Company of the week: Retro Biosciences
Retro Biosciences is a San Francisco-based longevity biotech founded in 2021 with a $180 million seed round funded entirely by OpenAI CEO Sam Altman, pursuing cellular reprogramming, autophagy, and cell replacement to add ten healthy years to the human lifespan.
Retro Biosciences is a San Francisco-based longevity biotechnology company founded in 2021 with a $180 million seed round funded entirely by OpenAI chief executive Sam Altman, pursuing the stated mission of adding ten healthy years to the human lifespan through cellular reprogramming, autophagy enhancement, and cell replacement therapies.
The company sits at the intersection of two of the most heavily capitalised narratives in current technology investing: the longevity moonshot and the application of large language models to drug discovery. As of May 2026 it has just closed new financing at a $1.8 billion valuation, dosed the first patient in its first human clinical trial, and published a widely discussed protein engineering collaboration with OpenAI. It has also fallen well short of the $5 billion valuation it was reportedly targeting only five months earlier.
This piece examines Retro's origins and single-benefactor capital structure, the four-program scientific pipeline, the OpenAI GPT-4b micro collaboration and what it does and does not demonstrate, the RTR242 Alzheimer's trial now running in Australia, the May 2026 financing and the gap between its target and realised valuation, the competitive longevity landscape, and what the cumulative picture signals about a pre-revenue, pre-efficacy company carrying a multi-billion-dollar mark.
The Origin Story: A Single Benefactor and an Audacious Mission
Retro Biosciences was founded in 2021 by Joe Betts-LaCroix, alongside co-founders Sheng Ding and Matt Buckley, and emerged from stealth in mid-2022 with an unusual capital structure: a single $180 million seed round provided in its entirety by Sam Altman, with no syndicate of venture investors.
The founding narrative has been told repeatedly by Betts-LaCroix. Altman, interested in plasma dilution work, suggested funding a company around it; Betts-LaCroix was already pursuing cellular reprogramming; and Altman's response was reportedly to suggest doing both at once inside a single multi-program company. That decision, to chase several aging hypotheses simultaneously rather than concentrating on one, has defined the company's structure ever since.
The early operational aesthetic was deliberately scrappy. Retro set up in a warehouse near San Francisco and bolted shipping containers to the concrete floor to create lab space quickly for the scientists it recruited. For a period the company also kept Altman's involvement quiet, on the reasoning that his name carried weight in the startup world but little in pharmaceutical and academic circles, where a scientific track record matters more than a venture reputation.
Betts-LaCroix himself is a non-traditional founder: an autodidact and serial entrepreneur, once the inventor of what was described as the world's smallest computer, with a background spanning Harvard and Caltech. The company recruited research talent from the reprogramming field, including Alejandro Ocampo, the University of Lausanne researcher whose 2016 mouse rejuvenation work helped catalyse the current wave of longevity investment.
The single-benefactor structure is the defining capital feature. Where competitor Altos Labs launched in 2022 with $3 billion from a syndicate including Jeff Bezos, Retro's entire seed was one cheque from one individual. That concentration gave the company latitude to pursue long-horizon science without the quarterly pressure of a broad investor base, but it also made Altman's continued participation central to every subsequent financing conversation.
The Scientific Pipeline: Four Programs Across Three Mechanisms
Retro's pipeline is organised around the thesis that aging is driven by identifiable, addressable cellular failures, and that intervening at the cellular level can restore tissues toward a more youthful state. As of late 2025 the company described four major programs, three of which involve cellular reprogramming.
The mechanisms span the longevity field's main bets:
| Program area | Mechanism | Approach | Stage |
|---|---|---|---|
| Autophagy (RTR242) | Restore lysosomal function and cellular recycling | Oral small molecule | Phase 1, dosing |
| In vivo reprogramming | Partial reprogramming inside the body | Gene therapy delivery | Preclinical |
| Cell replacement | Reprogram cells in the lab, return them rejuvenated | "Micro replacement" of T cells, hematopoietic stem cells | Preclinical |
| Microglia therapeutics | Reprogramming-derived CNS cell therapy | Cell therapy | Preclinical |
The reprogramming work rests on the foundational discovery by Shinya Yamanaka, who showed in 2006 that four transcription factors (OCT4, SOX2, KLF4, and c-MYC) could revert mature adult cells into pluripotent stem cells. Partial reprogramming aims to apply these factors transiently, resetting a cell toward a younger epigenetic state without erasing its identity entirely.
