HomeFull speed ahead. Keep track of advancements in every frontier

TheTechno Optimist.

RTR242 (autophagy)Retro Biosciences—Retro Biosciences—

GPT-4b micro (with OpenAI)

by Retro BiosciencesUS

ResearchStage 1 of 5

Research model used in-house to design better reprogramming factors.

Updated Jan 2025Checked 25 Sep0 updates this week

Milestones

Next · Use them to make cell therapies
  1. Collaboration revealedJan 2025Complete.
  2. Results for AI-designed factors sharedAug 2025Complete.
  3. Use them to make cell therapiesNowCurrent milestone.

Most important updates

  • Jan 2025

Physics limits

  • Protein space is astronomically largeA 300-amino-acid protein has 20^300 possible sequences. Models can only explore near proteins they have seen, and every design still needs a lab test to show it works.
  • Little training data for this exact taskFew experiments measure how well each factor variant reprograms cells, so predictions get unreliable far from known examples. Lab testing speed sets the pace.
  • Stronger factors carry the same risksFactors that reprogram faster also push cells faster toward tumour-forming stem cells. Better efficiency does not widen the safety window.

How it works

3 parts
An NVIDIA DGX B200 GPU server: the kind of hardware large AI models such as GPT-4b micro are trained and run on
An NVIDIA DGX B200 GPU server: the kind of hardware large AI models such as GPT-4b micro are trained and run onPhoto: Pokiiri · CC BY-SA 4.0 (opens commons.wikimedia.org)
Learn

Trained on protein sequences

A small version of GPT-4o is trained on protein sequences and biology text, learning which amino-acid changes tend to keep a protein working.

Propose

Rewrite the reprogramming factors

It proposes new versions of the SOX2 and KLF4 factors, some differing from nature by over 100 amino acids.

Test

Screen them in real cells

Retro tests designs in human cells; the best (RetroSOX, RetroKLF) raised stem-cell markers over 50-fold versus the natural factors.

Papers & demos

2 items
  1. Aug 2025blog
    Accelerating life sciences research with engineered reprogramming factors (OpenAI and Retro)AI-designed versions of two Yamanaka factors worked much better in lab cells.
  2. Jan 2025demo
    GPT-4b micro protein engineering model revealed (opens technologyreview.com)One of the first AI models OpenAI built just for biology.

Update log

1 update

Jan 2025

  • Major: PressResearch

About Retro Biosciences

The team behind GPT-4b micro (with OpenAI)

Retro Biosciences

Cell-cleanup pill, young cell transplants, AI

Bay Area company founded in 2021 by Joe Betts-LaCroix with $180M from Sam Altman, aiming to add ten healthy years to human life. It makes young replacement cells for blood and brain, and RTR242, a pill for Alzheimer's that entered trials in 2025.

  • RTR242 went from choosing a disease to the first human dose in about 15 months (trial in Adelaide, Australia).
  • In 2025 Retro and OpenAI reported AI-designed versions of two Yamanaka factors that worked better in lab cells.
  • May 2026: first close of a new round at a $1.8B valuation, led by 4P Capital.

retro.bio (opens retro.bio)Company profile →

Founded
20215 yrs
Headquarters
United States
Status
Private
Valuation
$1.8BprivateMay 2026
Raised
$180M2 rounds
Last round
New round (initial close)May 2026
Works in
LongevityCellular reprogrammingAI
Coverage
3 programs · 5 updateslatest 1 May 2026checked 25 Sep
Lead investors
4P CapitalSam Altman
Partners
OpenAI (opens openai.com)
People
Joe Betts-LaCroix (opens retro.bio)Co-founder and CEOSam Altman (opens en.wikipedia.org)Founding investor
Primer