Loading…
Protein-protein interaction embedding map that infers what a single interaction does from its neighbours among 199,137 human PPIs.
No providers recorded yet. Browse all providers
A protein's role in the cell depends on which partner it is engaged with, yet most functional interpretation of protein-protein interactions (PPIs) remains protein- or set-centric: the interactors of a bait protein, or the members of a network cluster, are tested for over-represented Gene Ontology, pathway, or domain terms. Because those strategies inherit pre-existing annotation, they are weakest exactly where knowledge is sparsest.
MAPPIE (Map of Protein-Protein Interaction Embeddings), from the Computational Biology and Data Mining group at Johannes Gutenberg University Mainz with the University Medical Center Mainz, makes the interaction itself the unit of representation. Protein language model embeddings for the two partners are merged into a single interaction vector, compressed by a trained autoencoder, and arranged into a fixed two-dimensional map built from 199,137 human interactions across 15,503 proteins. The same protein appears in different regions when paired with different interactors, making partner-specific context visible as geometry.
That framing separates MAPPIE from most sequence-based PPI work, which uses protein language models to decide whether two proteins interact. MAPPIE asks instead what an interaction means, and does so without end-to-end fine-tuning: the backbone stays frozen, so any candidate pair — including engineered, mutated, or fragment sequences — projects into the existing map without retraining.
Both partners are encoded with a frozen ESM-2 650M model and mean-pooled over the final layer to 1280 dimensions, then merged elementwise. The merged vector is standardized and compressed by a four-layer PyTorch MLP autoencoder with widths 512 and 256 and a 128-dimensional bottleneck, trained for 20 epochs (Adam, learning rate 1e-3, batch size 1024) on a 90/10 split under mean squared error. A grid search over embedding source (ESM-2 or ProtBERT), four merge operators, three bottleneck sizes, and two HIPPIE confidence thresholds selected elementwise multiplication at 128 dimensions, scoring 0.61 on the domain-domain compactness criterion. UMAP is fit once and used only for visualisation — retrieval and enrichment run in the 128-d latent space under cosine distance.
Benchmarks are annotation-recovery rather than binary classification. A query's own GO biological process terms are recovered at 72.8% from its ten nearest neighbours, with GO CC and KEGG reaching 73.8% and 74.0% at k = 500; the compressed space beats the uncompressed merged embedding in every category. For sparsely connected interactions (1–10 known partners, close homologues excluded), MAPPIE recovers 38.8% of KEGG and 26.3% of GO CC terms against 26.8% and 20.4% for a first-degree interactome baseline, though that advantage reverses once a query has more than ten known partners. Latent neighbours are enriched 19.7-fold for high-confidence STRING edges at k = 10, and across 302 CORUM complexes the pooled neighbourhoods recover curated GO terms at 72.0% mean recall.
MAPPIE targets interactions that conventional annotation transfer cannot reach. Applied to a human dark interactome of 1,925 high-confidence BioPlex pairs in which both partners lack usable annotation, it assigned experimentally corroborated roles to dark hub proteins: positive regulation of apoptosis for PNMA2 across 21 of its 26 interactions, p53 class mediator apoptotic signalling for MAGB2, and interferon-related regulation for RNF149. Because arbitrary pairs project into the fixed map, the same workflow interprets newly predicted interactions, engineered constructs, and individual edges from a proteomics screen.
MAPPIE reframes functional annotation of the interactome from a protein-level to an edge-level problem, and shows that a frozen protein language model plus a lightweight compression head makes that reframing tractable at interactome scale. The design carries boundaries the authors state directly: the reference map is human-only and built on an interactome estimated to cover well under a quarter of pairwise human interactions; the map places any submitted pair regardless of whether the two proteins genuinely bind, making it a tool for interpretation rather than interaction validation; and because most annotations are protein-level, the benchmarks use the union of both partners' terms. The MIT-licensed code follows the manuscript's Methods but ships training and projection scripts rather than packaged checkpoints, so local reuse means rerunning the autoencoder training and the UMAP fit. The preprint has not yet been peer reviewed.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.