Retro pursues two distinct delivery philosophies for this. In vivo reprogramming attempts to deliver the factors directly into tissues inside the body. Cell replacement, which Retro terms "micro replacement," instead removes cells, rejuvenates them in the lab, and returns them, an approach the company has discussed for T cells and hematopoietic stem cells. The microglia program extends reprogramming into central nervous system cell therapy.
On the manufacturing side, Retro signed a commercial and supply agreement valued at up to $85 million with Multiply Labs in May 2024, a robotics company building automated systems for cell therapy manufacturing, an early signal that at least one program was being pointed toward production scale.
The OpenAI Collaboration: GPT-4b micro and the 50x Result
The collaboration that brought Retro its widest attention is its joint work with OpenAI on a specialised protein engineering model. In August 2025, OpenAI published a closer look at GPT-4b micro, described as a miniature version of GPT-4o specialised for protein engineering, developed with Retro's Applied AI team.
The headline result is specific. The model designed novel variants of two Yamanaka factors, with the redesigned proteins named RetroSOX and RetroKLF, reported to achieve a roughly 50-fold increase in the expression of stem cell reprogramming markers compared with the standard proteins, alongside early signs of improved DNA damage repair.
What the model actually is
The mechanics are worth stating precisely, because the framing in popular coverage has tended to compress them. GPT-4b micro is a small, specialised protein language model, described as a miniature version of GPT-4o, rather than a general chatbot. It was trained on protein sequences, biological text, and tokenized 3D structural data, and researchers interact with it through few-shot prompting, supplying example sequences alongside the design task.
The technical premise is the one that underpins the entire protein-language-model field: an amino acid sequence is information-complete, storing structure and function in its ordering much as a sentence stores meaning, so a model trained to "speak" protein can in principle generate novel functional sequences. What separates GPT-4b micro from a structure predictor is the direction of the problem. AlphaFold is predictive: sequence in, 3D structure out. GPT-4b micro is generative: a desired function in, novel candidate sequences out. It belongs to the same category as ESM3 and ProGen, not the same category as AlphaFold.
The reported design results are specific. Over 30% of AI-generated RetroSOX variants showed higher reprogramming activity than native SOX2, some differing by more than 100 amino acids, and nearly 50% of RetroKLF variants exceeded manually engineered versions. In fibroblasts, the redesigned proteins produced pluripotency markers such as NANOG and TRA-1-60 days ahead of the usual timeline, and in mesenchymal stromal cells from older adult donors, over 85% of cells activated endogenous pluripotency genes within 12 days. The framing elsewhere has been that iPSC generation time fell from roughly three weeks to about seven days.
How it compares to the rest of the field
The protein-design model landscape as of May 2026 is more crowded, and in several respects more transparent, than Retro's positioning implies. Three reference points matter.
| Model | Developer | Type | Scale and disclosure | Notes |
|---|---|---|---|---|
| GPT-4b micro | OpenAI / Retro | Generative PLM | Undisclosed size; sequences and weights not released | Specialised for reprogramming-factor design |
| ESM3 | EvolutionaryScale | Generative, multimodal | Up to 98B params; published in Science, API access | Reasons jointly over sequence, structure, function |
| ProGen3 | Profluent | Generative PLM | Up to 46B params, 1.5T tokens; preprint with scaling laws | Wet-lab validated across model scale |
| AlphaFold 3 | Google DeepMind | Predictive | Open-sourced for non-commercial use | Structure prediction, not design |
The comparison cuts two ways. On capability, GPT-4b micro is not obviously ahead of the field. ESM3 reasons natively across sequence, structure, and function and was used to generate a novel fluorescent protein only 58% similar to any known one, a result published in Science. ProGen3 is a 46-billion-parameter family with published compute-optimal scaling laws and systematic wet-lab validation. Against these, GPT-4b micro is a narrowly scoped model whose distinctive feature is application focus, designing better Yamanaka factors specifically, rather than any disclosed architectural or scale advantage.
