Answers are grounded in this paper and cite their sources.
Predicting the optical properties of proteins is critical for the understanding and design of photoactive proteins. Red-shifting flavin-dependent photoenzymes’ spectra beyond the usual blue range has become an attractive goal for potential industrial use of photoenzymes for chemical synthesis. However, making such predictions using quantum chemical methods can be computationally intractable, especially if the correct description of the photoexcitation requires multireference methods. We present a sequence-transferable machine learning framework, LIGand-based PROtein-agnostic SPECTral prediction (LIG-PROSPECT), to predict the absorption spectra of flavin-dependent photoenzyme variants in the presence of specific substrates. We make use of structural descriptors of cofactor-substrate ensembles as a proxy to represent the 3D conformational features of protein sequences to enable the classification of cyan-light and red-light-absorbing peaks for different conformations and the prediction of spectral features of the proteins using linear regression, offering resource-efficient alternatives to multireference quantum mechanical calculations for screening protein sequences for desired photophysical properties, validated with variants of a flavoprotein, Gluconobacter ene-reductase (GluER). With models trained using data from only two protein variants, our approach allowed us to make predictions for a protein variant that represents an extrapolation beyond the training set.
No takes yet. Share an insight, caveat, or question.
Jayee et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: