Physics-Based Machine\nLearning Trains Hamiltonians\nand Decodes the Sequence–Conformation Relation in the Disordered\nProteome

Intrinsically disordered proteins and regions (IDPs)\nare involved\nin vital biological processes. To understand the IDP function, often\ncontrolled by conformation, we need to find the link between sequence\nand conformation. We decode this link by integrating theory, simulation,\nand machine learning (ML) where sequence-dependent electrostatics\nis modeled analytically while nonelectrostatic interaction is extracted\nfrom simulations for many sequences and subsequently trained using\nML. The resulting Hamiltonian, combining physics-based electrostatics\nand machine-learned nonelectrostatics, accurately predicts sequence-specific\nglobal and local measures of conformations beyond the original observable\nused from the simulation. This is in contrast to traditional ML approaches\nthat train and predict a specific observable, not a Hamiltonian. Our\nformalism reproduces experimental measurements, predicts multiple\nconformational features directly from sequence with high throughput\nthat will give insights into IDP design and evolution, and illustrates\nthe broad utility of using physics-based ML to train unknown parts\nof a Hamiltonian, rather than a specific observable, in combination\nwith known physics.