On transparency the gap is starker and runs the other way. ESM3 is published and API-accessible; ProGen3 has a public preprint; AlphaFold 3 was open-sourced for non-commercial research. GPT-4b micro's sequences, weights, and training data have not been released, and as of May 2026 the work remains a company and OpenAI write-up rather than a peer-reviewed paper, though the parties have said they intend to publish. The honest read is that GPT-4b micro's interest lies less in being a better protein model than in being a tightly integrated one: an AI capability built around a single company's specific biology, with outputs flowing straight into its wet lab.
What the science says, and where the caveats are
The underlying biology is real and active. Partial reprogramming with Yamanaka factors has a substantial peer-reviewed literature: transient OSKM or OSK expression has restored vision in mouse glaucoma models and extended lifespan in wild-type mice, and enhanced liver regeneration in animal studies. Safer, more efficient reprogramming factors are a genuinely useful goal, and an AI tool that broadens the searchable design space is a reasonable way to pursue it.
But the field's hard problems are exactly the ones an in vitro marker-expression result does not touch, and they are well documented:
- Generalisation across cell types and species is the central unsolved difficulty. Harvard aging researcher and Retro advisor Vadim Gladyshev has put it directly: skin cells are easy to reprogram, other cells are not, and moving to a new species often yields nothing. A large improvement in fibroblasts and aged-donor stromal cells is encouraging but does not establish that the gain survives the jump to therapeutically relevant tissues in vivo.
- In vivo reprogramming carries documented toxicity. Chronic OSKM expression causes teratoma formation and organ toxicity because different cell types reprogram at different rates. A January 2026 study from Gladyshev's own group found in vivo chemical reprogramming was associated with toxic lipid-droplet accumulation, with higher doses causing rapid weight loss requiring euthanasia in mice. Whether AI-improved factors ease this safety problem or simply make a hazardous process more potent is unknown.
- The result is in vitro and unpublished. The 50x figure is marker expression in a dish, not an animal rejuvenation or disease outcome, and the comparison to baseline is sensitive to how "efficiency" is defined, a quantity measured inconsistently across the literature. Without released sequences, independent benchmarking is not yet possible.
- The model is a black box. Researchers have acknowledged that how GPT-4b micro arrives at its designs is not well understood, which matters more for a therapeutic protein than for a generated image.
- There is an unavoidable conflict-of-interest optic. Altman is OpenAI's CEO and Retro's sole seed funder. OpenAI has stated no money changed hands and that Altman was not directly involved in the project, but the dual role means the most visible validation of an OpenAI bio-capability also lifts the fundraising profile of an Altman-backed company. That does not make the science wrong; it does mean external validation carries more weight than usual.
So, is it more interesting than it looks, or less?
Both, depending on which claim is being assessed. As a frontier protein-design model, GPT-4b micro is less interesting than the headlines suggest: it is smaller, more narrowly scoped, and far less transparent than ESM3 or ProGen3, and it has not demonstrated a capability the broader field lacks. As an integrated discovery loop, it is more interesting than a model card alone would convey: a specialised model whose outputs feed directly into one company's reprogramming programs and wet-lab validation is a concrete instance of the model-propose, lab-test, model-refine cycle that bio-AI has been promising in the abstract.
The fair verdict for an expert reader is that the collaboration is a credible and genuinely useful tool applied to a hard, legitimate problem, wrapped in a claim marketed well ahead of its evidentiary support. The narrative is doing more work than the data, which can be true and still leave the work worth watching. The decisive tests are not in the dish: they are whether AI-designed factors improve reprogramming safely in vivo, in tissues that matter, and whether the result holds up once the sequences are published and others can check it.
