Likelihood-based fine-tuning of protein language models for few-shot fitness prediction and design
Machine learning models trained on measurements of protein functional properties are widely used to accelerate laboratory-based protein design campaigns. To maximise the signal that can be extracted from limited experimental data, sequence embeddings produced by protein language models (PLMs) are often used as the basis of supervised fitness predictors. However, embedding-based predictors do not directly exploit the distributional information encoded in PLM likelihoods after self-supervised or generative pretraining on natural protein sequences. In contrast, likelihood-based fine-tuning approaches exploit this prior knowledge by directly updating pretrained PLM likelihoods to reflect observed fitness differences between sequences. While likelihood-based fine-tuning methods have been proposed previously, a conclusive comparison of their performance against state-of-the-art embedding-based methods has been lacking. To address this gap, we conduct a comprehensive empirical evaluation of both fine-tuning strategies on a representative set of protein fitness datasets from the ProteinGym benchmark. To ensure our evaluation is applicable across different PLM classes, we develop a simple, unified framework for likelihood-based fine-tuning that applies to models trained with various objectives. Across model classes and fitness datasets, likelihood-based fine-tuning consistently outperforms embedding-based methods previously reported as state-of-the-art, with the largest gains in low-data settings. Finally, to highlight the practical relevance of these findings, we demonstrate that the best-performing fine-tuning strategies can substantially improve the maximal fitness of designed sequences in multi-round in silico optimisation campaigns.