CellTOSG-FM

CellTOSG-FM is a multimodal graph foundation model for single-cell biology from the Fuhai Li lab at Washington University in St. Louis, released as a preprint in April 2025 alongside the dataset it is trained on. It addresses a structural limitation of transcriptomic foundation models such as scGPT and scFoundation: those models serialize a cell into an ordered sequence of genes and learn from expression values alone, with no explicit representation of what each gene does or how its protein product interacts with others. The signaling topology and the accumulated biomedical literature that a human analyst would bring to bear are simply absent from the input.

The proposed remedy is a data structure the authors call the Text-Omic Signaling Graph (TOSG), in which each node carries three aligned modalities — a curated textual description, a quantitative omic measurement, and a biological sequence — and edges encode transcript-to-protein relationships and protein-protein interactions. The accompanying corpus, OmniCellTOSG, was assembled from 79,195,364 single-cell and single-nucleus RNA-seq profiles drawn from CellxGene, the Brain Cell Atlas, GEO, the Single Cell Portal, and the Human Cell Atlas, then aggregated into 395,317 meta-cell graphs using SEACells and annotated against prior knowledge from BioMedGraphica.

CellTOSG-FM is the model pretrained on that resource. The same lab later built CellTosg2Sequence on this line of work, taking a different route by encoding a biomedical knowledge graph into virtual tokens for a general-purpose language model backbone.

#Key Features

• Three aligned modalities per entity: Textual entity names and descriptions, biological sequences, and numerical expression values are embedded separately and fused by a cross-modality encoder into a single entity representation.
• Frozen biological sequence encoders: RNA sequences are encoded with DNAGPT (with thymine substituted by uracil) and protein sequences with ProtGPT2; both are held frozen because BioMedGraphica entities are stable and reused across samples, which removes redundant computation and stabilizes reproducibility.
• Two-stage message passing: Signals propagate first within transcript-protein pairs at the nucleus level, then diffuse across the cell through the protein-protein interaction topology.
• Self-supervised graph pretraining: The model is trained by reconstructing stochastically masked PPI edges alongside a node-degree regression term that calibrates hub-versus-periphery structure, requiring no metadata labels.
• Interpretable signaling rationales: Because predictions are made over an explicit signaling graph, the model can rank disease-associated targets and pathways rather than returning an opaque embedding.
• CellTOSG_Loader: A companion package turns user-specified conditions (tissue, disease, cell type, sex) into stratified, class-balanced cohorts in a model-ready format.

#Technical Details

The full OmniCellTOSG graph comprises 533,458 entities and 16,637,405 relations, of which 152,585 are internal transcript-protein interactions and 16,484,820 are protein-protein interactions. Cell type labels are standardized to the Cell Ontology, covering 766 cell types across 65 tissues, and disease annotations are mapped to 140 states in the BioMedGraphica nomenclature. Pretraining uses a bi-encoder framework in which a trainable linear omic encoder and the frozen text and sequence encoders feed a cross-modality encoder, followed by graph encoders over the TOSG topology. At an edge-masking ratio of 10^-5, the model recovers roughly 80% of masked edges and converges to an AUC near 0.85 — achieved while pretraining on only 5% of OmniCellTOSG.

For downstream use the pretrained architecture and weights serve as a transferable initialization, with a task-specific decoder fine-tuned on top. Evaluation covers cell type annotation, disease-versus-normal classification, and sex classification across four cohorts: Alzheimer's disease, lung adenocarcinoma, atrial fibrillation, and systemic lupus erythematosus. Roughly 1,000 meta-cells were subsampled per task, with training and test sets partitioned by donor identity to avoid leakage. Across all four cohorts CellTOSG-FM matched or exceeded feedforward, GCN, GAT, and UniMP baselines as well as scGPT, scFoundation, and scCello.

#Applications

The model targets disease-focused single-cell analysis where interpretability matters as much as accuracy: annotating cell types in patient cohorts, separating diseased from control cells within a tissue, and ranking candidate targets and signaling pathways implicated in a condition. Researchers working with public atlas data benefit most directly, since CellTOSG_Loader handles retrieval and cohort balancing across the assembled corpus, mitigating the confounding by sex, age, and cell-type composition that complicates cross-study single-cell comparisons.

#Impact

OmniCellTOSG and CellTOSG-FM together argue that structured biological priors — literature text, gene function, and signaling topology — are a usable input modality for cell-level models rather than context that has to be reintroduced during interpretation. The dataset is the more immediately reusable artifact: it is distributed on Hugging Face with a substantive data card, while no pretrained CellTOSG-FM checkpoint is hosted, so reproducing the model requires running the released pretraining pipeline. The evaluation is also modest in scale, with roughly 1,000 meta-cells per downstream task, and the work remains a preprint whose comparisons against established single-cell foundation models have not been independently reproduced.