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

This paper introduces IRIS, an integrated experimental-computational framework that leverages high-throughput in vitro perturbation screens and transfer learning to accurately reconstruct in vivo single-cell signaling histories, revealing combinatorial signaling codes and accelerating stem cell differentiation optimization.

The Gist

Imagine you are trying to figure out exactly what a chef is cooking in a massive, bustling kitchen, but you can't walk inside the kitchen to see them. You can only look at the finished dishes on the table and try to guess what ingredients were used and in what order.

This is essentially the challenge biologists face when studying cells. Cells are constantly changing their "state" (what they are becoming, like a heart cell or a lung cell) based on a complex recipe of chemical signals (ingredients) they receive from their environment. For a long time, scientists could guess the recipe by looking at the ingredients (ligands) floating around, but they couldn't be sure if the cell actually used them. It's like seeing flour on the counter and assuming a cake is being baked, even if the baker is actually making bread.

This paper introduces a new, super-smart detective tool called IRIS (Intracellular Response Inferred Signaling States) that solves this mystery. Here is how it works, broken down into simple concepts:

1. The "Training Camp" (The Perturbation Atlas)

First, the scientists needed to teach IRIS how to recognize the "fingerprints" of different chemical signals. They couldn't just guess; they had to run a massive experiment.

• The Analogy: Imagine a cooking school where they take a batch of raw dough (stem cells) and force them to try every possible combination of ingredients. They mix in salt, sugar, yeast, or nothing at all, in every conceivable order.
• The Result: They created a giant "atlas" (a map) of how cells react when fed specific combinations of signals. They recorded the cells' "transcriptome" (the list of all active genes), which is like taking a photo of the cell's internal machinery to see exactly what changed.

2. The "Super-Student" (Transfer Learning)

Here is the clever part. Usually, scientists think that a cell's reaction to a signal is unique to that specific cell type. They thought a liver cell reacts to sugar differently than a brain cell.

• The Discovery: The researchers found that cells actually share a common "language" or "fingerprint" for how they react to specific signals.
• The Analogy: Think of it like learning a language. You might think that a French speaker and a Japanese speaker have totally different ways of saying "Hello." But IRIS discovered that there is a universal grammar. If you teach a computer (IRIS) how a mouse cell reacts to a signal in a lab, it can actually understand how a human cell reacts to the same signal, even though they are different species!
• The Tool: IRIS is a neural network (a type of AI) that learned these universal "fingerprints" from the lab experiments. It didn't just memorize a few genes; it learned the pattern across thousands of genes, making it incredibly robust.

3. The "Time Travel" (Reconstructing Histories)

Once IRIS was trained, the scientists took it to the real world: a developing mouse embryo.

• The Challenge: You can't stick a probe into a living mouse embryo to ask, "What signals are you feeling right now?"
• The Solution: The scientists fed IRIS the genetic data from the mouse embryo. Because IRIS knew the "fingerprints" of the signals, it could look at a single cell and say, "Ah, this cell was just exposed to Signal A, followed by Signal B."
• The Analogy: It's like looking at a person's clothes and knowing exactly where they've been today. "You have mud on your shoes (Signal A) and a ticket stub from the train station (Signal B), so you must have walked through the park and then taken the train." IRIS reconstructed the entire "travel history" of the cells as they developed from a single fertilized egg into a complex embryo.

4. The "Recipe Optimizer" (Fixing Stem Cell Therapies)

Finally, they used IRIS to fix a real-world problem: making stem cells turn into specific organ parts (like lung tissue) in a lab dish.

• The Problem: Making stem cells turn into lung tissue is like trying to bake a perfect soufflé. Scientists had a recipe, but it wasn't working well. They were guessing the timing and amounts of ingredients, which is slow and expensive.
• The IRIS Fix: IRIS looked at the "history" of how lung cells naturally develop in a mouse embryo. It realized, "Hey, the original recipe adds the 'WNT' ingredient too late!"
• The Result: They changed the lab recipe to add the WNT signal earlier. The result? They successfully created much more lung tissue than before. IRIS saved them from having to test thousands of random combinations; it told them exactly what to do.

Why This Matters

This paper is a game-changer because it turns "guessing" into "knowing."

• Before: We could guess what signals might be active, but we couldn't be sure, and we couldn't see the history of how a cell got to where it is.
• Now: We have a tool that can read the "diary" of a cell's life just by looking at its genes. This helps us understand diseases, figure out how embryos grow, and design better medicines and lab-grown organs without needing to test every single possibility in a petri dish.

In short, IRIS is a translator that lets us read the secret language of cells, understanding not just what they are doing, but exactly how they got there.

Technical Summary

1. Problem Statement

Understanding how cells transition between states during development, homeostasis, and disease relies on knowing their signaling states and histories (i.e., which combinations of external signals they have received).

• Limitations of Current Methods:
  • Genetic/Reporter Approaches: Low-throughput, difficult to apply to human tissues, and lack temporal precision.
  • Ligand-Receptor Inference (e.g., CellChat, NicheNet): These infer "signaling potential" based on gene expression of ligands and receptors but do not guarantee pathway activation or capture the actual intracellular signaling state.
  • Response Gene Scoring: Traditional methods rely on a small set of known marker genes (e.g., Axin2 for Wnt). These are often cell-type specific, sensitive to dropout in scRNA-seq data, and fail to generalize across different cell types or batches.
• The Core Question: Can we learn generalizable, transcriptome-wide "signaling response signatures" from in vitro perturbation data that allow us to accurately infer signaling states in diverse, unobserved in vivo cell types?

2. Methodology

The authors developed an integrated experimental-computational framework centered on a deep learning model called IRIS (Intracellular Response Inferred Signaling States).

