A reusable graph model that turns each node's structural role into a common language—then carries that knowledge from citation and commerce networks into unseen biological systems.
Every graph may use different features and labels. Topology supplies the stable interface.
Convert topology—degree, centrality, ego networks, communities and diffusion—into a natural-language node profile.
The backbone learns across citation, co-authorship, hyperlink, and product co-purchase networks. It is then adapted without labels to protein interaction graphs it has never seen.
Fine-tuned GFM substantially exceeds the strongest supervised message-passing baseline. The frozen zero-shot backbone also surpasses every supervised baseline.
Species-stratified testing asks the model to predict proteins from organisms absent during training. Performance increases as GO terms become more specific.
The zero-shot model exceeds the strongest supervised baseline across molecular function, biological process, and cellular component prediction.
Frozen structural embeddings outperform supervised models on test fold classes that share no identity with the classes used for training.
GO function prediction
GFM leads across MF, BP and CCZero-shot lightweight evaluation and fine-tuning exceed all four supervised GNNs on accuracy, AUROC and macro F1.
Regulatory-network function
GFM leads every ontologyTransfer holds on transcription-factor-to-target edges, rather than physical protein interactions.
GO function prediction
+6 to +14 accuracy pointsThe GFM variants outperform GIN across all three ontologies on a substantially smaller vertebrate proteome.
GO function prediction
+7 to +11 AUROC pointsThe advantage over GIN persists across MF, BP and CC in an invertebrate interactome.
GO function prediction
Top or tied across panelsTransfer extends to a second invertebrate clade across accuracy, AUROC and macro F1.
GO function prediction
Matches or exceeds every baselineThe pretrained representations remain strongest despite a smaller filtered GO label vocabulary.
GO function prediction
Highest trained AUROCFine-tuned GFM achieves the highest mean AUROC on molecular function, biological process and cellular component.
Independent human interactome
86.3 · 90.4 · 82.0 AUROCFine-tuned GFM exceeds the strongest baseline by more than 21 points on MF, BP and CC.
GO function prediction
Top AUROC in all ontologiesFrozen GFM is already at or near the task ceiling and exceeds the trained model on BP and CC AUROC.
Music preference
Matches or exceeds best baselineTopology-based transfer remains competitive outside protein networks on a single-label social task.
TF versus target
Zero-shot leads accuracyA frozen GFM embedding with a lightweight head captures regulatory role and direction.
Music preference
Matches or exceeds best baselineThe shared structural representation transfers to a second music-oriented social network.
Page category
Within 2 points of the bestGFM remains competitive on graph families related to its non-biological pretraining distribution.
Developer category
Within 2 points of the bestNode representations preserve performance on web-versus-machine-learning developer prediction.
Paper topic
Within 2 points of the bestThe reusable backbone remains competitive with specialized supervised models on a citation graph.
Molecular function
The strongest story is not only benchmark performance. Topology-based pretraining recovered functional and spatial organization that was never supplied to the model.
In SagePPI, protein label count positively tracks local GFM embedding density. The strongest supervised baseline shows the opposite trend.
StringGO embeddings separate mitochondrial, nuclear/cytoplasmic, and membrane compartments without localization annotations.
Performance rises with hierarchy depth, consistent with structural prompts preserving long-range and community-level signals.
On Fold-PPI's disjoint test classes, frozen topology-based embeddings avoid the class-specific bias introduced by supervised fine-tuning.
The Graph Foundation Model converts feature-agnostic topology into language-encoded structural prompts, aligns them with message-passing embeddings across heterogeneous graphs, and reuses the pretrained backbone on unseen networks. The submitted study evaluates 18 held-out biological graphs and 6 non-biological graphs across 37 downstream tasks.
The models and interactive workspace are intended for research. Uploaded data should be de-identified and users remain responsible for validating outputs in their own experimental context.