The collaboration matters strategically regardless of how the specific result resolves. It is the clearest worked example to date of an Altman-affiliated AI capability applied directly to an Altman-funded biology company, and it is the narrative engine behind Retro's positioning as an AI-powered longevity company rather than a conventional one.
RTR242: The First Clinical Trial
In late December 2025, Retro dosed the first participant in the first human trial of its history, crossing what Betts-LaCroix called a binary transition from a discovery-stage to a clinical-stage company.
The candidate is RTR242, a small-molecule oral therapy. Its mechanism is distinct from the reprogramming programs: rather than resetting cell identity, it is designed to restore lysosomal function, a core component of autophagy, the cell's waste-handling and recycling system. In younger cells, lysosomes maintain an acidic environment that lets autophagy break down damaged proteins and debris. With age, and particularly in neurodegenerative disease, lysosomes lose acidity and efficiency, and toxic protein aggregates accumulate. RTR242 is intended to reactivate that cleanup machinery.
The opening indication is Alzheimer's disease, chosen as a wedge into the broader aging thesis: a defined disease with measurable endpoints serving as the proving ground for an autophagy mechanism with putative whole-body relevance.
| RTR242 trial | Detail |
|---|---|
| Phase | Phase 1, first-in-human |
| Design | Randomised, double-blind, placebo-controlled |
| Population | Healthy volunteers |
| Site | Early-phase clinical unit, Adelaide, Australia |
| Endpoints | Safety, tolerability, plus exploratory autophagy and lysosomal biomarkers |
| First dosing | December 2025 |
| Data | First readout anticipated around August 2026 |
The choice of Australia is pragmatic and common among early-stage biotechs: first-in-human approvals are faster there, letting a company move from lab to human safety data without prolonged regulatory delay.
At STAT's Breakthrough Summit West in May 2026, Betts-LaCroix said the trial was going well and that researchers had not seen dose-limiting toxicities, with some data expected around August. As a Phase 1 in healthy volunteers, the trial is designed to read out on safety and tolerability rather than efficacy; cognitive or disease-modifying signals are not its endpoint, and any biomarker observations are exploratory.
The May 2026 Financing: $1.8 Billion, and the Gap to $5 Billion
On May 22, 2026, Retro announced new financing at a $1.8 billion valuation. Around the same window, Altman was reported to have personally invested close to a further $180 million into the company, echoing the size of his original seed.
The valuation number is most informative when read against the company's own prior ambition. In December 2025, STAT reported that Retro was chasing a $5 billion valuation on a roughly $1 billion Series A, having circulated fundraising decks, one of which reportedly projected the company's eventual market value approaching that of large-cap technology firms. The round had first surfaced in January 2025 as a $1 billion Series A that Altman was joining, with the company said to be in discussions with venture investors, family offices, and sovereign wealth funds.
The progression is worth laying out plainly:
| Date | Reported event | Valuation reference |
|---|---|---|
| 2021 to 2022 | Seed round, funded entirely by Altman | $180M raised |
| Jan 2025 | $1B Series A reported, Altman joining | Target round forming |
| Aug 2025 | GPT-4b micro 50x result published with OpenAI | (narrative catalyst) |
| Dec 2025 | Reported to be chasing $5B valuation | $5B target |
| Dec 2025 | First RTR242 patient dosed | (clinical milestone) |
| May 2026 | Financing announced | $1.8B realised |
A $1.8 billion mark is a large valuation for a company with no approved products, no revenue, and no human efficacy data. It is also materially below the $5 billion the company was reportedly targeting five months earlier. The gap can be read several ways, and a neutral account should hold them together rather than pick one.
One reading is a broad repricing of speculative, pre-clinical longevity assets, in line with the wider biotech funding environment. Another is investor discipline specific to the category: the longevity field's most heavily valued companies have repeatedly been marked on narrative and backer prestige rather than clinical validation, and the distance between a $5 billion ask and a $1.8 billion close is the kind of correction that follows when that gap is tested against an actual book of committed capital. A third reading is simply that the company chose to take available capital at a defensible mark rather than hold out for a number the market would not support. The reported continuation of Altman's personal participation is consistent with any of these.
