediction of molecular properties by integrating information from sigma profiles with these advanced deep learning techniques. Our methodology involves fine-tuning the Uni-Mol model to accurately predict sigma profiles, which capture detailed molecular structural information important for determining molecular interactions. We then utilize transfer learning to apply the learned weights to predict specific molecular properties, replacing the final output layer to adapt to each new task. The results demonstrate improvements in predictive accuracy across various datasets, showcasing the effectiveness of combining sigma profiles with state-of-the-art machine learning models and demonstrating a path forward for leveraging theory-driven descriptor development to enhance large-scale data-driven molecular property modeling.