Join us Wednesday December 10th for an amazing seminar by @ChoYehlin to cap off 2025. See you at 7pm EST in Room 6055, Longwood Center @DanaFarber "How AF3-Style Structure Prediction Models Can Be Used for Protein Design: BoltzDesign and Protein Hunter" https://t.co/E8bDyGzirb

Another example showing how a common SH3 fold may only match at the CATH "class" level. If you're splitting your train/test by topology...you've got data leakage 🚰

One example from CIRPIN ( https://t.co/SMO3AOX9xe) where two very similar structures have entirely different CATH classifications

@niopeklab We've also developed a Colab Notebook where you can try out CIRPIN for yourself! If you're looking for remote homologs that may be related by CP, insertions, extensions, rewirings, or other rearrangements, try it out and let us know how it worked! https://t.co/hiONSCJWE9

Excited to share this work with @yoakiyama @ChoYehlin @jajoosam @sokrypton We find that protein language models trained solely on individual protein sequences, implicitly learn the interface contacts of homo-oligomeric assemblies! As the model scales up, more interface signals

🚀Sergey Ovchinnikov's paper is here!! Can a deep learning model be trained to recognize proteins with identical 3D structures but different sequence connectivity, revealing thousands of hidden evolutionary relationships? "CIRPIN: Learning Circular Permutation-Invariant

@niopeklab If you're interested in circular permutation and remote homolog search, there's more to the story here:

@niopeklab In addition to CPs, CIRPIN allowed us to uncover more complex topological rearrangements. We identified cases of rewiring, where the connectivity of secondary structures differs, as well as pairs of similar proteins obscured by insertions/extensions:

@niopeklab Investigating the node-level embeddings of CIRPIN/Progres revealed that CIRPIN captures tertiary motifs. CIRPIN identifies regions of similarity within proteins that Progres missed. 3cbn is an interesting example since the two halves of the protein are nearly identical:

@niopeklab Shifting to the model interp side of things: PCA of CIRPIN/Progres embeddings shows how CIRPIN groups CPs of the same structure together. This can be thought of as a "folding" of the embedding space:

@niopeklab There's a lot of interesting evolutionary q's to dive into here, saved for another thread. But it's worth highlighting how there's been decades of very detailed work on PDZ domains. How might knowledge of these four CPs change what we know about PDZ form and function?

@niopeklab We then used CIRPIN/Progres to investigate the PDZ topology in the AFDB cluster representatives and found that PDZs exist in 4 circularly permuted forms:

Among the pairs we discovered in SCOP, were PDZ domains, recently reported to be the most frequently inserted domains in the AFDB by the @niopeklab

We could then use a contrastive approach to identify novel CPs by searching for pairs with high CIRPIN scores and low Progres scores

Training with synCPs allows our model, CIRPIN, to find similarity between known cases of circular permutants which Progres previously failed to identify:

Think of how comma placement in English can drastically change the meaning of a sentence. A panda can be a cute docile animal or a serial killer, depending on your inclusion of a comma.

Under the hood, synCP generation is simply shifting the positional information of each structure. But the consequences are significant.

What happens during training is that random synCPs of input structures are generated, forcing the model to learn that structures related by CP are similar:

Building off the Progres model by @jgreener64 @jamaliki1998, we introduced a novel data augmentation strategy using synthetic circular permutations (synCPs)

We wondered if there was a way to leverage the speed of deep learning based search tools, with the sensitivity of traditional structural alignment.