🛠️ Method & tooling: FM-os — the SLM/foundation-model-ops hub and the closed-loop machinery this project runs on. longevity-loop is FM-os applied to a real mission.
A solo builder's public, self-improving loop for AI × longevity: pick a falsifiable question → analyze open aging data → score on a public verifier → write it up honestly → share → let the artifact recruit people, feedback, and funding → repeat, harder. Built in public, verified not vibed.
North star: climb a credibility-gated, code-only leaderboard (the Biomarkers of Aging Challenge on the open Biolearn platform) — real signal in aging science with no wet lab.
Not medical advice. Computational results on public data are labeled as such and kept strictly separate from any wet-lab/therapeutic claim (which requires independent validation). No evidence ⇒ no claim.
Each turn is falsifiable and ends in a shared, verifiable artifact. No evidence ⇒ no claim.
| # | Stage | What happens |
|---|---|---|
| 1 | QUESTION | State one falsifiable question + the metric that settles it (and the null you'd accept). |
| 2 | DATA | Pull an OPEN aging dataset (CELLxGENE, Tabula Muris Senis, GEO, GTEx, the Biomarkers challenge set). |
| 3 | MODEL | Analyze / fine-tune an open bio-FM (pyaging, Geneformer, scGPT) — cheaply, reproducibly. |
| 4 | VERIFY | Score on a public verifier — the Biolearn leaderboard or a held-out benchmark. No evidence ⇒ no claim. |
| 5 | WRITE-UP | Honest report: result AND the null/failure, threats-to-validity, a reproduce command. |
| 6 | SHARE | Build in public — repo + thread + explainer; route to the human hubs (VitaDAO, LBF, Foresight). |
| 7 | COMPOUND | The artifact recruits feedback, collaborators, and funding → they unlock the next, harder question. |
⑦ COMPOUND feeds back into ① — each turn adds data, a tool, or a connection, so the next question is bigger.
Every turn of the loop, logged honestly. done requires a PROOF (result incl. the null + a reproduce command).
| Turn | Question | Stage | Status |
|---|---|---|---|
| turn-01-biolearn-baseline | Motivated by Levine's Systems Age (2025): does cross-clock DISAGREEMENT (heterogeneity across a clock panel) add predictive signal over the best single clock on the Biomarkers-of-Aging open data? | VERIFY | 🧩 scaffolded |
| turn-02-biofm-finetune | Does fine-tuning an open single-cell FM beat a frozen-embedding linear probe at predicting age? | MODEL | 🧩 scaffolded |
The frontier groundbreakers' most recent deep works — verified, with a link + a real quote. Refreshed weekly by scripts/track.py (arXiv + GitHub).
- Alex Zhavoronkov — The End of Aging Clocks: Training Foundation Models to Reason in Aging and Longevity (2026)
Longevity-LLM v0.1 (fine-tuned Qwen3-14B) hit 4.34-yr MAE epigenetic-age prediction (beating Horvath) across methylation/proteomics/clinical/RNA, and handled multiple longevity tasks.
"These results demonstrate that a single modestly sized LLM can match or replace purpose-built aging clocks across data modalities." → Future: Interim report from Insilico's Multi-Modal AI Gym for Science (MMAI) — foundation models for drug discovery + aging.
- Morgan Levine — Systems Age: one blood methylation test quantifying aging across 11 physiological systems (2025)
DNA-methylation clocks that score aging separately per system (heart, lung, brain, immune…) from one blood draw, beating global clocks at system-relevant disease prediction.
"most epigenetic clocks provide a single age estimate, overlooking within-person variation." → Future: System-specific clocks usable clinically to track how interventions shift aging in individual organ systems.
- Tony Wyss-Coray — Plasma proteomic signatures of cellular aging predict human disease (2026)
From >7,000 plasma proteins in 60,542 people, ML models estimate biological age of 40+ cell types, linking cell-type-specific aging to disease and mortality.
"Aging is asynchronous across cells and organs." → Future: Cell-type-resolved plasma proteomic clocks as clinical biomarkers from a single blood test.
- Vadim Gladyshev — Mammalian aging involves genome-wide splicing degeneration leading to functional decline (2026)
Integrative mouse/human analysis shows aging systematically loses RNA-splicing fidelity ('splicing degeneration'), rising with age but alleviated by calorie restriction or rapamycin — a proposed new hallmark.
