henhaobio
Sign in

Molecular Systems BiologySystems BiologyPeer-reviewed journal article

hu.MAP3.0: Atlas of human protein complexes by integration of > 25,000 proteomic experiments

Samantha N FischeriD, Erin R. ClausseniD, Savvas KourtisiD, Sara SdelciiD, Sandra E. OrchardiD, Henning HermjakobiD, Georg KustatscheriD, Kevin DrewiD

PDFDOIPreprintCodeData

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (15 Oct 2024) · CC BY 4.0

Machine learning integration of over 25,000 mass spectrometry experiments yields 15,326 human protein complexes covering 13,769 proteins, plus 511 AlphaFold-derived mutually exclusive subunit pairs.

Many human protein complexes still lack a known subunit composition, and the previous map, hu.MAP2.0, placed only about half of human proteins (about 10,000) into complexes. This update adds about 10k new BioPlex3 AP-MS experiments and an AutoGluon ensemble classifier, then uses protein covariation and AlphaFold models to annotate understudied proteins and flag subunits that cannot share a complex at once.

Key results

  1. ShownOn held-out Complex Portal interactions, the AutoGluon co-complex network is more accurate than earlier maps. Clustering it gives a much larger complex map than hu.MAP2.0 or BioPlex3.

    AUPRC 0.68 for hu.MAP3.0 vs 0.61 for hu.MAP2.0 and 0.38 for BioPlex3 on leave-out test pairs. The final map has 15,326 complexes, 13,769 proteins and 159,451 scored interactions, about 40% more proteins than hu.MAP2.0. · Results; Construction of machine learning model to…; Discussion

  2. ShownProteomeHD.2 protein covariation comes from a separate set of expression experiments, so it gives independent support. Pairs that score high in both resources are strongly enriched among OpenCell interactions, which neither resource used as input.

    About 15 million protein pairs are scored by both resources. OpenCell recovers 29% of pairs that are high in both hu.MAP3.0 and ProHD.2, vs 0.3% of random pairs, and more than pairs that are high in only one resource. Complex covariation correlates only mildly with confidence level. · Protein covariation provides orthogonal evidence…; Covariation provides complementary insights into…

  3. SuggestedTransferring annotations from enriched complexes to their members (guilt by association) gives functional hypotheses for understudied proteins. Selected pairs are backed by covariation and AlphaFold3 models, e.g. REX1BD with the 19S proteasome lid and TMEM167A with the disease gene IER3IP1.

    472 uncharacterized proteins (UniProt annotation score <=3) annotated, vs 274 in hu.MAP2.0. REX1BD-PSMD13: score 0.99, covariation 0.43 (top 0.6%), AF3 pTM 0.75, ipTM 0.67. IER3IP1-TMEM167A: score 0.992, covariation 0.78, ipTM 0.77, with the MEDS mutations Val21Gly and Leu78Pro at the interface. · Uncharacterized protein REX1BD interacts with…; hu.MAP3.0 provides testable hypotheses to…; Annotation enrichment analysis for hu.MAP 3.0…

  4. ShownConfident AlphaFold2 pairwise models are aligned on a shared third subunit, and each co-complex pair is called mutually exclusive or structurally consistent. Mutually exclusive pairs are heavily enriched for paralogs, which points to gene duplication as a route to swappable subunits.

    511 mutually exclusive vs 421 structurally consistent pairs, from more than 10,000 models filtered at pDockQ > 0.23 and IF_pLDDT > 70. Paralogs (eggNOG): 201 of 511 (~39%) mutually exclusive pairs vs 9 of 482 (~2%) structurally consistent pairs. The method recovers the known PSMB5/PSMB8 swap. · Identification of mutually exclusive subunits in…; Identification and characterization of mutually…

  5. ShownMutually exclusive partners differ in expression rather than in subcellular localization. An expression-based ALL-IN score built from cancer cell line proteomes separates them, most clearly when the shared subunit does not form homomultimers.

    Compared with structurally consistent pairs: lower ProteomeHD.2 covariation (Welch's t-test P < 0.0001), lower tissue RNA correlation (Mann-Whitney P < 0.0001), and no significant difference in HPA localization Jaccard similarity. ALL-IN is higher for mutually exclusive pairs (P < 0.0001) across 949 cancer cell lines. · Identification of mutually exclusive subunits in…; Lineage - specific expression of protein…; Figure legends

Methods, data and reproducibility

Each experiment type contributes its own features: CF-MS elution similarity, AP-MS and proximity-labeling bait-prey scores, and hypergeometric weighted-matrix co-occurrence. These features for 25,992,008 protein pairs train an AutoGluon classifier on Complex Portal labels. The scored network is clustered with ClusterOne and then MCL, testing 1,080 parameter combinations.

