Key points from article :
The researchers developed a new AI model, dubbed GPT‑4b micro, that is optimized for protein engineering. It was built by starting from a smaller version of GPT‑4 and then trained specially on large datasets of protein sequences, structural‑data and biological text to target difficult tasks such as engineering the famous “Yamanaka factors” (e.g., SOX2 and KLF4) that can reprogram adult cells into induced pluripotent stem cells (iPSCs).
Using this AI model in partnership with Retro Biosciences, the team generated novel variants of SOX2 and KLF4, referred to as “RetroSOX” and “RetroKLF”. These variants showed over a 50‑fold increase in expression of stem‑cell reprogramming markers in vitro compared with wild type controls, and exhibited enhanced DNA‑damage‑repair capability — a hallmark relevant for rejuvenation and longevity.
This work suggests that domain‑specific AI tools can dramatically accelerate biological research: what might have taken years by traditional methods may now be achieved in a matter of months. The article highlights that the model’s hit rate (percentage of successful variants) was much higher than conventional directed‑evolution methods.
The authors consider this a proof‑of‑concept and stress that further work is needed before clinical applications — but the results point to a new paradigm for integrating AI and biotechnology.
