CIFM

A drug does not act on one cell. Knock a target down in a tumour section and what changes first is what the tumour cells secrete; then how the macrophages beside them respond; then whether T cells arrive at all. A model that scores a perturbation cell by cell stops at the first of those steps, the one that matters least.

CIFM does not have to stop there. Developed at Caltech by Matt Thomson's group — styled CI-FM and expanded as the Cellular Interaction Foundation Model in the preprint that released it, and as the Cell Interaction Foundation Model in the follow-up — it is a 100-million-parameter geometric graph neural network trained on one objective: mask a cell and predict its transcriptome from the cells around it. Because its output has the shape of its input, the predictor runs on itself, every cell resampled from its current neighbourhood and the result becoming the next step's tissue. A perturbation held fixed at the start spreads outward through cell-cell interactions instead of staying where it was applied.

That separates CIFM from single-cell foundation models such as scGPT, from spatially-aware models like Nicheformer that treat position as an auxiliary signal, and from perturbation-response models such as STATE, which predict outcomes for dissociated cells. Here the tissue is the unit of prediction.

#Key Features

• Microenvironment-conditioned expression inference: Given a neighbourhood, the model predicts the full expression profile of a cell at a specified location, exposed as predict_cells_at_locations(adata, target_locs).
• Autoregressive tissue play-out: Monte-Carlo iteration of the masked-cell update produces trajectories, typically run for thirty steps until cell-state distributions stabilise. Steps are ordinal and are not calibrated to physical time.
• Combinatorial perturbation screening: A virtual perturbation pins one or more genes to a fixed expression value across the tissue, and every run is scored against a matched unperturbed simulation of the same section.
• Transferable embeddings: A single embed(adata) call turns a spatial sample into microenvironment embeddings that downstream classifiers consume with no fine-tuning of the backbone.
• Cross-panel gene matching: Gene channels are keyed to Ensembl identifiers and matched at load time, letting the same checkpoint be applied to datasets whose gene panels differ from the training data.

#Technical Details

A void-invariant E(n)-equivariant graph network passes messages over a spatial radius graph from MLP-encoded expression vectors, and a second equivariant network acts as the masked-cell decoder, with separate heads for per-gene expression level and detection. Training used 22,951,580 cells across 56 sections and 18,289 genes, a three-platform subset of a 49,953,298-cell curated corpus spanning 281 sections. Averaged over eight tissues, reconstruction error was 0.689 against 0.773 for SpatialPCA, with a noise floor of 0.517. Cell types called from predicted profiles matched measured calls for 64% of cells, 89% of the accuracy the data allows. Frozen embeddings assign held-out whole samples to the correct one of eleven tissue and disease classes at 0.597 accuracy against 0.091 for chance. Configurations from 100M to 4B parameters raised validation correlation from 0.282 to 0.328; the 100M configuration is used for every reported simulation and is the only size released.

#Applications

The design workflow runs a screen, then a trajectory. Across tumour samples the authors scored over 200,000 single and combinatorial perturbations of 487 mostly secreted signaling genes (the abstract puts the count above a million), ranked them on paired objectives — raising CD8A without raising FOXP3 — and mapped the results into eight destination tissue states. Identical interventions change sign between patients, separating tumours into resistant, responsive and intermediate classes from architecture alone. Fine-tuned onto a 980-gene panel, the checkpoint transfers to colonic biopsies from ulcerative colitis and Crohn's disease, where it ranks IL-23 p19 suppression above TNF-alpha blockade. Groups without the data to train their own can use it as a fixed featurizer.

#Impact

A CIFM-designed intervention has been tested prospectively: its top-ranked combination for luminal B breast cancer, CXCL10 induction with INHBC suppression, increased T cell migration toward tumour cells in a transwell assay when delivered as CXCL10 plus a blocking antibody against INHBC, as did CXCL10 paired instead with an antibody against the receptor ACVR1C, where CXCL10 alone and anti-PD-1 alone did not. Those replicates are wells within a single experiment. Retrospective comparisons cover simulated NSD2 inhibition against published prostate cancer organoid and xenograft data — from a Nature paper since retracted over animal-welfare protocol breaches, whose retraction note states no other data were affected — and virtual T cell-tumour dyads against cell-cell sequencing, at Pearson r = 0.47. Suppressing a ligand transcript is not antibody neutralisation of the secreted protein, and the authors read their rankings as relative ordering rather than predicted efficacy. Weights and inference code are on Hugging Face under MIT with a tutorial notebook; the simulation and screening code behind the newer results is available from the authors on request only, and neither preprint has been peer reviewed.