AI designs proteins to accelerate stem cell reprogramming

OpenAI and biotech company Retro Biosciences used an AI model to redesign proteins that turn adult cells into stem cells. In laboratory tests, some variants activated reprogramming markers more than 50 times higher than natural proteins, according to results published by both companies.

The model is called GPT-4b micro and was created specifically for protein engineering. It is not publicly available, nor is it a tool that lets you make treatments at home: it proposed protein sequences that were then synthesized and tested by Retro scientists in human cells.

Which proteins the AI redesigned

The work focuses on the so-called Yamanaka factors: OCT4, SOX2, KLF4, and MYC. These four proteins can return an adult cell to a state similar to that of a pluripotent stem cell, which can develop into different types of tissue.

The original discovery, made by Shinya Yamanaka, received the Nobel Prize in Physiology or Medicine in 2012. Since then, researchers have studied these proteins to create cells for studying diseases, testing drugs, and developing potential regenerative therapies.

The problem is that the process remains slow and inefficient. With natural factors, usually less than 0.1% of cells are reprogrammed, and the procedure can take three weeks or more. Efficiency drops even further when the cells come from older people or people with diseases.

How the model contributed

GPT-4b micro is based on a reduced version of GPT-4o, but it was trained mainly on protein sequences, biological texts, and three-dimensional structure data. It also received information about protein evolution, their functions, and the molecules they interact with.

This makes it possible to ask for something more specific than a simple prediction. For example, the model can generate variants of SOX2 that preserve its function while improving the activation of certain cellular markers.

The model worked with contexts of up to 64,000 tokens, an especially large amount of information for a protein sequence model. That helps it analyze a protein alongside data on related proteins and possible interactions, rather than treating it as an isolated chain of amino acids.

Results in human cells

Retro first tested model-generated SOX2 variants in fibroblasts, cells found in the skin and connective tissue. More than 30% of the proposals outperformed the natural protein in the expression of markers associated with pluripotency. In traditional screenings, success rates are usually below 10%, according to the announcement.

The team then did the same with KLF4, another Yamanaka factor. Nearly 50% of the AI-generated variants outperformed the best combinations obtained in the previous phase.

When the best versions of SOX2 and KLF4 were combined, pluripotency markers appeared several days earlier than they did with the natural factors. In another test, conducted with messenger RNA instead of viral vectors and with cells from three donors over 50 years old, more than 30% of the cells began expressing key markers within seven days.

The researchers also confirmed that the resulting cells could develop into the three major types of embryonic tissue. In addition, the cell lines maintained healthy karyotypes and genomic stability over several rounds of culturing, two important checks before considering medical applications.

They also observed signs of rejuvenation

The team analyzed whether the redesigned proteins could reduce DNA damage, one of the signals associated with cellular aging. Cells treated with the improved combination showed lower γ-H2AX intensity, a marker of DNA breaks, than cells treated with the standard factors.

That result points to a possible use in cellular rejuvenation research. But it does not show that a safe or effective anti-aging therapy exists for people. The tests described are laboratory experiments, and risks such as unintended changes, tumors, immune responses, and how the cells behave inside an organism still need to be evaluated.

What actually changes

The importance of the work lies in the scale of the search. A protein such as KLF4 can contain hundreds of amino acids, and the number of possible combinations is so large that testing them one by one is impractical. AI can propose designs that differ substantially from natural proteins and reduce the number of experiments needed to find useful candidates.

That does not eliminate the laboratory. AI suggests sequences, but scientists must produce them, test them in cells, discard those that fail, and verify that the ones that work do not introduce safety problems.

OpenAI says the results were replicated across several donors, cell types, and delivery methods. Even so, this is an announcement from the organizations that conducted the study. Sam Altman, OpenAI's chief executive, is also an investor in Retro Biosciences, a relationship worth considering when assessing the scope of the conclusions.

The next step will be to determine whether these variants retain their advantages in animal models and, much later, in clinical trials. For now, the most relevant signal is specific: a specialized model can greatly expand the search for useful proteins, but the leap from a laboratory dish to a treatment still requires years of validation.