structural-biology/ai produced the result/Frontiers in Bioinformatics 2026 · v2
Software reads AlphaFold predictions to infer how protein complexes assemble
Researchers built MultimerMapper, a pipeline that compares protein structures predicted by AlphaFold2-multimer and AlphaFold3 when the same proteins are modelled in different combinations, to infer which proteins bind, in what numbers, and how their shapes shift.
spectrum · one line per step, placed by what the step does · bright lines used AI
Context-dependent protein structure prediction analysis and complex stoichiometry inference with MultimerMapper
Frontiers in Bioinformatics, 2026
doi:10.3389/fbinf.2026.1914731 · record aix-00277 v2 · checked 2026-11-04
- AI was for
- Structure determination
- Model family
- Transformer
- Checked by
- Held-out20 tested
- Code
- available
The finding the paper is about came from the AI.
What this research was about
Most proteins do not work alone. They stick together into complexes, and the number of copies of each partner in a finished complex, its stoichiometry, matters for how it works. A protein's shape can also change depending on who it is bound to, so the same chain may look different in different company. Working this out in the laboratory is slow, because each complex has to be purified and imaged or crystallised. Computer programs that predict a protein's three-dimensional shape from its amino acid sequence can model a given combination of chains, but a single model says nothing about which combinations are the right ones.
The authors set out to make the comparison itself the subject of analysis. Instead of judging one predicted model, MultimerMapper takes many predictions in which the same proteins appear in different groupings, splits multi-chain models into their component chain pairs, and reads the prediction software's own confidence scores to decide which pairs genuinely touch. It groups contact patterns into binding modes, sorts single chains by how far their shapes differ, and records which partners were present in each case. An exploration algorithm then proposes which combinations to predict next, stopping when nothing suggested remains untried.
Where AI came in
The AI here is the source of all the structural data. Every structure the paper analyses came from running two released protein structure prediction models without further training: AlphaFold2-multimer version 3 on rented cloud graphics processors, and the AlphaFold3 web server. Five models were produced per protein combination. What the AI stood in for is laboratory structure determination: the shapes and contacts examined were computed from sequence rather than measured from purified protein.
MultimerMapper itself is presented as a non-AI tool. Its interaction calls, clustering, graph analysis and statistics are conventional calculations applied to the AlphaFold output, so the AI supplies the data and the pipeline interprets it. The loop closes when the exploration step hands its suggested combinations back to AlphaFold for another round. On 20 complexes drawn from a public database of known complexes, the expected stoichiometry was among those called stable in 80% of cases and singled out unambiguously in 60%.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
MultimerMapper is a software pipeline that analyses ensembles of AlphaFold predictions in which the same proteins are modelled in different combinations. AlphaFold2-multimer v3 and the AlphaFold3 server produced all the structural models; MultimerMapper then uses statistics, graph theory and clustering, rather than learned models, to decompose multi-chain models into chain pairs, call interactions from confidence metrics, group contact maps into binding modes, and sort isolated chains into RMSD pseudo-trajectories annotated with which partners were present. An iterative stoichiometric space exploration algorithm proposes which new protein combinations the user should predict next, stopping when no suggested combination remains unpredicted. On 20 hetero-oligomeric complexes sampled from the Complex Portal the expected stoichiometry was recovered among the stable stoichiometries in 80% of cases and unambiguously differentiated in 60%, while expected conformational changes were recovered in 6 of 7 test systems and expected PPI dynamics in 8 of 12.
How AI was used
All structural data analysed in this work were produced by running released, pre-trained protein structure prediction models: AlphaFold2-multimer v3 on local GPU instances, with input MSAs from a modified Discoba database concatenated with ColabFold MSA server alignments and no templates, and the AlphaFold3 server. Five models were generated per protein combination using a maximum of 20 recycles. No model was trained or fine-tuned in this study. Everything downstream is conventional, non-learned computation applied to those predictions: decomposition of N-mer models into all pairwise sub-models, extraction of pLDDT, pTM, ipTM, PAE, pDockQ and miPAE, ROC-based selection of miPAE and model-count cutoffs against a manually curated interaction benchmark, residue-residue contact calls from inter-centroid distance and PAE thresholds, Leiden community clustering of the intra-chain PAE matrix for domain segmentation, hierarchical clustering of contact matrices with multidimensional scaling, radius-based symmetry-fallback detection, and phi and point-biserial association tests relating partner presence to interaction and conformational state. The stoichiometric space exploration algorithm closes the loop by handing its suggested protein combinations back to AlphaFold for the next round of prediction.
The shape of the work
Structural · the record, drawn
Curate PPI benchmark pairs
Obtaining raw data, whether by measurement, download or retrieval.
a dataset was generated through manual curation of the literaturewhere the paper describes this · verbatim
Predict 2-mer and N-mer structures with AlphaFold
Running a trained model over new data to predict, classify or score. The AI stood in for physical experiment.
Structure predictions were performed with AlphaFold3 server and a local instance of AlphaFold2-multimer v3.where the paper describes this · verbatim
Decompose models into chain pairs and extract metrics
Cleaning, filtering, normalising or labelling data already obtained.
For N-mers, models are decomposed into its component pairs.where the paper describes this · verbatim
Select interaction-detection cutoffs by ROC analysis
Iterative search over a space.
we performed a receiver operating characteristic (ROC) analysis using the interaction metrics (pDockQ, ipTM and miPAE)where the paper describes this · verbatim
Call interactions and residue-residue contacts
Reducing a candidate set by filtering or ranking, in a single pass.
the PPI between P and Q was considered positive in a prediction if at least 4 models had miPAE≤miPAEcutoffwhere the paper describes this · verbatim
Analyse conformations and binding modes across contexts
Extracting understanding from model behaviour.
MultimerMapper constructs pseudo-trajectories by sorting the isolated conformations in ascending RMSD order relative to a reference.where the paper describes this · verbatim
Iterative stoichiometric space exploration
Iterative search over a space. Its result feeds back into an earlier step.
The process stopped when the system reached convergence, that is, when no suggested stoichiometries remained unpredicted.where the paper describes this · verbatim
Test predictive layers on non-trypanosomatid systems
Testing outputs against ground truth.
we tested all three predictive layers of MultimerMapper against systems outside trypanosomatidswhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
Every structure analysed was produced by AlphaFold2-multimer or AlphaFold3; the paper's interaction, conformational-change and stoichiometry calls exist only because those models generated the ensembles. MultimerMapper itself is presented by the authors as a non-AI tool, so the AI is the data source rather than the analyser.
using 20 hetero-oligomeric complexes randomly sampled from the Complex Portal database across five species, the expected stoichiometry was recoveredwhere the paper describes this · verbatim
The source code of MultimerMapper can be found at https://github.com/elviorodriguez/MultimerMapperwhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Trained model weightsWhether the trained model is available is not stated.
- Version of AlphaFold3 (server)Which version of the model was used is not stated.
About this article
Record aix-00277, version 2, checked by a person on 2026-11-04. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error
Cite this record: AIxSci (2026). Software reads AlphaFold predictions to infer how protein complexes assemble. Record aix-00277, version 2, checked 2026-11-04. https://aixsci.org/articles/aix-00277/