Single-cell model inferring which developmental signaling pathways are active from scRNA-seq, trained on combinatorial stem-cell perturbation screens.
No providers recorded yet. Browse all providers
Cells decide what to become by integrating combinations of external signals, so knowing which pathways were active in a cell — and in what order — is central to understanding development and to engineering it. Genetic perturbations and transgenic reporters are low-throughput and largely unavailable in human tissue, while the computational alternatives infer communication from ligand-receptor co-expression, which reports signaling potential rather than whether a cell responded. The transcriptional response itself is the obvious readout, but the field has long assumed those responses are specific to each cell type, which would mean perturbing every cell type of interest.
IRIS, from Pulin Li's lab at the Whitehead Institute and MIT with a collaborator at the Medical University of Vienna, tests that assumption and finds it too strong. The group built an in vitro perturbation atlas by running sequential combinatorial signal screens on human pluripotent stem cells across six developmental pathways — TGF-beta, FGF, BMP, Hedgehog, retinoic acid, and Wnt — then trained a neural network to classify pathway activation from a cell's full transcriptome. The resulting signatures behave like a per-pathway fingerprint that transfers between unrelated cell types, batches, and species.
Applied without retraining to a mouse gastrulation atlas, IRIS mapped signaling states across roughly 40 annotated cell types from E6.5 to E8.5, reconstructed the sequence of signal combinations along endodermal and cardiac lineages, and recovered morphogen gradients across the somitic mesoderm. Its predictions are cell-autonomous, making it complementary to ligand-receptor methods such as CellPhoneDB, NicheNet, and CellChat rather than a replacement.
IRIS is a semi-supervised model built on the scANVI architecture from scvi-tools, a conditional variational autoencoder of the same family as scVI. An encoder maps raw counts and one-hot covariates — batch, cell type, species — into a latent space under a zero-inflated negative binomial likelihood, while a feedforward classifier on that latent space is trained jointly against binary pathway labels, with a sigmoid output thresholded at 0.5. Hyperparameter search spanned 1 to 3 hidden layers, 32 to 1,024 hidden units, and 10 to 60 latent dimensions; shallow architectures performed best, and models trained for 200 epochs on a single NVIDIA A6000.
Training data combines a published mouse embryonic stem cell screen with three new human ESC screens: mesodermal cells given 46 signal combinations before a second round yielding 96 final states, four splanchnic mesoderm subtypes each given all 16 combinations of TGF-beta, Wnt, FGF, and BMP, and definitive endoderm given all 64 combinations of the six pathways. Against elastic net, random forest, and two tuned SVMs, IRIS was comparable in-sample but held its advantage in the data-limited cross-species setting. A gene ablation test showed that thousands of genes must be randomized before accuracy falls to chance, indicating a transcriptome-wide signature rather than one carried by a handful of canonical response genes.
The immediate use is annotating signaling state in existing scRNA-seq atlases, where IRIS recovered known biology — BMP along the cardiac but not endodermal lineage, Wnt inactivation preceding cardiomyocyte differentiation, retinoic acid marking atrial cardiomyocytes — and surfaced predictions in poorly characterized lineages. It predicted that Wnt and BMP are enriched in respiratory mesenchyme, confirmed in ex vivo mouse foregut culture: a Wnt agonist expanded Tbx4 and Foxf1 expression while an inhibitor eliminated Tbx4. Activating Wnt earlier and longer than the published protocol then significantly improved the efficiency of deriving respiratory mesenchyme from hESCs. That loop is the payoff for stem cell and organoid groups facing a combinatorial space too large to screen.
The central claim is a biological one with methodological consequences: signal response signatures are shared across cell types far enough apart to make transfer learning viable, so a screen run in accessible cells can annotate atlases of cells that cannot be perturbed. How far that transfer extends is not settled — performance improves markedly when a small labeled sample from the target species is included, and FGF transfers less reliably than Wnt or retinoic acid, plausibly because it shares transduction machinery with other receptor tyrosine kinase pathways. Signal carry-over between sequential stimulation steps raises false-positive rates for BMP, Wnt, and TGF-beta, though these stay below 0.3 in most conditions, and reconstructing histories inherits whatever error the underlying lineage inference carries. The code is MIT-licensed and the perturbation atlas became publicly downloadable from GEO as GSE289836 in April 2026, though no trained checkpoint has been released, so reproducing the transfer results means retraining from the screens. The work remains a preprint, posted under bioRxiv's no-reuse terms.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.