Home Research Services
Online Services
campaignUpdated October 2026: new co-assembly and homo-oligomer scores with adjustable thresholds, and an optional PDB template search. What’s new / release history

info About StoPred

StoPred is a deep learning-based tool that predicts the stoichiometry of protein complexes from sequence alone, using a combination of protein language model embeddings and a graph attention neural network.

StoPred can handle both homomeric and heteromeric complexes, without requiring template assemblies or predefined stoichiometry; an optional template search can be added for comparison. The model outputs the most probable stoichiometry and can suggest a small set of candidates to guide downstream structure prediction.

StoPred achieves state-of-the-art performance compared to Template-based and AlphaFold ranking score-based methods on large-scale benchmark datasets.

StoPred’s stoichiometry prediction assumes that the submitted proteins form a complex. We recommend submitting proteins with experimental support for complex formation. Where such evidence is not available, StoPred also reports assembly-state scores (see below) to help assess whether the input is likely to assemble, but these are model estimates and do not replace experimental support.

upload_file Submit Protein Sequences

The date is the training cutoff (latest PDB release in the training data). Use the manuscript release only to reproduce the paper.
One record per distinct protein, each with a unique header.
join_innerTwo-chain complex
2 sequences
View result
hubLolCDE–Pal (PDB 9RLF)
4 sequences
View result
looks_oneSingle chain
1 sequence
View result
optional
Searches PDB biological assemblies (snapshot 2026-05-05) and lists each template’s copy numbers, structure and alignments. Leave the date empty to use all entries, or set one (e.g. a model’s training cutoff) to keep only earlier templates.

rule What StoPred Reports

bar_chartStoichiometry (every job)
Ranked copy numbers for each submitted sequence, e.g. A2B1 for two copies of chain A and one of chain B. This prediction assumes the sequences form a complex.
looks_oneOne sequence
A homo-oligomer score from a monomer-aware model: the probability of any copy number other than 1. Drag the threshold to see its recall, precision and false-positive rate.
hub2–6 distinct sequences
A co-assembly score: how likely the sequences occur in one assembly. For a pair, drag the threshold to see the trade-off; for 3–6 sequences, every pair and every leave-one-out subset is scored too.
travel_exploreTemplates (optional)
PDB biological assemblies similar to your chains, with each template’s copy numbers, the 3D structure and the alignments.
  • Submit each distinct sequence once. For a homo-oligomer, submit one record; StoPred predicts its copy number. A sequence repeated under two headers is not accepted.
  • At most six distinct sequences per job.
  • Scores are model estimates from sequence alone, not experimental evidence. Recall, precision and false-positive rate come from our evaluation data; sets of 3–6 proteins have no curated negatives, so no such rates are given for them.

download Download

download
Download Page
Container image, software package, model releases, and model naming documentation
Open Downloads
storage
Model Releases
Available StoPred model archives and model naming documentation
View Models

menu_book Citation

Liu, Q., Peng, C., Zheng, W., Zhang, C., & Freddolino, L. (2025). StoPred: Accurate Stoichiometry Prediction for Protein Complexes Using Protein Language Models and Graph Attention. bioRxiv. https://doi.org/10.1101/2025.10.20.683515

FreddolinoLabumich.edu | 1150 W. Medical Center Dr, Ann Arbor, MI 48109-0600