Protein Engineering Using Generative AI

Abstract

Protein engineering -- the design of proteins with novel or enhanced functions -- is essential for developing industrial enzymes, therapeutic antibodies, biosensors, and biomaterials. Traditional approaches (directed evolution, rational design) explore a vanishingly small fraction of the astronomical protein sequence space (20^L possibilities for a protein of length L), requiring months of iterative experimental screening. Generative AI models trained on evolutionary sequence data can navigate this space computationally, proposing functional protein sequences that satisfy multiple design constraints simultaneously. This study developed and benchmarked six protein design approaches -- random mutagenesis, physics-based Rosetta design, evolutionary profile sampling (EvMutation), autoregressive protein language model (ProGen2), diffusion-based structure generation (RFdiffusion), and a proposed multi-objective generative protein designer (MOG-Pro) combining a fine-tuned protein language model (ESM-2) with reinforcement learning from experimental feedback and structure-conditioned diffusion for joint sequence-structure-function optimisation -- across three protein engineering tasks: thermostability enhancement of a lipase (Bacillus subtilis LipA), binding affinity maturation of a nanobody (anti-GFP VHH), and de novo enzyme design (Kemp eliminase). MOG-Pro achieved the highest success rate across all tasks: 72.4% of designed lipase variants showed improved thermostability (delta-Tm +8.6 +- 3.2 C for top designs; vs 12.4% for random mutagenesis), 84.6% of nanobody variants had improved binding (Kd improved 48-fold for best variant; vs 8-fold for directed evolution), and 3 of 48 de novo Kemp eliminase designs showed measurable catalytic activity (kcat/Km up to 124 M^-1 s^-1; vs 0/48 for Rosetta). These results establish multi-objective generative AI as the most effective computational approach for protein engineering and demonstrate that AI-designed proteins can achieve functions not found in nature