Westlake University Joins Hands with DAMO Academy to Release “Guiyuan”: Letting AI Do 4 Million Experiments for Stem Cell Researchers

On July 14, Westlake University and Alibaba DAMO Academy jointly released an AI model called “Guiyuan”, designed specifically to predict stem cell fate. The name is meaningful—“Guiyuan” means “returning to the origin,” which perfectly mirrors what stem cell reprogramming tries to do: pull an already differentiated cell back to a pluripotent starting point with multiple developmental potentials.

It might sound abstract, but its significance in biology is profound. In 2006, Shinya Yamanaka reprogrammed skin cells into induced pluripotent stem cells (iPSC) with four transcription factors and later won the Nobel Prize. Two decades later, finding a precise way to guide cells into a specific state remains one of the most time-consuming and costly tasks in the lab.

4 Million Combinations That Would Take Decades for Humans to Test

First, let’s be clear about how hard this problem is.

Stem cell reprogramming involves 25 lineage-regulating factors, including 17 small-molecule compounds and 8 protein growth factors. These factors can be combined in pairs, triplets, or quartets, with concentration adjustments—leading theoretically to nearly 4 million possible experimental setups.

Gu Fei, a senior algorithm expert at DAMO Academy, gave an intuitive description: if we were to test each combination experimentally, it could take decades. Moreover, the traditional approach relies heavily on researchers’ experience and intuition—try A+B today, fail; then A+C tomorrow, still fail. It burns money, time, and graduate students’ patience.

This is a textbook combinatorial explosion problem—and solving such explosions is exactly what deep learning excels at. AlphaFold’s protein structure prediction, in essence, searches for answers within astronomically large conformational spaces. The idea behind Guiyuan is, in a sense, to transfer that same methodology to predicting cell fate.

Dual-Modality Encoding: How to Put Small Molecules and Proteins into the Same Model

The core technical design of Guiyuan is its dual-modality encoding strategy.

Here’s an unavoidable engineering challenge: small-molecule drugs and protein growth factors are completely different types of substances. Small molecules have simple structures—tens of atoms forming a backbone that can be represented by a chemical formula or SMILES string—while proteins consist of hundreds or thousands of amino acids folded into complex 3D structures, with properties determined by sequence.

Guiyuan’s approach is:

• For small molecules, it uses molecular structural representation encoding—essentially converting the chemical graph structure into a vector.
• For growth factors and cytokines, it uses a protein language model—similar to ESM or ProtBERT—to encode amino acid sequences into semantic vectors.
• The outputs of these two pathways are projected into the same high-dimensional representation space.

Once unified in the same space, the model can understand how combinations like “small molecule A + protein B” jointly affect cells. While this multimodal fusion idea isn’t entirely new—CLIP and Flamingo already use such techniques for vision-language tasks—applying it to predict combinatorial perturbations in stem cells required heavy re-engineering efforts.

Interpretability: Not a Black Box—It Can Tell You “Why It Works”

This point deserves special emphasis.

In biomedicine, the acceptance of AI is completely different from the internet world. In online recommendation systems, if the model works well, nobody cares why a user clicked a video. But biology is different—no matter how accurate a predicted combination is, if it can’t explain which signaling pathways are involved, researchers can’t design follow-up experiments, and peer reviewers won’t be convinced.

Guiyuan includes an interpretability module that maps predictions to known biological signaling pathways. In other words, AI doesn’t just say “the A+B+C combination can push cells toward the hypoblast state,” but also explains “because it activates pathway XXX and inhibits pathway YYY.”

This interpretability is even more crucial for practical use than accuracy. For researchers, a model that explains mechanisms is a real tool that can integrate into the research workflow—rather than an inscrutable oracle.

Experimental Validation: Cultured Hypoblast-like Cells Can Be Passaged 50 Generations

A model alone isn’t enough; experimental validation is essential.

The research team had Guiyuan simulate all ~4 million combinations, selected the optimal ones, and then actually carried out the lab experiments—producing hypoblast-like stem cells.

The hypoblast is rather special. In natural embryos, it briefly appears between days 5–7 after fertilization, handling key tasks like nutrient support, signal transmission, implantation, and early yolk sac formation—before rapidly differentiating and disappearing. Because it’s so transient, researchers have long struggled to study it systematically.

The hypoblast-like stem cells cultivated under Guiyuan’s guidance underwent several key tests:

1. Molecular characteristics highly resembled natural hypoblast cells.
2. They stably expressed key pluripotency factors.
3. They maintained stem-cell functionality after 50 successive passages in vitro.

That last point is particularly important. Being able to passage for 50 generations means these cells can serve as long-term, stable research material—not just a one-time experiment. Researcher Liu Xiaodong from Westlake University noted that hypoblast abnormalities are closely linked to recurrent miscarriages and implantation failures. In the past, such cells couldn’t be stably cultured in vitro; now, with a reliable source, follow-up studies on in vitro hematopoiesis, embryo-like structure construction, and cell therapy finally have a concrete starting point.

“Foundation Model” Positioning: A Big Ambition

Notably, the team describes Guiyuan as a foundation model, not a specialized model limited to the hypoblast.

Here’s the distinction: a specialized model solves a specific task, while a foundation model learns general principles of cell fate transitions, theoretically enabling transfer to other reprogramming tasks. The hypoblast result is just the first; applications for cardiac, neural, or hematopoietic directions are likely next.

From this perspective, Guiyuan’s strategy is conceptually aligned with the AlphaFold family—first solving a fundamental representation and prediction challenge, then letting it branch into downstream applications. For a Chinese team to deliver a foundation model of this caliber in bio-AI is quite impressive.

Some Observations

AI for Science has been a hot topic in recent years, but few projects have delivered truly convincing results. Guiyuan stands out for several solid reasons:

• Well-defined problem: 4 million combination optimization, where AI has clear advantages.
• Self-built dataset: The team constructed a large-scale combinatorial perturbation dataset rather than relying on public data.
• Closed-loop validation: Model prediction → experimental verification → real, viable cells—a complete chain.
• Interpretability: Not just “good performance,” but grounded in biological mechanisms.

Of course, there are still aspects to watch—such as how well the model generalizes to unseen factor combinations, and whether the pathway analyses from the interpretability module can truly withstand expert scrutiny. These will require more publications and experimental data.

For developers interested in bio-AI, Guiyuan sends a clear signal: multimodal fusion + domain-specific priors + large-scale perturbation data is a working recipe for scientific domains. Similar ideas could be applied to drug screening, material synthesis, or enzyme design—any field facing combinatorial explosion challenges.

Next time a combinatorial optimization problem lands on your desk, don’t test combinations one by one.

References

• Westlake University and Alibaba DAMO Academy Launch Stem Cell AI Model “Guiyuan,” Enabling Precise Regulation of Cellular “Reverse Growth” - IT Home: IT Home’s original report on the Guiyuan model release, including technical details and interviews with the research team.