Designing diverse and high-performance proteins with a large language model in the loop.

We present a protein engineering approach to directed evolution with machine learning that integrates a new semi-supervised neural network fitness prediction model, Seq2Fitness, and an innovative optimization algorithm, biphasic annealing for diverse and adaptive sequence sampling (BADASS) to design sequences. Seq2Fitness leverages protein language models to predict fitness landscapes, combining evolutionary data with experimental labels, while BADASS efficiently explores these landscapes by dynamically adjusting temperature and mutation energies to prevent premature convergence and to generate diverse high-fitness sequences. Compared to alternative models, Seq2Fitness improves Spearman correlation with experimental fitness measurements, increasing from 0.34 to 0.55 on sequences containing mutations at positions entirely not seen during training. BADASS requires less memory and computation compared to gradient-based Markov Chain Monte Carlo methods, while generating more high-fitness and diverse sequences across two protein families. For both families, 100% of the top 10,000 sequences identified by BADASS exceed the wildtype in predicted fitness, whereas competing methods range from 3% to 99%, often producing far fewer than 10,000 sequences. BADASS also finds higher-fitness sequences at every cutoff (top 1, 100, and 10,000). Additionally, we provide a theoretical framework explaining BADASS's underlying mechanism and behavior. While we focus on amino acid sequences, BADASS may generalize to other sequence spaces, such as DNA and RNA.