Machine Learning Optimization of non-Kasha Behavior and of Transient Dynamics in Model Retinal Isomerization

Abstract

Designing a model of retinal isomerization in Rhodopsin, the first step in vision, that accounts for both experimental transient and stationary state observables is challenging. Here, multi-objective Bayesian optimization is employed to refine the parameters of a minimal two-state-two-mode (TM) model describing the photoisomerization of retinal in Rhodopsin. With an appropriate selection of objectives, the optimized retinal model predicts excitation wavelength-dependent fluorescence spectra that closely align with experimentally observed non-Kasha behavior in the non-equilibrium steady state. Further, adjustments to the potential energy surface within the TM model reduce the discrepancies across the time domain. Overall, agreement with experimental data is excellent.