Machine Learning Models for Local Optimization of Red Fluorescent Protein Variants in a Low-Data Setting.

Fluorescent proteins (FPs) are widely used reporters for visualizing cellular structures and processes. Traditional wet-lab strategies for FP engineering (rational design and directed evolution) have enabled substantial improvements in photophysical performance but are limited by their requirement for deep expert knowledge or labor-intensive screening. AI-driven approaches have recently gained traction for engineering variants of green FPs, yet applications to red fluorescent proteins (RFPs) remain scarce. Here, we demonstrate an application-focused approach using machine learning (ML) for the local optimization of RFP variants in a low-data setting. Using a data set of over 150 reported RFP sequences, we trained lightweight descriptor-based ML models to prioritize variants within the local sequence space of the state-of-the-art RFP mScarlet-I3. Guided by model predictions, we identified variants exhibiting red-shifted emission peaks, large Stokes shifts, or brightness comparable to the parental protein. Our findings show that these interpretable, data-efficient models can serve as effective auxiliary guides for targeted local engineering, and provide a practical framework for protein optimization.