Using *Stoic*-predicted stoichiometries improves downstream structure prediction - including accuracy on CAMEO targets and metrics like iLDDT and QS-global - while scaling efficiently to large complexes where brute-force enumeration becomes computationally infeasible. (5/10)
Daniil Litvinov
Big thanks to my co-authors: @lorenzopantolini.bsky.social, @peterskrinjar.bsky.social, @swiss-model.bsky.social , @computingcaitie.bsky.social, @cellarchlab.com, @torstenschwede.bsky.social, the rest of Schwede & Engel groups, and particularly @ninjani.bsky.social. (10/10)
Because *Stoic* excels at interface residue classification, we can reuse these outputs as an interpretability signal for predicted structures and a confidence check to verify AlphaFold3 models used the correct stoichiometry. (7/10)
Why it matters: knowing the correct stoichiometry - the number of copies of each protein entity in a complex - is a prerequisite for structure prediction, but current combinatorial approaches are often slow and computationally expensive. (2/10)
*Stoic* uses protein language model embeddings, learns interface‑aware residue features, and feeds them into a graph neural network to capture complex‑level context. (3/10)
Daniil Litvinov
Try *Stoic* yourself! GitHub: github.com/PickyBinders...
Web version (Hugging Face Space): huggingface.co/spaces/Picky...
Colab: colab.research.google.com/github/Picky...
(9/10)
In summary, *Stoic* makes stoichiometry prediction easy and fast directly from sequence, scales to large complexes, boosts AF3 structural prediction, returns ready-to-use AF3 input JSON, and provides interpretable interface residue signals as a confidence bonus. (8/10)
Daniil Litvinov
I'm excited to share *Stoic*, a method for fast and accurate protein complex stoichiometry prediction directly from sequence. Preprint: www.biorxiv.org/content/10.6... 🧵👇(1/10)
*Stoic* achieves high sequence‑level and stoichiometry accuracy across both homomeric and heteromeric complexes, outperforming both homology‑based (SWISS-MODEL) and deep learning-based (Seq2Symm) approaches. (4/10)
Daniil Litvinov
Ablations show that all three main components - GNN architecture, weighted embeddings pooling, and the auxiliary loss - contribute meaningfully to performance. *Stoic* also delivers solid interface residue predictions. (6/10)
