MULTICOM4

AlphaFold solved single-chain structure prediction to a degree that shifted the field's open problem to multimers: complexes whose subunit counts are unknown, whose interfaces depend on paired evolutionary signal across chains, and for which the top-ranked model out of a large sampling run is frequently not the best one produced. MULTICOM4 is the latest version of the MULTICOM system, built by Jianlin Cheng's Bioinformatics and Machine Learning lab at the University of Missouri to attack exactly those three gaps.

It is not a repackaged AlphaFold configuration. MULTICOM4 runs AlphaFold 2, AlphaFold-Multimer, and AlphaFold 3 as complementary structure generators, then wraps them in components the group trained itself: PreStoi for predicting complex stoichiometry, sequence- and structure-based MSA engineering that goes well beyond the stock AlphaFold search, modeling exception handling for targets where the standard pipeline fails, and a suite of deep-learning quality-assessment models used to rank the resulting ensemble. The repository's own download_database_and_tools.py fetcher installs the sequence databases and third-party structure tools the pipeline depends on, separately from the AlphaFold parameter downloads. The learned quality-assessment and stoichiometry components described in the paper are not part of that fetch: the ranking shipped in the repository is pLDDT-based with pairwise structural consensus, and GATE is maintained as a separate, unlicensed repository.

The system was evaluated blind across all CASP16 targets in 2024, with a single configuration applied uniformly rather than tuned per target, and the results were published in Proteins: Structure, Function, and Bioinformatics in 2025.

#Key Features

• Two complementary folding engines: The transformer-based AlphaFold2 and AlphaFold-Multimer and the diffusion-based AlphaFold 3 fail on different targets, so models from both are pooled before ranking.
• Stoichiometry prediction: PreStoi infers subunit counts from structure prediction and template evidence, which is what makes prediction possible when the assembly state is not given.
• MSA engineering: Alignments are diversified using both sequence search across large metagenomic and complex-specific databases and structure comparison, targeting the shallow-MSA cases where a single default alignment collapses.
• Deep-learning model ranking: GATE, a graph transformer for complex quality assessment, is combined with VoroMQA, GCPNet-EMA, EnQA, and pairwise structural consensus to pick the model to submit.
• Exception handling: Explicit fallbacks recover targets — very large assemblies, unusual inputs — on which a plain AlphaFold run does not return a usable model.

#Technical Details

The pipeline runs monomer alignment generation, template search, per-subunit structure generation, alignment and template concatenation, multimer generation, and multi-stage evaluation, with rankings emitted from pLDDT, pairwise MMalign consensus, and their average. CASP16 sampling used 100 predictions per model per AlphaFold predictor (500 per predictor across five checkpoints). Beyond the standard AlphaFold databases, it installs roughly 3 TB more: AlphaFold DB, ESM Atlas, Metaclust, STRING, Uniclust30, pdb_complex and pdb_sort90, DHR, and JGIclust. It also calls ESMFold, AFsample, and Dense Homolog Retrieval as auxiliary tools.

In CASP16 Phase 0, where stoichiometry was withheld, MULTICOM predictors performed best, with MULTICOM_human averaging a TM-score of 0.752 and a DockQ score of 0.584 on top-ranked predictions across 26 targets; standard AlphaFold 3 placed 34th in that phase. In Phase 1, with stoichiometry supplied, MULTICOM_human averaged a TM-score of 0.797 and a DockQ score of 0.558 over 39 targets with native structures. Five MULTICOM predictors competed, three as automated servers and two on the human deadline.

#Applications

MULTICOM4 suits structural biologists modeling assemblies where the subunit stoichiometry is genuinely unknown, where MSAs are shallow, or where a single AlphaFold run gives inconsistent answers and a principled way to choose among many samples matters more than raw throughput. The quality-assessment stage is also useful on its own for scoring externally generated complex models. The trade-off is cost: the databases run into terabytes and the sampling regime is orders of magnitude heavier than a default AlphaFold job, so it fits considered modeling of specific complexes rather than proteome-scale screening.

#Impact

MULTICOM4 documents where the remaining headroom in complex prediction sits — in stoichiometry, alignment construction, and model selection rather than in the folding network — and gives a concrete, reproducible measure of how much each contributes, since its AlphaFold 3 baseline was run under the same conditions. The full system is public on GitHub under an MIT license, including the installer for the MULTICOM-specific databases and trained sub-models, though those sub-models ship bundled through that installer rather than as individually documented, separately downloadable artifacts.