Benchmarking Molecular Methods of Protein Engineering and Surveying Protein Search Space to Bolster Artificial Intelligence-Based Designs

Abstract

Proteins are complex macromolecules that drive functionality in cells across all life on Earth. Protein engineering and AI-based designs facilitate exploration of the vast search space. Herein, we explore and compare mutagenic methodologies in combination with fast and efficient phenotypic evaluation of a pink fluorophore (Magenta) to map protein search space and inform AI-based designer algorithms. We binned mutated variants as neutral, deleterious, or catastrophic based on the observed phenotype and fitness measurement (relative fluorescent units, RFUs). Importantly, we found that for a given fluorophore, about 80% of random mutations were deleterious for the artificially selected phenotype resulting in decreased fitness, 20% were neutral mutations, and 0.8% enhanced the artificially selected phenotype. Using error prone PCR (epPCR), we were able to identify four candidates in the first round and one candidate in the second round of epPCR for enhanced fluorescence and in toto enhanced RFUs by 35% after two rounds of epPCR. We compared these rates of true positive enhanced RFU generated sequences against a state of the art evolutionary algorithm approach and found that biological computation beats it in its current state. We conclude with suggestions for key improvements to enhance algorithm performance in the future. Taken together, we demonstrated that iterative rounds of variant prediction improved our algorithm, and our study can serve as a reference point for future AI driven modeling studies in red fluorescent protein engineering.