"aging is characterized by systematic deterioration of the fidelity of RNA splicing, here termed splicing degeneration" → Future: Splicing degeneration as 'a promising target for aging interventions acting to reverse' it.
- João Pedro de Magalhães — Translational toolkit for reproducible, cross-study profiling of human ageing hallmarks (2026)
A validated assay toolkit to simultaneously quantify 8+ ageing hallmarks (senescence, immune ageing, mTOR, autophagy, genomic instability…) in clinically accessible human blood and tissue.
"a validated, high-resolution toolkit for the simultaneous quantification of multiple ageing hallmarks in clinically accessible human samples" → Future: Standardize hallmark measurement to overcome methodological heterogeneity and translate into human clinical studies.
- Peter Fedichev — A Minimal Model Explains Aging Regimes and Guides Intervention Strategies (2025)
Reduces aging physiology to three variables (resilience, entropic damage, regulatory noise), yielding two regimes: linear damage-driven aging in stable species like humans vs intrinsic instability in mice/flies.
"In stable species, including humans, aging is driven by linear damage accumulation that gradually erodes resilience" → Future: A three-level intervention roadmap: target dynamic hallmarks, reduce physiological noise, slow/reverse entropic damage.
- Jacob Kimmel — In silico design of epigenetic reprogramming payloads (2025)
NewLimit's generative model (protein-foundation-model transfer learning) designs transcription-factor reprogramming payloads from sparse sampling of the combinatorial TF space, in a lab-in-the-loop.
"Through diverse epigenetic codes, human cells execute distinct programs from a common genome" → Future: Run the model lab-in-the-loop to design reprogramming interventions far faster than pure experiments.
- George Church — Replacement as an aging intervention (Nature Aging Perspective) (2025)
Argues replacing aged cells/tissues/organs is an underappreciated, near-term-feasible strategy where drug interventions have not yet proven durable in humans.
"there is a lack of interventions conclusively shown to attenuate the processes of aging in humans" → Future: Develop replacement-based interventions (cell/tissue/organ) alongside reprogramming and gene therapy.
- Steve Horvath — Epigenetic ageing clocks: statistical methods and emerging computational challenges (2025)
Nature Reviews Genetics review (Teschendorff & Horvath) on the statistical foundations of epigenetic clocks and open problems in interpretation, cell-type heterogeneity, and single-cell methods.
"many computational and statistical challenges remain that limit our understanding, interpretation and application of epigenetic clocks" → Future: Interpretable clocks built at cell-type and single-cell resolution to make epigenetic age causally + clinically meaningful.
- Matt Kaeberlein — Exercise and Weekly Sirolimus (Rapamycin) in Older Adults: RAPA-EX-01 RCT (2026)
RCT (40 adults, 65-85) found once-weekly 6 mg rapamycin did NOT boost — and may have slightly blunted — functional gains from a 13-week exercise program, with more adverse events. A clean, useful negative result.
"did not enhance, and in sensitivity analyses, it may have modestly attenuated short-term functional improvements from a home exercise programme" → Future: Test alternative rapamycin dosing/timing (e.g. mTORC1 cycling) before combining with exercise in older adults.
🕸️ Congregational view: the field as a spatiotemporal knowledge graph (modeled after getzep/graphiti) — open the field graph →.
Reflections — what else could be important? (synthesis, not claims)
- Clocks are collapsing into foundation models: Zhavoronkov's Longevity-LLM replacing purpose-built clocks + Kimmel's protein-FM reprogramming design → the field's own 'FM-ops' moment. A unified multimodal aging FM is the obvious open target.
- Aging is resolving from one number to many: Systems Age (11 systems), Wyss-Coray (40+ cell types), Gladyshev (splicing as a new hallmark). The gap: a shared, cell/system-resolved BENCHMARK so these aren't incomparable — a natural longevity-loop contribution.
- Negative results are becoming first-class (Kaeberlein's rapamycin+exercise null). An open registry of honest longevity nulls would be high-trust signal and is exactly the no-evidence⇒No discipline the field needs.
- Correlation ≠ cause is the recurring caveat (Horvath): the frontier wants INTERPRETABLE, causal clocks. A loop turn probing whether an intervention moves a clock in a held-out, pre-registered way is more valuable than a new clock.
- The data wall is inverting in biology (multiomic tokens > internet-text tokens): the scarce input is now well-curated, standardized human hallmark data (de Magalhães' toolkit) — curation, not compute, is the bottleneck to attack.
