IRIS model from the Whitehead Institute decodes cell signals during embryo development

On September 8, a team from the Whitehead Institute for Biomedical Research, affiliated with MIT, published a description of an artificial intelligence model called IRIS in the journal Nature Methods. The tool recognizes the distinctive pattern of gene activity that each signaling pathway leaves in a cell, and uses it to reconstruct which signals reached that cell at each stage of embryonic development.

Cells communicate with each other through signaling pathways, chains of molecules that carry information from outside a cell into its interior and switch on specific genes. The problem is that the same signaling pathway triggers different responses in different cell types, so until now it has been difficult to determine which signal was responsible for a given effect by looking only at gene activity.

A fingerprint for every signal

The Whitehead Institute team found that despite these differences, each signaling pathway leaves a unique pattern of gene activity in the cell, a kind of fingerprint that recurs regardless of cell type. This finding became the basis for building a model that can recognize that pattern and attribute it to a specific signal.

To teach the model to distinguish between these patterns, the researchers first built an extensive experimental atlas. They exposed thousands of human embryonic stem cells to various combinations of six major signaling pathways at multiple stages of development and recorded how their gene activity changed. This dataset served as the training material for the IRIS neural network.

Transferring knowledge across cell types

A key challenge was that a model trained on stem cells also had to work on entirely different cell types it had never seen before. Dr. Pulin Li compared this mechanism to speech recognition systems.

Think of voice recognition systems like Siri, which are trained mostly in English but then use that training to recognize other languages. That's called transfer learning, and it's why IRIS can work across many different cell types - Dr. Pulin Li, Whitehead Institute

Thanks to this approach, the model, after being trained on stem cells under laboratory conditions, also proved effective when analyzing individual cells taken from mouse embryos at the gastrulation stage, the point at which cells begin to differentiate into the tissues that form the heart, gut, muscles or spinal cord.

Why this matters to scientists

Until now, determining which combinations of signals lead to the formation of a specific tissue type required painstaking testing of successive variants in the lab. IRIS is meant to drastically shorten this work by narrowing down the combinations that actually need to be tested experimentally.

That's how IRIS is helping us decode the language cells use to talk to each other, much faster than we could realistically achieve through experiments alone - Nicholas Hutchins, graduate student, Whitehead Institute

Practical applications include, above all, the design of organoids, miniature lab-grown models of tissues and organs used to study diseases and test drugs without experimenting on patients. The authors point in particular to respiratory diseases such as asthma, lung cancer and pulmonary fibrosis as areas where a better understanding of cell signaling could translate into new therapies.

Implications for biotechnology

The model could also speed up the development of stem-cell-based regenerative therapies, where precisely controlling what type of tissue a given cell turns into is crucial. Instead of testing dozens of signal combinations by trial and error, researchers could use IRIS's predictions as a starting point.

For Polish research centers and biotech companies working in tissue engineering or drug testing, this is a sign that AI tools supporting the design of biological experiments are becoming increasingly precise and available as scientific publications with accompanying code, making them easier to adapt outside their home laboratory.

The Whitehead Institute team has made the model's code publicly available, allowing other labs to run their own tests and further develop the tool. It's another example of how machine learning models are entering basic biological research, bringing methods known from speech or image recognition to the analysis of processes taking place inside individual cells.