The central obstacle the researchers confronted is a familiar one in modern biology. The richest data on cellular signaling come from perturbation screens performed in controlled in vitro settings, where scientists can apply defined stimuli — a dose of interferon, a jolt of EGF, a targeted kinase inhibitor — and observe, often at the protein level, how signaling networks respond within minutes. These experiments are clean, interpretable and information-dense. But cells in a dish are not cells in a body. In vivo contexts are shaped by tissue architecture, immune interactions, developmental cues and a constantly shifting microenvironment, and perturbation data of comparable quality simply cannot be gathered there. The question was whether the deep mechanistic knowledge embedded in in vitro screens could be transferred to the messy, heterogeneous world of living organisms.

IRIS — the framework’s name reflects its role as an interpreter of signaling — answers that question affirmatively by learning what its developers call conserved signaling representations. The method is trained on large collections of perturbation screening data, spanning many stimuli, many cell types and many experimental conditions. From this corpus it distills a shared mathematical space in which the essential features of signaling activity are encoded: which pathways are active, in what combinations, and with what intensity. Crucially, the representation is designed to be context-invariant. It abstracts away from the particulars of any single experiment and captures the recurring logic of signal transduction itself — the way receptor activation cascades into kinase activity, transcription factor engagement and downstream gene expression. Once this conserved space has been learned, the framework can map cells from an entirely different setting, such as an in vivo tissue atlas, into the same coordinate system and read off their signaling states.

Technically, the approach draws on transfer learning, one of the most consequential ideas in contemporary machine learning. Rather than training a model from scratch on scarce in vivo data, IRIS transfers knowledge acquired in the data-rich in vitro domain to the data-poor in vivo domain. The perturbation screens serve as a kind of ground-truth curriculum: because the stimuli are known, the model can learn the explicit correspondence between an applied perturbation and the resulting signaling response. That supervised grounding is what distinguishes IRIS from purely unsupervised approaches, which can cluster cells by similarity but cannot attribute those similarities to specific signaling events. With the transfer learned, the model performs what is effectively computational inference across domains — taking the gene expression profile of a single cell from a tumor, an embryo or an inflamed organ and reconstructing the signaling activity that most plausibly produced it.

The implications of reconstructing signaling histories extend well beyond technical elegance. Many of the most important decisions a cell makes are driven by transient signals that leave subtle traces in gene expression. An immune cell that encountered its antigen hours ago, a stem cell that received a morphogen pulse during development, a cancer cell that survived a burst of chemotherapy-induced stress — all of these carry records of their past exposures in their present molecular state, but those records are encrypted. By decoding them, IRIS allows researchers to ask retrospective questions that were previously unanswerable: which cells in a tissue recently received a particular signal, which pathways were engaged in sequence, and how the history of stimulation differs between neighboring cells that look superficially identical.

In disease research, this capability is especially potent. Tumor microenvironments are theaters of continuous signaling crosstalk, with malignant cells, immune infiltrates and stromal cells exchanging signals that determine immune evasion, therapy resistance and metastatic potential. Bulk genomic methods average away this complexity, and even single-cell atlases typically report endpoint states. A framework that infers signaling histories cell by cell can reveal, for example, which subpopulations of tumor cells have been receiving survival signals from their surroundings, or which immune cells show the signaling signature of recent activation versus chronic exhaustion. Such information could sharpen the design of combination therapies, identifying not just which pathways are active in a tumor but which were active in the events leading to the observed cellular landscape.

The method also addresses a persistent reproducibility problem in single-cell biology. Signaling measurements are notoriously difficult to standardize across laboratories, platforms and tissue types, which has slowed the accumulation of comparable in vivo signaling data. By anchoring interpretation in a conserved representation learned from perturbation data, IRIS provides a common reference frame. Two studies of different tissues, generated with different protocols, can be projected into the same signaling space and compared directly. This harmonizing function may prove as valuable as the reconstruction itself, because it converts fragmented, experiment-specific observations into a cumulative, comparable body of knowledge about how signaling operates across the body.

Like any ambitious computational method, IRIS rests on assumptions that will require continued scrutiny. Transfer learning works best when the source and target domains genuinely share underlying structure, and while core signaling circuits are indeed conserved, in vivo contexts introduce regulatory layers — mechanical forces, metabolic gradients, three-dimensional tissue organization — that in vitro screens cannot fully represent. The framework’s predictions are inferences, not direct measurements, and experimental validation in specific biological systems remains essential. The authors position the tool as a hypothesis-generating engine: it nominates the signaling events most likely to have shaped a cell, which experimentalists can then test with targeted perturbations, phosphoproteomics or lineage tracing. In this sense, IRIS does not replace laboratory investigation but directs it, concentrating experimental effort on the most informative hypotheses.

The broader significance of the work lies in what it signals about the trajectory of computational biology. The field is moving from descriptive cataloguing — enumerating the cell types present in a tissue — toward mechanistic reconstruction, in which algorithms infer the dynamic processes that generated the observed states. Perturbation screens supply the causal grammar; transfer learning supplies the bridge from controlled experiments to natural contexts. Together they suggest a future in which a routine single-cell dataset can be reinterpreted not as a static census but as a record of conversations between cells and their environments. For a discipline built on snapshots, the ability to reconstruct history from a single frame is a genuinely transformative shift, and IRIS offers a concrete, testable route toward that goal.

Subject of Research: Reconstructing single-cell signaling states and histories using perturbation screens and transfer learning.

Article Title: Reconstructing signaling histories of single cells via perturbation screens and transfer learning

Article References: Hutchins, N. T., Meziane, M., Lu, C., Mitalipova, M., Fischer, D. S., & Li, P. (2026). Reconstructing signaling histories of single cells via perturbation screens and transfer learning. Nature Methods. https://doi.org/10.1038/s41592-026-03213-8

Image Credits: AI Generated

DOI: 10.1038/s41592-026-03213-8

Keywords: single-cell sequencing, signaling pathways, perturbation screens, transfer learning, IRIS, in vivo cellular contexts, computational biology, tumor microenvironment, signaling history reconstruction, Nature Methods, Reconstructing, signaling