- Two camps to bridge: 'repair/replace damage' (Church, Fedichev's entropic damage) vs 'reprogram/rejuvenate' (Kimmel, Levine). A model that predicts which regime a given tissue is in could route interventions — an unclaimed synthesis.
Three tracks every week — full weekly plan in docs/ROADMAP.md:
- 🧠 Knowledge — ramp on aging biology fast; verify every claim against a primary paper (TRUE = evidenced).
- 🛠️ Tooling — each week ships runnable, gated code (the loop); cheap fine-tunes on Tinker/Modal.
- 🤝 Connections — build in public; reach the hubs + people with an artifact in hand, never empty-handed.
Stand up the loop, ramp on aging biology, ship the first VERIFIABLE result.
GATE 1 — public repo live · ≥1 leaderboard submission · first grant application in · 5 researcher touches.
A genuine finding on OPEN data + an adopted open tool + a real collaborator.
GATE 2 — a reproducible finding write-up · an open tool with ≥1 external user · 1 named collaborator · a micro-grant funded OR strong grant progress.
Convert signal into a preprint, non-dilutive funding, and a deliberate fork in the road.
GATE 3 — a preprint OR top-decile leaderboard · non-dilutive funding · a named collaborator · a clear next-90 plan.
Signal ladder (each rung recruits the next):
- Public repo + honest launch thread
- First open-leaderboard submission (code-only, credibility-gated)
- A reproduced aging clock + a shipped open tool
- A fine-tuned bio-FM finding on open data (research-loop write-up)
- A named academic/industry collaborator
- Non-dilutive micro-grant (VitaDAO / Foresight)
- A preprint or top-decile leaderboard finish
- A talk/podcast + a company-or-lab decision
🤖 = AI-forward · 💬 = active in the open community (good first contacts).
- João Pedro de Magalhães 🤖💬 — University of Birmingham: HAGR aging databases; computational biogerontology
- Alex Zhavoronkov 🤖💬 — Insilico Medicine: deep-learning aging clocks + generative AI drug discovery
- Peter Fedichev 🤖💬 — Gero: physics/AI dynamical models of aging
- Jacob Kimmel 🤖💬 — NewLimit (fmr Calico): ML-designed reprogramming payloads; bio foundation models
- Morgan Levine 🤖💬 — Altos Labs (fmr Yale): epigenetic aging clocks (PhenoAge)
- Tony Wyss-Coray 🤖💬 — Stanford: plasma-proteomic organ aging clocks
- Vadim Gladyshev 🤖 — Harvard Medical School: mouse + single-cell (scAge) aging clocks
- George Church 🤖💬 — Harvard / Wyss Institute: gene-therapy longevity; synthetic biology
- Nir Barzilai 💬 — Albert Einstein College of Medicine: centenarian genetics; TAME metformin trial
- Matt Kaeberlein 💬 — Optispan (fmr U. Washington): rapamycin geroscience; Dog Aging Project
- David Sinclair 💬 — Harvard Medical School: information theory of aging; reprogramming; sirtuins
- Eric Verdin 💬 — Buck Institute: geroscience; ketone bodies; immune aging
- Andrew Steele 💬 — Independent (author of 'Ageless'): longevity science communication
- Charles Brenner 💬 — City of Hope: NAD+ metabolism; vocal longevity-hype skeptic (a good reality check)
- Joe Betts-LaCroix 💬 — Retro Biosciences: reprogramming + autophagy longevity company
- Kristen Fortney 🤖 — BioAge Labs: ML on longitudinal human data for aging drug discovery
🤖 = AI-native platform.