What the financing does establish is runway. Combined with the clinical entry of RTR242, it funds the company through the period in which its first human data, and the durability of the OpenAI-driven reprogramming story, will be tested.
The Competitive Longevity Landscape, May 2026
Retro operates in a small field of well-capitalised companies pursuing cellular rejuvenation, where competition for a limited pool of reprogramming scientists has been as intense as competition for capital.
| Company | Founded | Lead backer(s) | Approximate valuation | Approach |
|---|---|---|---|---|
| Altos Labs | 2022 | Jeff Bezos and others | ~$6B+ | Cellular reprogramming, rejuvenation |
| Retro Biosciences | 2021 | Sam Altman | ~$1.8B | Reprogramming, autophagy, cell replacement |
| NewLimit | 2021 | Brian Armstrong | private | Epigenetic reprogramming |
| Unity Biotechnology | 2011 | Bezos, Thiel (early) | public, small-cap | Senolytics |
Altos remains the field's heavyweight by capital, having launched with $3 billion and recruited a large share of academic reprogramming leaders. NewLimit, founded by Coinbase's Brian Armstrong, pursues a similar epigenetic reprogramming thesis. Across the group, the common position as of May 2026 is the same one that frames Retro: large valuations, deep scientific ambition, and an absence of human efficacy data.
Retro's claimed differentiators within this set are its multi-program breadth, its early clinical entry with RTR242, and the AI tooling from the OpenAI relationship. Each is genuine and each is contestable. Breadth spreads risk but also spreads a fixed research budget across more fronts. Early clinical entry is real but sits at the safety-only end of the development path. The AI advantage is the most distinctive and the least independently validated.
Red Team vs Blue Team Analysis
Risk Analysis (Red Team)
No human efficacy data anywhere in the portfolio. As of May 2026 the company's clinical footprint is a single Phase 1 safety trial in healthy volunteers. Every therapeutic claim, including the reprogramming programs that justify most of the scientific narrative, rests on preclinical and in vitro results. The $1.8 billion valuation prices a thesis, not an outcome.
Single-benefactor concentration. Altman's funding has been central from the seed through the latest round. That concentration has given Retro freedom from quarterly pressure, but it also means the company's financing has not, to date, been broadly stress-tested by an arms-length syndicate, and its fortunes are unusually tied to one individual's continued conviction and capacity.
The valuation trajectory is a caution signal. Closing at $1.8 billion after reportedly targeting $5 billion is a meaningful miss on the company's own stated ambition. Whether driven by market conditions or investor scepticism, it indicates that the prestige-and-narrative valuation model that has carried the longevity field is being repriced.
The 50x result is unverified externally and several steps from a drug. The GPT-4b micro reprogramming result is an in vitro marker-expression multiple, with sequences and training data not publicly released. It is a laboratory throughput improvement, not a demonstrated therapeutic, and its translation into the in vivo and cell-replacement programs is unproven.
Multi-program breadth dilutes focus. Pursuing autophagy, in vivo reprogramming, cell replacement, and microglia therapeutics simultaneously spreads a finite budget and management attention. The strategy is deliberate, but it is the opposite of the single-asset focus that typically de-risks an early biotech for later investors.
Reprogramming carries intrinsic oncogenic risk. Partial reprogramming sits close to a known hazard: full reprogramming and the c-MYC factor are associated with cancer risk. Any in vivo program must demonstrate it can rejuvenate cells without tipping them toward malignancy, a safety bar that has constrained the entire field.
Opportunities and Mitigants (Blue Team)
Clinical-stage status is a genuine inflection. Dosing the first RTR242 patient moved Retro from discovery to the clinic, a transition many longevity companies have not reached. A clean safety readout around August 2026 would be the first human data point in the portfolio and would de-risk the autophagy program specifically.
The autophagy mechanism is well-grounded and indication-specific. RTR242 targets lysosomal function in Alzheimer's, a mechanism supported by a substantial body of literature on autophagy in neurodegeneration. Unlike the reprogramming programs, it has a defined disease, defined endpoints, and a conventional small-molecule development path.
