Deep-learning-enhanced exploration of peptide conformational space with high fidelity using hydrogen bond information

Neural network potential models were trained using density functional theory (DFT) data for singly protonated hexapeptide, DYYVVR, previously studied through cryogenic ion spectroscopy and applied for its conformational analysis. A fragmentation-based approach was employed, in which the training datasets included capped dipeptides and capped single-residue clusters. The fragmentation approach effectively reduced energy prediction errors at reduced computational costs. To better capture a wider range of conformational space, all hydrogen bond types present in the peptide were included in the training dataset. As a result, the neural network potential model achieved a mean absolute error of 4.79 kJ mol^-1 in energy predictions compared to the DFT calculations. The model was further patched through an active learning scheme during basin-hopping simulations. The structures discovered during the simulations were optimized using the neural network model, leading to the identification of new conformational minima. The newly found structures successfully explained the experimental IR-UV depletion spectra obtained via cryogenic ion spectroscopy.