Here, customization primarily benefits challenging, out-of-distribution proteins that are poorly represented in sequence databases (as measured by MSA size)

For example, ProteinTTT applied to ESMFold improves 19% of AlphaFold2-predicted viral protein structures in BFVD

This consistently improves performance of various protein language models across protein structure, fitness and function prediction, particularly on challenging targets

ProteinTTT enables *customizing* protein language models to one target protein at a time without assuming any additional data via on-the-fly self-supervised fine-tuning on the single protein

We train machine learning models on millions of proteins. But when it comes to making predictions, do we need them to understand all proteins at once? Often, we need an accurate model for the specific protein we are studying or designing. We address this with ProteinTTT arxiv.org/abs/2411.02109 1/🧵