Green fluorescent proteins: Examining the underlying factors of brightness using machine learning

Green fluorescent protein (GFP), originally discovered in the jellyfish Aequorea victoria, is a protein that emits green fluorescence when its internal fluorophore absorbs blue light. GFPs have high utility in scientific research as fluorescent markers that are used to visualize biological processes, structures, and interactions. Engineering of fluorescent proteins (FPs) generated color variants through mutation of the wild-type sequence that subsequently shifted excitation wavelength. However, it is poorly understood how sequence mutations influence fluorescence at 405 nm versus 488 nm, which represent the two predominant excitation peaks of GFP and related proteins. To elucidate how sequence mutations shape the GFP fluorescence spectra, we developed a novel hybrid neural network model that combines a black-box deep network with a biochemical model of fluorescence, including parameters for protein folding, quantum yield, and fluorophore pKa to predict fluorescence intensity from the amino acid sequence. We trained the model on a published dataset consisting of paired sequence-function measurements for thousands of FP variants which critically included excitation at both 405 and 488 nm. The model accurately predicted fluorescence measurements at each excitation wavelength from sequence alone. The interpretability of the model parameters allows for inference and assessment of the biochemical factors underlying shifts in excitation wavelength. Further improvements can be made by extending the model to also fit in vitro measurements.