BPS2026 – Unraveling spectral shifts in microbial rhodopsin through machine learning predictions and molecular dynamics simulations

Microbial rhodopsins are light-sensitive proteins that mediate ion transport, phototaxis, and photoswitching in microorganisms and are widely used as optogenetic tools. These proteins operate most efficiently at their absorption maximum (λ max ). Accurate prediction of λ max enables rational tuning of these proteins to function at desired wavelengths, advancing applications in neuroscience and bioengineering. However, due to rhodopsin’s structural complexity, determining λ max experimentally has been challenging, time-consuming, and laborious. Existing predictive models have limited accuracy, and no explainable model currently exists that provides mechanistic insights into the molecular basis of spectral shift. To address these challenges, we developed a pipeline that leverages machine learning, feature importance analysis, and molecular dynamics simulations to gain mechanistic insights. First, we trained several classical machine learning regressors on 878 microbial rhodopsin sequences to accurately predict the λ max . With the best-performing model, we achieved a coefficient of determination of 0.9 on the test set. Then SHAP-based feature importance analysis of the best performing model identified key residues near chromophore responsible for spectral shift. From that, we selected mutations at two key residue sites: S254M that causes a bathochromic (red) shift, and G153V responsible for hypsochromic (blue) shift, to obtain mechanistic insights using molecular dynamics simulations. Analyses revealed that the pocket in S254M was more open and hydrated, causing the chromophore to adopt a planar geometry, whereas G153V resulted in a compact pocket that imposed steric strain and twisted the chromophore. These features could be further correlated with quantum mechanics to explain the mechanistic basis of spectral shift more robustly, particularly by examining conjugation, chromophore delocalization, and energy gaps. Our sequence-based predictive and explainable framework will enable rational engineering of microbial rhodopsins and potential extension to human opsins for vision science.