Skip to main content
Back to timeline
bioRxivSource publication:

StoiAgent uses an AI agent to work out how many copies each subunit has: 75.0% top-1 across 56 blind CASP17 complexes

Synopsis

The copy number of each subunit in a protein complex, its stoichiometry, is often unknown, yet structure predictors such as AlphaFold3 need it as input. StoiAgent turns the open-weight Qwen3.6-27B model into a two-stage AI agent: stage one retrieves PDB, UniProt and literature evidence and combines it with structural templates to generate and rank candidate stoichiometries, and stage two reranks the top five using AlphaFold3 scores while keeping reasoning traces. On 56 CASP17 targets it reaches 75.0% top-1 accuracy, 23.2 percentage points above PreStoi with AlphaFold3 ranking, matching StoPred, with 85.7% top-3 and 87.5% top-5.

Source-provided article image: Accurate and interpretable protein complex stoichiometry prediction through agentic AI with StoiAgent
Fig. 1 · bioRxiv · Page 3

Interpretation

A protein complex's stoichiometry can now be inferred automatically by an AI agent that leaves a checkable record of the evidence behind each prediction. The earlier PreStoi also used homologous structural templates and AlphaFold3 ranking scores, but integrating evidence from different sources and resolving conflicting predictions still required human judgment, limiting automation and scalability. On 56 CASP17 complex targets, StoiAgent stage two reaches 75.0% top-1 accuracy (42/56), 23.2 percentage points above PreStoi at 51.8% (29/56) and level with StoPred at 75.0%.

The agent works in two steps: it first proposes candidate stoichiometries, then reranks a fixed shortlist using structural scores. In stage one the language model retrieves evidence from PDB, UniProt and PubMed, combines it with sequences and structural templates to generate candidates with support scores, and passes the top five to stage two. Adding AlphaFold3 ranking scores in stage two raises top-1 from 58.9% (33/56) in stage one to 75.0% (42/56), fixing ten errors and adding one; ranking the same candidates by maximum or mean AlphaFold3 score alone gives only 60.7% and 62.5% top-1.

Perspective

The evaluation is limited to 56 CASP17 multimer targets (43 heteromeric and 13 homomeric, spanning 34 biological systems); predictions were made after the official CASP17 experiment concluded and used only evidence available before each target's release date, making this a retrospectively blind test rather than a prospectively blinded assessment.

The authors note that the evaluation covers only 56 targets across 34 biological systems, with some systems contributing multiple targets, so larger and more diverse benchmarks plus uncertainty estimates accounting for biological-system dependence are needed; the contribution of individual evidence sources to specific ranking decisions is not systematically isolated; and computational cost comes mainly from running AlphaFold3 structure prediction for each candidate stoichiometry.