The AI tooling is a real differentiator if it generalises. If GPT-4b micro-style sequence design extends beyond Yamanaka factors to other therapeutic proteins, Retro would hold a structural advantage in protein engineering throughput that conventional longevity competitors lack. The OpenAI relationship is difficult for others to replicate.
Capital concentration enabled long-horizon science. The single-benefactor structure freed Retro from the milestone-by-milestone financing treadmill, letting it pursue multiple aging hypotheses and build manufacturing capacity ahead of clinical need, optionality a syndicate-funded company might not have tolerated.
A defensible mark with fresh capital and runway. Even at $1.8 billion rather than $5 billion, the company has secured a large valuation and new funding, reportedly including continued Altman participation, sufficient to carry it through the period in which its first human data arrives.
Manufacturing groundwork is already laid. The Multiply Labs agreement positions at least one cell-therapy program toward automated production, reducing a downstream scale-up risk that often surprises early cell-therapy developers.
| Risk Category | Key Concern |
|---|---|
| No human efficacy data | Entire portfolio priced on preclinical and in vitro results |
| Single-benefactor concentration | Financing centred on one individual, not arms-length tested |
| Valuation miss | $1.8B close versus $5B target signals category repricing |
| Unverified AI result | 50x figure is in vitro, sequences and data not released |
| Multi-program dilution | Finite budget spread across four programs |
| Oncogenic risk | Reprogramming intrinsically near cancer-related pathways |
| Opportunity | Observation |
|---|---|
| Clinical-stage entry | First human data point expected around August 2026 |
| Grounded autophagy mechanism | RTR242 has a defined disease and conventional path |
| AI differentiation | OpenAI tooling hard for competitors to replicate |
| Long-horizon capital | Concentration enabled multi-program, manufacturing-ahead strategy |
| Fresh runway | New funding carries the company to first data |
| Manufacturing readiness | Multiply Labs deal de-risks cell-therapy scale-up |
Conclusion
Retro Biosciences is the clearest single example of the longevity moonshot fused with the AI-for-biology thesis: founded in 2021 on a $180 million cheque from one person, structured to chase several aging hypotheses at once, and narratively anchored by a protein engineering result produced with OpenAI.
As of May 2026 the company has reached three genuine milestones. It has entered the clinic with RTR242, an autophagy-targeting Alzheimer's candidate now in a Phase 1 safety trial in Australia with first data expected around August. It has published a reprogramming result with OpenAI that, whatever its eventual independent standing, is the field's most prominent demonstration of LLM-driven protein design applied to a funded biology program. And it has closed new financing at a $1.8 billion valuation, with reported continued backing from Altman.
The same period also delivered the clearest caution. A $1.8 billion close, against a reported $5 billion target only five months earlier, is the kind of correction that arrives when narrative valuations meet a real book of committed capital. For a company with no approved product, no revenue, and no human efficacy data, the gap is the most concrete market signal available about how the longevity category is now being priced.
The near-term test is straightforward. The RTR242 safety readout is the first hard data point the company will produce, and the durability of the reprogramming story depends on whether the GPT-4b micro work translates beyond in vitro marker expression. Both will be clearer within the next twelve to eighteen months. Until then, Retro remains what it has been since 2022: a well-capitalised, scientifically ambitious bet on the proposition that aging is addressable, carrying a valuation that prices the ambition rather than the evidence.
All information in this article was accurate as of the time of the research and is derived from publicly available sources including company statements, OpenAI's published collaboration write-up, financial news reporting, and interviews with company leadership. Specific terms of Retro's May 2026 financing (round size, investor syndicate, and the precise structure of Sam Altman's reported additional investment) are not fully publicly disclosed. The reported 50x reprogramming result has not been independently verified, as the underlying protein sequences and training data have not been released. Information may have changed since publication. This content is for informational purposes only and does not constitute investment, legal, or financial advice. The author is not a lawyer or financial adviser.