When AI takes up the brush of life | OpenAI italiano | OpenAI Chat | ChatGPT 4 | Turtles AI

When AI takes up the brush of life
New variants of Yamanaka factors created from a GPT template multiply reprogramming efficiency and enhance DNA repair
Editorial Team27 August 2025

 

OpenAI and Retro Biosciences announce engineered variants of Yamanaka factors, designed from a GPT-4b micro model, that increase reprogramming markers more than fiftyfold and improve DNA repair, opening up avenues for clinical regenerative medicine and anti-aging.

Key Points:

  • 50x increase in expression of reprogramming markers compared to "natural" factors.
  • "RetroSOX/RetroKLF" variants with reduced γ-H2AX signaling after DNA damage.
  • GPT-4b micro model, initialized from GPT-4o and trained on sequences, biological tests, and structural data.
  • Replicates across different donors and cells; validation with mRNA, full pluripotency, and healthy karyotype.


It’s as if someone has removed the handbrake on cellular reprogramming: in collaboration with Retro Biosciences, a custom OpenAI GPT-4b micro model has proposed new sequences for two of Yamanaka’s famous factors, SOX2 and KLF4, producing a more than fifty-fold increase in pluripotency markers in vitro compared to "wild-type" controls; not a simple statistical embellishment, but a visible acceleration of the cells’ entry into the iPSC state, historically slow and inefficient (often on the order of 0.01–1%). The announcement was made by OpenAI’s head of applied research, Boris Power, who leads the company’s scientific partnerships in a project that puts AI at the forefront of experimental biology. OpenAI’s tech blog explains that GPT-4b micro is derived from a stripped-down version of GPT-4o, then refined with a corpus largely of protein sequences, biological texts, and even "tokenized 3D" structures, enriched with evolutionary and functional context to guide the generation of proteins with targeted properties; a useful approach for "messy" targets like OSKM factors, whose activity arises from networks of transient interactions rather than a single stable fold. The practical result is the variants dubbed RetroSOX and RetroKLF: in addition to unlocking late markers like TRA-1-60 and NANOG as early as around day 10, they have passed independent testbeds with different delivery approaches (not only viral vectors but also mRNA) and in mesenchymal cells from middle-aged human donors, demonstrating that the advantage is not a laboratory chimera. To crown the identity test, the generated colonies activated endogenous pluripotency genes (OCT4, NANOG, SOX2, TRA-1-60), differentiated into the three germ layers, and retained normal karyotypes, i.e., genomic stability, a prerequisite for any clinical perspective. The most intriguing piece of the puzzle, however, concerns rejuvenation: in a doxorubicin-induced genotoxic damage assay, cells treated with RetroSOX/KLF showed a significantly lower γ-H2AX signal compared to standard factors or the fluorescent control, a clue that the new "recipe" enhances the ability to repair double-strand breaks, a cornerstone of cellular aging. Placed within the broader framework, this work touches a sensitive nerve in regenerative medicine: the slowness and chronic inefficiency of reprogramming, known for years, have hindered the industrial adoption of iPSCs; If a targeted model can generate dozens of candidates and achieve high hit rates with even profound modifications (tens or hundreds of amino acids compared to natural ones) and cross-platform validation, then AI stops being just a "textbook brain" and becomes a co-designer of functional biology. Of course, we’re still in the early stages: "in vitro" doesn’t mean therapy; safety, off-target, oncogenicity, and the ability to scale to GMP processes still need to be mapped; but the indication is clear: combining expert domain (wet lab) and vertical models (protein LMs) can transform multi-year problems into cycles of days or weeks. Meanwhile, history puts the numbers into perspective: from the 0.01–0.1% of the pioneering era, to today’s wave of boosters (small molecules, barrier-busting strategies, transient protocols), up to these factors "reimagined" by AI that multiply reprogramming signals and mitigate DNA damage, opening concrete avenues for tissue rejuvenation therapies, degenerative diseases, and faster production of quality iPSC lines.

If Retro’s stated goal is to gain years of healthy life, this could be one of the first screws tightened in the pipeline.