Contrastive training lands a protein's sequence and its structure in the same spot on the map

Sequence and structure do not track each other cleanly. Proteins with unrelated sequences sometimes adopt similar structures, and similar or even identical sequences can produce very different ones ^[4]. Most protein language models have handled the two separately, and even the models that take both in do not necessarily put the sequence and the structure of one protein at the same location on a protein map ^[6].

CLSS is trained to close that gap directly. During training the model receives sequences and their corresponding structures, and it learns to produce similar embeddings for the true sequence-structure pairs while separating unrelated pairs ^[5]. Co-location is therefore the objective the model was optimized for.

That is the second result, and it is the one carrying weight. The CLSS representations closely reproduced relationships recorded in ECOD and CATH, two expert-curated classification systems, neither of which was provided to the model during training ^[8]. Those hierarchies hold decades of scientific knowledge ^[9]. A model that lands on them unprompted has learned something about relatedness.

Then there are the fragments. Most existing protein language models need a complete sequence or structure to produce a meaningful representation; CLSS in many cases positioned short sequence fragments meaningfully alongside complete sequences and structures ^[10]. The team notes that fragments are particularly important for understanding evolution ^[11].

The work was led by Rachel Kolodny and the PhD candidate Guy Yanai at the University of Haifa, Nir Ben-Tal and Gabriel Axel at Tel Aviv University, and Liam Longo at the Earth-Life Science Institute in Tokyo, where Kolodny spent five months as a visiting researcher developing ways to analyze the model ^[2][15]. "This gives us a way to look at the protein universe through sequence and structure at the same time, rather than treating them as separate worlds," said Longo ^[12].

The thing this doesn't tell you is whether the shared map recovers homology that sequence-only models miss. Longo puts that in the conditional: "What is particularly exciting for us is the possibility of using these maps to uncover large-scale evolutionary patterns that are difficult to recognize using conventional approaches" ^[13]. The phys.org account reports that CLSS performed strongly in classification tests and that it brought sequence and structure into a cohesive map when compared with other state-of-the-art models, without naming a benchmark, an accuracy or a baseline ^[14][17].

So far, the redrawing of protein family relationships remains a prospect. Reproducing ECOD and CATH tests the map against relationships curators have already recorded. That is the right first test for a new representation. The evolutionary payoff would be a relationship they have not recorded, and this report presents that as a possibility.