CellNavi

Most single-cell perturbation models run forward: given a gene to knock out, predict the transcriptome that results. CellNavi runs the other direction. Given a source cell and a target cell — an unperturbed T cell and its restimulated counterpart, a healthy neuron and a stressed one — it ranks candidate genes by how likely each is to be the driver of that transition. This inverse framing maps directly onto questions experimentalists actually ask: which gene should I target to push these cells where I want them to go?

CellNavi was developed at Microsoft Research AI for Science with collaborators at the University of Washington, Tsinghua University, and Beijing Normal University, released as a bioRxiv preprint in October 2024 and published in Nature Cell Biology in October 2025. It is a two-stage system. A cell manifold model (CMM) is pretrained self-supervised on roughly 20 million human single-cell transcriptomes to embed each cell as a coordinate on a low-dimensional manifold where distance reflects biological similarity. A driver gene predictor (DGP) then reads a pair of those coordinates and scores candidate genes, fine-tuned on single-cell CRISPR screens where the perturbed gene is known.

What separates the CMM from expression-token pretraining in scGPT or Geneformer is that a prior gene graph is built into the attention mechanism rather than applied afterward. Where GEARS uses a gene-gene graph to predict expression responses, CellNavi uses a directional regulatory graph to shape the representation the driver-gene classifier consumes.

#Key Features

• Inverse perturbation inference: Ranks which gene most plausibly caused an observed transition between two cell states, rather than predicting the effect of a chosen perturbation.
• GeneGraph attention: Standard transformer encoder layers are replaced with layers that inject the NicheNet directional gene graph through centrality encoding (in/out-degree) and spatial encoding (graph-distance attention biases).
• Pretrained cell state manifold: Self-supervised pretraining on Human Cell Atlas data yields a reusable cell representation; ablations show removing it hurts out-of-domain splits far more than in-domain ones.
• Generalizes across perturbation modalities: Evaluated on genetic (CRISPR), chemical, and cytokine-induced transitions spanning differentiation, disease progression, and drug response, across different cell types and independent studies.
• Released weights and code: The pretrained checkpoint, training and inference package, and worked tutorials are public under GPL-3.0; NVIDIA GPUs (V100/A100 class) are required.

#Technical Details

The CMM stacks six GeneGraph attention layers over 256-dimensional gene embeddings formed by summing a learnable gene-name embedding with an expression embedding. During training at most 2,048 non-zero genes are sampled per cell for speed and augmentation; at inference all non-zero genes are used. The CLS representation is projected from 256 to 2,048 dimensions to form the cell coordinate, and a decoder reconstructs full expression from a downsampled profile as the pretraining objective, then is discarded. Pretraining used ~20 million single-cell and single-nucleus profiles from the Human Cell Atlas (roughly 1.5 TB of raw data through March 2023). The DGP is a two-layer MLP over the source and target coordinate pair.

On the Schmidt CRISPR screen in primary T cells, CellNavi reaches 0.621 top-1 and 0.733 top-5 accuracy across 23,047 source-target cell pairs. On the Norman K562 dataset with dual-gene perturbations held out, it ranks the first and second perturbed genes at average positions 7.9 and 31.2 of 105 candidates. Baselines including SCENIC and SCENIC+, GEARS in-silico perturbation, GRN inference by GENIE3, GRNBoost2 and RENGE, differential expression analysis, and linear models were all outperformed; SCENIC could not score 61.6% of samples for lack of matching regulons. Substituting inferred GRNs, sparsified graphs, or random graphs for NicheNet degraded accuracy.

#Applications

The framework targets four workflows named by its authors: nominating CRISPR screen targets, predicting master regulators of differentiation, discovering disease drivers, and elucidating drug mechanism of action. In iPS-derived dopaminergic neurons under tunicamycin-induced ER stress — a perturbation that touches no single gene directly — CellNavi placed the ER stress genes EIF2S1, BAX and HSPA5 at median ranks 3, 7 and 16 among 184 candidates. Applied to K562 cells treated with 17 HDAC inhibitors, its top-scoring genes separated compounds into clusters whose enriched pathways tracked reported potency differences.

#Impact

CellNavi reframes a question that Perturb-seq analysis usually answers with differential expression heuristics into a supervised ranking task grounded in a pretrained representation, and its ablations make a concrete case that biologically curated gene graphs beat inferred or random ones for this purpose. The practical caveats are real: the driver gene predictor needs labeled CRISPR screen data to fine-tune, candidate genes must be enumerated in advance, results reported are computational rather than prospectively validated at the bench, and the released pretraining corpus is human-only. The pretrained manifold is nonetheless directly reusable, and the GPL-3.0 release with tutorials lowers the barrier for groups wanting to apply it to their own paired cell-state data.