A Foundation Model for Accurate Atomistic Simulations in Drug Design

Neural network potentials now offer robust alternatives to electronic structure and empirical force fields computations for the on-the-fly production of the potential energy surfaces required in atomistic Molecular Dynamics (MD) simulations. However, widespread application in Chemistry and Biology faces several challenges: the need for fast inference and economical training; stringent model transferability requirements, particularly including charged-species interactions. Trained exclusively on synthetic quantum chemistry data, FeNNix-Bio1 sets a new standard for Foundation machine learning Models to provide predictive condensed-phase MD simulations including quantum nuclear effects. Its full-range of capabilities is demonstrated by modelling: water properties, ions in solution, small molecules hydration free energies, complex folding free-energy landscapes, large-scale protein dynamics, protein-ligand binding, chemical reactions and by coupling it to protein structure prediction foundation models' outputs for further refinement. FeNNix-Bio1 is accurate and systematically improvable while limiting human parametrization efforts: it is likely to have a strong impact in Drug Design.