One Protein Is All You Need

This video summarises the paper “One Protein Is All You Need” (2025, arXiv:2411.02109). Protein language models such as ESM2 and SaProt are trained to perform well on average across vast sequence collections. But biologists rarely care about average-case performance. In practice, a researcher often needs accurate structure, fitness, or function predictions for one specific protein, often one that is out-of-distribution or poorly represented in the training set. This paper introduces a simple but surprisingly powerful idea to close that gap: Protein Test-Time Training (ProteinTTT), a method that customises a protein model to a single sequence at inference time, without any additional data.