CIFM

Tissues are not bags of independent cells. What a cell expresses depends on which cells it touches, and spatial genomics now measures that arrangement at transcriptome scale across millions of cells. Turning those measurements into a model of cellular interaction is harder than it sounds: the problem is multi-scale, with gene circuits inside each cell driving interactions among many diverse cell types, and the data volumes rule out methods that do not scale.

CIFM — the Cellular Interaction Foundation Model, written CI-FM in the paper — is the Thomson Lab's answer, developed at Caltech by Yuning You, Zitong Jerry Wang, Kevin Fleisher, Rex Liu and Matt Thomson. It is a 100-million-parameter geometric graph neural network trained on spatial genomics data covering roughly 23 million cells. The pretraining objective is deliberately simple and entirely self-supervised: mask a cell and predict its gene expression from the cells in its spatial microenvironment. A model that can do this well has, by construction, learned how neighbourhood composition determines cell state.

That framing puts CIFM in a different place from single-cell foundation models such as scGPT or spatially-aware models like Nicheformer, which learn from expression profiles with spatial context as an auxiliary signal. CIFM makes the microenvironment the input and the cell the target, which is what allows it to be run generatively — placing hypothetical cells at chosen coordinates and reading out the tissue's predicted response.

#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).
• Transferable embeddings: A single embed(adata) call turns a spatial sample into microenvironment embeddings that downstream classifiers consume directly, with no fine-tuning of the backbone.
• In-silico perturbation: Because generation is conditioned on composition, the authors simulate T cell infiltration and read out how the surrounding cells shift — a step toward what they call an AI virtual tissue.
• Geometry-aware architecture: Interactions are modelled on a radius graph with equivariant message passing, so a prediction depends on the relative arrangement of neighbours rather than on absolute coordinates or orientation.
• 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

An MLP gene encoder lifts each cell's expression vector into the hidden space; a void-invariant E(n)-equivariant graph network then passes messages over a spatial radius graph, and a second equivariant network acts as the masked-cell decoder. Two MLP heads emit per-gene expression and a dropout probability, the standard treatment for the zero inflation of spatial transcriptomic counts. On held-out data the model reaches a mean squared error of 1.1% relative to the square of median expression, and 79.4% of cells on average are assigned the same cell type from predicted expression as from measured expression. Applied to tumor samples it was not trained on, frozen CIFM embeddings support linear-probe classification of sample condition at ROC-AUC 0.76 while surfacing signatures shared across samples.

#Applications

The clearest use is tumor microenvironment analysis: embed a cohort of spatial samples, probe the embeddings for a clinical or treatment variable, and inspect the interaction signatures that carry the signal. The generative side supports hypothesis generation for immunotherapy — asking what a niche is predicted to do when T cells arrive, before running the experiment. More broadly, any group with spatial transcriptomics or proteomics data can use the checkpoint as a fixed featurizer, which is attractive for the many studies whose sample counts are too small to train a model of their own.

#Impact

CIFM is a preprint and has not yet been peer reviewed, and the checkpoint carries no download or citation history that would demonstrate broad uptake. What it does contribute is a clean formulation — expression-from-microenvironment as a self-supervised objective on geometric graphs — that makes cellular interaction the modelled quantity rather than a by-product. The weights and inference code are released under the MIT license on Hugging Face with a worked tutorial notebook, while the preprint text itself is CC-BY-NC. Reported validation is entirely computational: the T cell infiltration simulation is compared against expectation rather than against a matched perturbation experiment, and no training dataset card accompanies the release.