  • hu.MAP2.0 feature matrix of CF-MS, AP-MS and proximity labeling features, reused as the base of the hu.MAP3.0 feature matrix · human · n = 292 features; 220 from 55 CF-MS datasets
  • BioPlex 3.0 AP-MS experiments in HEK293T and HCT116 cells (newly integrated features and WMM scores) · human · n = over 10,000 baits (about 10k new AP-MS experiments)
  • Complex Portal manually curated human complexes, split into gold-standard train and test sets · human · Complex Portal version 2022_07_11
  • ProteomeHD.2 covariation map from reprocessed SILAC raw files in PRIDE, covering many biological perturbations · human (pairs matched to hu.MAP3.0 proteins) · n = 23,000 raw files; 2,498 perturbations
  • OpenCell experimentally determined pairwise interactions, used as held-out orthogonal validation
  • AlphaFold2 models of co-complex human protein pairs, used for mutual exclusivity analysis · human · n = more than 10,000 structures
  • Cancer cell line proteomes (Gonçalves et al.), used for lineage expression and the ALL-IN score · human · n = 949 cell lines; 28 tissues; over 40 cancer types
  • Human Protein Atlas RNA tissue consensus, tissue specificity and subcellular localization annotations · human
Methods & tools
AutoGluon TabularPredictor, XGBoost (sole member of final ensemble), ClusterOne, MCL (Markov clustering), Weighted matrix model (hypergeometric co-occurrence), k-clique weighted precision/recall evaluation, g:Profiler, AlphaFold3 web server, AlphaFold2 pairwise models filtered by pDockQ and IF_pLDDT, PyMOL align, eggNOG (Eukaryota ortholog groups), treeClust and Random Forest (ProteomeHD.2)
Reproducibility gaps
Pairwise ProteomeHD.2 scores are available only on request, and ProteomeHD.2 itself is deferred to a separate manuscript in preparation. No software versions, containers or compute resources are reported.
Limitations (6)
  • Mutual exclusivity calls assume one binding site on the common subunit per competing partner. Common proteins that homomultimerize or form fibers can host both partners, so this assumption may not hold for them. · stated by the authors
  • Only steric overlap at the interface is detected. Partners that exclude each other allosterically, by reshaping the common subunit's surface, would need modeling of alternative conformations to be found. · stated by the authors
  • Calls depend on chosen cut-offs (pDockQ > 0.23, IF_pLDDT > 70, more than 10 overlapping interface residues). Pairs with some overlap but fewer than 10 residues were left inconclusive because of expected AlphaFold2 error. · stated by the authors
  • Low subunit covariation is hard to interpret. It may reflect missed covariation in ProteomeHD.2, false complexes in hu.MAP3.0, or subunits that also have functions outside the complex. · stated by the authors
  • The new interactions (REX1BD-PSMD13, IER3IP1-TMEM167A, DIPK2B-NDST2, LNP1-YWHAE) rest on reanalyzed pulldowns, covariation and AlphaFold3 confidence. The text reports no new targeted biochemical or cellular validation. · noted in this summary
  • Some counts disagree within the text: 421 structurally consistent pairs in one place vs 482 in the paralog comparison, and six selected cluster sets later described as five merged clusterings. This makes exact reproduction harder. · noted in this summary

Read this if you need human protein complex memberships with confidence levels, function hypotheses for understudied proteins, or a structure-plus-expression workflow for finding mutually exclusive complex subunits.

Citations
24
FWCI
5.79
Field percentile
top 2%

Abstract

Macromolecular protein complexes carry out most functions in the cell including essential functions required for cell survival. Unfortunately, we lack the subunit composition for all human protein complexes. To address this gap we integrated >25,000 mass spectrometry experiments using a machine learning approach to identify > 15,000 human protein complexes. We show our map of protein complexes is highly accurate and more comprehensive than previous maps, placing [~]75% of human proteins into their physical contexts. We globally characterize our complexes using protein co-variation data (ProteomeHD.2) and identify co-varying complexes suggesting common functional associations. Our map also generates testable functional hypotheses for 472 uncharacterized proteins which we support using AlphaFold modeling. Additionally, we use AlphaFold modeling to identify 511 mutually exclusive protein pairs in hu.MAP3.0 complexes suggesting complexes serve different functional roles depending on their subunit composition. We identify expression as the primary way cells and organisms relieve the conflict of mutually exclusive subunits. Finally, we import our complexes to EMBL-EBIs Complex Portal (https://www.ebi.ac.uk/complexportal/home) as well as provide complexes through our hu.MAP3.0 web interface (https://humap3.proteincomplexes.org/). We expect our resource to be highly impactful to the broader research community.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISDrew '21Drew '17Pan '23Uckelmann '25Oghbaie '19Hajkarim '19
81% semantic matchSystems Biology
Kevin Drew, John B. Wallingford, Edward M. Marcotte·Molecular Systems Biology·May 1, 2021· 136 citations