Transferring chemical and energetic knowledge between molecular systems with machine learning

Predicting the structural and energetic properties of molecular systems is crucial in molecular simulations, with applications spanning chemistry, biology, and medicine. This paper introduces a novel machine learning methodology for knowledge transfer from simple to complex molecular systems, focusing on classifying high and low free-energy conformations. The approach utilizes a novel hypergraph representation of molecules, encoding multi-atom interactions, and incorporates novel message passing and pooling layers for free-energy prediction. Transfer learning from tri-alanine to deca-alanine achieved a remarkable Area Under the Curve (AUC) of 0.92. The method also successfully grouped chemically related secondary structures of deca-alanine based on similar free-energy values. This work demonstrates the potential of reliable transfer learning models for predicting structural and energetic properties of biologically relevant systems.

Sajjad Heydari, Stefano Raniolo, Lorenzo Livi, Vittorio Limongelli