- ReproPerturb predicts optimal drug and growth factor combinations from nearly 4 million possibilities, dramatically reducing trial-and-error in stem cell research
- The explainable AI model has already enabled scientists to create long-term extraembryonic endoderm-like stem cells and optimize cell therapy development
Alibaba’s research arm DAMO Academy and Westlake University on July 14 jointly unveiled ReproPerturb (归元), an AI model designed to predict how combinations of small-molecule drugs and growth factors can reprogram cells into specific stem cell states.
This release could potentially transform one of regenerative medicine’s most labor-intensive research processes.
The model searches through nearly 4 million possible combinations of small molecules and growth factors to identify the most effective recipe for steering cells toward a desired stem cell fate.
The work builds on the landmark 2006 discovery that mature skin cells can be reprogrammed into induced pluripotent stem cells (iPSCs) — effectively returning specialized adult cells to a primitive state capable of developing into many different cell types.
During this reprogramming process, cells remain highly plastic. Introducing different combinations of small-molecule compounds and protein growth factors can direct them toward distinct stem cell lineages.
Finding those combinations, however, has traditionally relied on researchers’ experience and years of trial and error. The study involved 25 lineage-regulating factors, creating a theoretical search space of nearly 4 million combinations.
“If every combination had to be validated experimentally using conventional methods, it could take decades, with enormous costs and a low success rate,” said Gu Fei, a senior algorithm expert at DAMO Academy.
Protein language models
The model’s key innovation lies in its dual-modal encoding strategy. It represents small-molecule drugs through molecular structure encoding while using protein language models to encode growth factors and cytokines.
Both are projected into the same high-dimensional representation space, allowing the model to predict how different combinations influence cell fate.
Equally important, ReproPerturb incorporates an explainability module that links its predictions to known biological signaling pathways.
Rather than simply identifying promising combinations, it also provides clues as to why they are likely to work.
Using AI-recommended protocols, the research team successfully generated long-term expandable extraembryonic endoderm-like stem cells in vitro for the first time.
The cells closely resemble their natural counterparts at the molecular level and retained their stem cell characteristics after 50 generations of passaging.
Extraembryonic endoderm cells exist naturally only during a brief window — roughly five to seven days after fertilization — making them difficult to study systematically.
Liu Xiaodong, a researcher at Westlake University, said the breakthrough could improve understanding of early human embryonic development while advancing research into in vitro blood formation, embryo-like structures and cell therapies.
The team has already begun applying ReproPerturb to other cell fate engineering tasks, including generating dopaminergic neuronal precursor cells for Parkinson’s disease therapies and optimizing manufacturing processes for complex cell therapy products.
Accelerator for stem cell research
Stem cell research has long faced a fundamental bottleneck: the combinatorial space governing cell fate decisions is simply too vast for conventional experiments to explore efficiently.
Globally, the convergence of AI and stem cell biology has accelerated rapidly. A study covering 2,866 papers published between 1994 and 2025 found annual growth of about 20%, with the United States, China, Japan, Germany and the United Kingdom emerging as the leading contributors.
Unlike many AI applications focused on downstream cell identification or quality control, ReproPerturb intervenes at an earlier stage by helping design optimal reprogramming strategies before cells acquire their final identities.
Why it matters for global readers
Twenty years after Nobel laureate Shinya Yamanaka’s discovery of iPSC technology, cell reprogramming remains as much an art as a science.
ReproPerturb represents one of the first attempts to combine large-scale combinatorial perturbation experiments, explainable AI and stem cell biology into a reusable framework for controlling complex cell fate decisions.
By replacing “needle-in-a-haystack” experimentation with computational prediction, the model could help shift stem cell research from an experience-driven craft toward a more predictable, data-driven engineering discipline.
For the global regenerative medicine and cell therapy community, its significance extends beyond identifying effective molecular combinations.
By also explaining why particular combinations work, the model addresses one of the biggest challenges facing future clinical applications, where reproducibility, controllability and biological interpretability will be critical.
