Thread by @bryan_johnson

This is big. OpenAI and Retro used a custom model to make cellular reprogramming into stem cells ~50× better, faster, and safer. Similar Wright brothers’ glider to a jet engine overnight.

We may be the first generation who won't die.

Let's take a look at what they did. 🧵

0/ AI Redesigns the Engines of Cellular Reprogramming

OpenAI and Retro Biosciences reported a landmark achievement: using a domain-specialized protein design model, GPT-4b micro, they created engineered reprogramming factors that deliver over 50× higher efficiency in generating induced pluripotent stem cells (iPSCs), with broad validation across donors and cell types. These AI-designed proteins not only accelerate reprogramming but also enhance DNA repair, overcoming DNA damage as one cellular hallmark of aging hinting at relevance for aging biology.

1/ 0/ From Yamanaka’s Discovery to Today

In 2006, Shinya Yamanaka identified four transcription factors; Oct4, Sox2, Klf4, and c-Myc (OSKM), capable of resetting differentiated adult cells into a pluripotent, embryonic-like state. This breakthrough opened two revolutionary doors in biology:

1) Full reprogramming: completely erasing cell memory, identity, to generate pluripotent stem cells (iPSCs) for regenerative medicine, tissue engineering, and disease modeling.

2) Partial reprogramming: transient expression to roll back cellular age without erasing cell identity.

This remarkable discovery won Yamaka the Nobel Prize for Physiology or Medicine in 2012.

Yet, in practice, OSKM are astonishingly inefficient (<0.1% conversion, requiring +3 wees of culture) and carry oncogenic risks, particularly from c-Myc. These limitations have long restricted their therapeutic potential. These limitations have limited the progress in regenerative medicine due to the difficulty of scaling the safe and reliable production induced pluripotent stem cells.

2/ 1/ OpenAI × Retro: Breaking the Bottleneck

To overcome these barriers, OpenAI built GPT-4b micro, a compact model derived from GPT-4o but specialized for protein design.

Initialized from a scaled-down GPT-4o, GPT-4b micro was trained by blending protein sequences, structural embeddings (3D folds), evolutionary relationships, and biological text into a single representational space. It was then instruction-tuned to follow biological design prompts (“create SOX2 variants with improved pluripotency induction”) and iteratively refined through a design–test–feedback loop with Retro’s wet labs.

Crucially, this allowed the model to sample protein sequence space probabilistically, something humans cannot do. For perspective:

1) SOX2 (317 amino acids) and KLF4 (513 aa), design space in the magnitude of 10^1000.

2) Traditional protein engineering might explore dozens to hundreds of variants, tweaking a handful of amino acids, and even directed evolution only explores a small slither of the space as it only blindly mutates a handful of residues at each time.

3) GPT-4b micro generated hundreds to thousands of variants per factor, some diverging by >100 amino acids. This means the model intelligently sampled regions of sequence space ~10^100 orders of magnitude further from wild-type than classical methods while still delivering a 30–50% functional hit rate.

This training paradigm transforms protein design from incremental trial-and-error into exploratory leaps guided by structure, evolution, and function.

3/ Key Results (AI-generated KLF4 and SOX2 vs. traditional OSKM Controls)

Variant hit rate
+ Expert-driven design of OSKM: < 10% of engineered variants outperform wild-type.
+ GPT-4b micro: 30% of SOX2 and 50% of KLF4 variants outperformed wild-type.

Pluripotency induction efficiency
+ Wild-type: < 0.1% of cells reprogramed; weak marker expression after ~3 weeks.
+ Retro variants: > 50× higher pluripotency marker expression, with ~30% conversion rates in aged donor cells.

Early pluripotency markers (Day 7)
+ Wild-type: negligible expression of SSEA-4 and TRA-1-60.
+ Retro variants: > 30% of cells expressed SSEA-4 and TRA-1-60.

Late pluripotency markers (Day 12)
+ Wild-type: weak, heterogeneous activation of OCT4, NANOG, SOX2, TRA-1-60.
+ Retro variants: ≥ 85% of cells activated these markers, showing robust and synchronized pluripotency.

Colony formation
+ Wild-type: sparse alkaline phosphatase+ colonies, emerging after ~21 days.
+ Retro variants: dense colonies by Day 10, more than twice as fast.

DNA damage repair
+ Wild-type OSKM: strong γ-H2AX signal under genotoxic stress (double-strand breaks).
+ RetroSOX/KLF: significantly reduced γ-H2AX, indicating up to 4x more efficient DNA repair and genomic stability.

Safety point: generate stem cells showed normal chromosomal structure (caryotype), a good early proof of the safety of the generated stem cells, which also converges with tighter DNA maintenance and repair. Cancerous transformation is a big concern when reprogramming cells back to stem cells, chromosome anomalies are key hallmarks of cancer.

4/ Significance and caveats of the OpenAI-Retro approach

Advantages to regenerative medicine: The discovery removes long-standing barriers to regenerative medicine by enabling a scalable, efficient, and safer supply of iPSCs for use in tissue, organ, and cell therapy development.

Caveats
+ Still too strong for in vivo rejuvenation: These redesigned factors are best suited for ex vivo regenerative medicine. Direct systemic use risks loss of cell identity and tumorigenesis.

+ Narrow hallmark validation: The aging link was shown via improved DNA damage repair, important but not a full model of biological aging.