Learning to Evolve Structural Ensembles of Unfolded and Disordered Proteins Using Experimental Solution Data

Abstract

We have developed a Generative Recurrent Neural Networks (GRNN) that learns the probability of the next residue torsions $X_{i+1}=\ [\phi_{i+1},\psi_{i+1},\omega _{i+1}, \chi_{i+1}]$ from the previous residue in the sequence $X_i$ to generate new IDP conformations. In addition, we couple the GRNN with a Bayesian model, X-EISD, in a reinforcement learning step that biases the probability distributions of torsions to take advantage of experimental data types such as J-couplingss, NOEs and PREs. We show that updating the generative model parameters according to the reward feedback on the basis of the agreement between structures and data improves upon existing approaches that simply reweight static structural pools for disordered proteins. Instead the GRNN "DynamICE" model learns to physically change the conformations of the underlying pool to those that better agree with experiment.