I’m presenting this work at the EMBO Computational Structural Biology Workshop in Heidelberg #EMBOComp3D this week, and @workshopmlsb.bsky.social in Copenhagen over the weekend. Let’s connect!

Huge thanks to my co-authors: @lauraengist.bsky.social, @ievapudz.bsky.social @martinsteinegger.bsky.social, @torstenschwede.bsky.social, especially to @ninjani.bsky.social! Couldn't have done this without the whole team, including the Swiss-Model development team and the rest of the Schwede group.

Try it out yourself! github.com/PickyBinders/tea. A web-service for search is coming soon at alphabet.scicore.unibas.ch.

Ultimately, TEA brings deep learning representation to protein sequence bioinformatics algorithms, such as profiles, phylogenetic trees, motif finding, multiple sequence alignments, and more, all while maintaining the speed and low resource consumption of amino acid sequences. (6/n)

We used TEA to connect >1.5 million singletons in AFDB Clusters, proteins which slipped past structure-based clustering approaches due to disordered or repetitive regions or simply because of low confidence structure predictions. (5/n)

TEA sequences come with a built-in confidence metric in the form of Shannon entropy, which we saw correlates with pLDDT, and can be used to filter out uncertain predictions. (4/n)

Running MMseqs2 with TEA gives extremely fast and highly sensitive results, similar to structural searches, even on unseen folds! Check out our ablations to see how we ended up with the final architecture. (3/n)

By using a contrastive objective, we trained an alphabet enriched with structural information, without the need for the actual structure. This approach ensures that remote homologs expressed with TEA maintain high sequence identity. (2/n)