A. Experimental Design: High-Throughput Perturbation Atlas

To generate ground-truth training data, the authors performed multiplexed sequential combinatorial screens on Human Embryonic Stem Cells (hESCs) and utilized existing Mouse Embryonic Stem Cell (mESC) data.

• Target Pathways: Six major developmental pathways: TGF-β, FGF, BMP, Hedgehog (HH), Retinoic Acid (RA), and Wnt.
• Screen Design:
  • hM_d4 & hM_d7 (Mesoderm): Cells were differentiated into primitive streak and then subjected to sequential combinatorial stimulation. The hM_d7 screen specifically varied RA and SHH first to generate splanchnic mesoderm subtypes, then applied all 16 combinations of the remaining four pathways.
  • hE_d8 (Endoderm): Cells were differentiated to definitive endoderm and exposed to all 64 possible combinations of the six pathways.
  • Multiplexing: Each condition was barcoded and pooled for single-cell RNA sequencing (scRNA-seq), creating a rich atlas of ~40,000 cells with known signaling histories.

B. Computational Model: IRIS

IRIS is a semi-supervised Conditional Variational Autoencoder (CVAE) architecture (based on scANVI).

• Architecture:
  • Encoder: Maps input gene expression ($X$) and conditioning variables (batch, cell type, species) to a latent space ($Z$).
  • Decoder: Reconstructs the gene expression from $Z$ and conditions.
  • Classifier: A feedforward neural network (FNN) trained jointly on the latent space to predict binary signaling states (ON/OFF) for each pathway.
• Training Strategy:
  • The model learns to remove batch effects and embed cells with similar expression patterns together.
  • It is trained on the perturbation atlas where signal conditions are known (labeled) to learn the transcriptome-wide signature of each pathway's activation.
  • Transfer Learning: Once trained, IRIS is applied to unlabeled in vivo datasets (e.g., mouse gastrulation atlas) to infer signaling states without prior knowledge of the signal conditions.

C. Validation & Benchmarking

• Cross-Batch/Cell-Type/Species: Tested by holding out entire screens (e.g., training on mouse, testing on human) or specific cell lineages (endoderm vs. mesoderm).
• Baselines: Compared against simple response-gene scoring, Elastic Net, Random Forests, and SVMs.
• Feature Importance: Used gene ablation tests and saliency maps (gradients) to determine if the model relies on a few genes or the whole transcriptome.

3. Key Contributions

1. Discovery of Transferable Signaling Signatures: The study challenges the dogma that signaling responses are strictly cell-type specific. It demonstrates that transcriptome-wide response signatures for major pathways (Wnt, RA, BMP, etc.) are conserved and transferable across species (mouse/human) and distinct cell lineages (mesoderm/endoderm).
2. The IRIS Algorithm: A robust deep learning framework that outperforms traditional machine learning models and response-gene methods, particularly in data-limited, cross-species scenarios.
3. High-Resolution Signaling Histories: The ability to reconstruct the temporal sequence of combinatorial signaling codes along developmental trajectories at single-cell resolution.
4. Protocol Optimization: A practical application where IRIS predictions were used to reverse-engineer and optimize stem cell differentiation protocols for specific tissue types.

4. Key Results

• Superior Accuracy: IRIS achieved significantly higher F1 scores and AUROC compared to response-gene methods and other ML models (SVM, RF) in cross-validation and cross-species tests.
• Generalization:
  • Cross-Species: Models trained on mouse data successfully inferred signaling states in human cells (and vice versa), especially when a small amount of in-species data was included for fine-tuning.
  • Cross-Lineage: Models trained on endodermal screens accurately predicted signaling states in mesodermal cells.
• Transcriptome-Wide Features: Gene ablation tests showed that thousands of genes contribute to the prediction accuracy. The model does not rely on a small set of "context-independent" markers but captures distributed information across the gene regulatory network.
• Reconstruction of Developmental Dynamics:
  • Applied to the mouse gastrulation atlas (E6.5–E8.5), IRIS revealed a global transition in signaling codes around E7.0.
  • It successfully reconstructed known temporal trends (e.g., Wnt inactivation preceding cardiomyocyte differentiation) and identified heterogeneity within cell clusters (e.g., RA+ cells in the cardiomyocyte cluster were enriched for atrial markers).
  • It reconstructed spatial gradients in the somitic mesoderm, correlating predicted signaling probabilities with the anterior-posterior axis.
• Optimization of Respiratory Mesenchyme Differentiation:
  • IRIS predicted that Wnt signaling is required earlier in the differentiation of respiratory mesenchyme than previously thought.
  • Experimental Validation: Modifying the hESC differentiation protocol to include earlier and longer Wnt stimulation (using CHIR) significantly increased the yield of respiratory mesenchyme (marked by TBX4 and FOXF1), validated via HCR-FISH in both mouse explants and human stem cells.

5. Significance

• Conceptual Leap: This work shifts the paradigm from inferring signaling "potential" (ligand-receptor pairs) to inferring actual "signaling state" based on learned transcriptomic responses.
• Scalability: It provides a scalable solution to map complex signaling interactions in in vivo tissues without the need for expensive, low-throughput perturbation experiments in every specific cell type of interest.
• Therapeutic & Engineering Impact: The framework accelerates the design of stem cell differentiation protocols (e.g., for organoids) by drastically reducing the combinatorial space that needs to be experimentally tested.
• Foundation for Future Models: The findings suggest that foundational models for cell biology can be built by learning generalizable principles of signal transduction across diverse contexts, paving the way for predictive modeling of cell-cell communication and disease mechanisms.

In summary, the paper presents a powerful tool (IRIS) that bridges the gap between in vitro perturbation screens and in vivo biological complexity, enabling the reconstruction of cellular signaling histories with high accuracy and temporal resolution.