- NewLimit 🤖 — AI-guided epigenetic reprogramming to restore youthful cell function (well-funded-private)
- Retro Biosciences 🤖 — reverse cellular aging (reprogramming, autophagy) + AI protein design (well-funded-private)
- Gero 🤖 — physics-based AI modeling of aging + generative molecule design (well-funded-private)
- Shift Bioscience 🤖 — AI 'virtual cell' for safe single-gene cellular reprogramming (early)
- BioAge Labs 🤖 — aging-biology multi-omics ML platform (metabolic disease) (public)
- Insilico Medicine 🤖 — generative AI drug discovery (PandaOmics, Chemistry42) + clinical pipeline (public)
- Altos Labs — cellular rejuvenation via partial reprogramming (+ growing computation arm) (well-funded-private)
- Calico Life Sciences — basic biology of aging → age-related-disease medicines (well-funded-private)
- Recursion 🤖 — AI drug discovery via cellular imaging (Recursion OS; merged Exscientia) (public)
- Isomorphic Labs 🤖 — AI drug design built on AlphaFold (well-funded-private)
- Cellarity 🤖 — AI on single-cell transcriptomics to design cell-state-correcting medicines (well-funded-private)
- Xaira Therapeutics 🤖 — AI-native drug discovery on David Baker's generative protein models (well-funded-private)
- Rubedo Life Sciences 🤖 — AI-driven senolytics discovery (ALEMBIC platform) (well-funded-private)
- Gordian Biotechnology — high-throughput in-vivo pooled screening (Mosaic) for diseases of aging (early)
- Loyal — canine longevity — lifespan-extension drugs for dogs (well-funded-private)
- Centenara Labs — hallmarks-of-aging therapeutics portfolio (fmr Rejuveron) (well-funded-private)
- Geneformer — Transformer on ~30-104M single-cell transcriptomes; fine-tune for cell-type / perturbation / aging tasks.
- scGPT — GPT-style single-cell foundation model; fine-tune for annotation, integration, perturbation prediction.
- ESM-2 — Protein LM (8M-650M) that runs on a laptop; embeddings for variant-effect on longevity genes.
- TxGNN — GNN over a drug/disease knowledge graph for zero-shot repurposing; rank candidate longevity interventions.
- AltumAge — Deep pan-tissue methylation clock (MLP over ~20k CpGs) to fine-tune/benchmark vs ElasticNet clocks.
- AlphaFold — Structure prediction; usually query the precomputed AlphaFold DB rather than run locally.
- scanpy — Standard single-cell preprocessing (QC/clustering/DE) before any aging model.
- CZ CELLxGENE Census — API to slice ~33M+ standardized cells by tissue/age/disease in seconds — fastest cohort pull.
- pyaging — PyTorch package bundling 50+ aging clocks with one API — score biological age on a laptop.
- Biolearn — Open standardized platform for the Biomarkers of Aging Challenge — the code-only leaderboard to compete on.
- scAge — Epigenetic age from sparse single-cell methylation — detect cell-level aging + rejuvenation.
- DunedinPACE — Pace-of-aging (rate, not age) from 450k/EPIC — a strong intervention outcome variable.
- Tabula Muris Senis — Mouse aging cell atlas (~500k cells, 18 organs) — the go-to open single-cell aging benchmark.
- GTEx Portal — Human multi-tissue expression with donor age; expression matrices download freely (genotypes gated).
- NCBI GEO — Largest public expression/epigenomics archive; search 'age'-flagged series to build custom aging sets.
- Human Cell Atlas — Open multi-omic reference maps (70M+ cells) for age-stratified tissue baselines.
- HAGR (GenAge / CellAge) — Curated aging/senescence gene sets (GenAge 307, CellAge 866) — priors / feature filters.
- UK Biobank — 500k multi-omic + health cohort — ACCESS-GATED (application/fee/cloud-only), not laptop-downloadable.
- Biomarkers of Aging Challenge — Open competition + curated dataset (methylation/proteomics/outcomes, 500+ people) — the North-Star code-only leaderboard.
- Open Problems in Single-Cell — Community benchmarking harness (tasks/datasets/metrics) to fairly evaluate a model vs baselines.
- LAB-Bench — 2,457-question benchmark for LLMs on biology research tasks (literature, sequences, DBs).
- BixBench — Benchmark for LLM agents on real computational-biology analysis workflows.
✅ = realistically open to a solo/independent builder.
- Biomarkers of Aging Challenge / Longevity Prize ✅ grant — Open, code-only competition + curated dataset (Biolearn) — the IDEAL first credibility-gated signal, no wet lab.
- VitaDAO / VitaLabs ✅ community — DeSci collective; fast fellowship grants (~$65K), light application, active Discord — most accessible funding+community on-ramp.
- Foresight Institute — Longevity Grants ✅ grant — Monthly-deadline frontier grants (AI-for-science + longevity), unusually open to non-traditional applicants.
- Foresight Fellowship ✅ fellowship — Year-long fellowship; mentorship + intros to funders/senior scientists; global, independent-friendly.
- Longevity Biotech Fellowship (LBF) ✅ fellowship — The main 'how do I get into longevity biotech' front door (ODLB merged in); cohort + retreat + community.
- age1 ✅ accelerator — Laura Deming's longevity accelerator (~$500K, 4-mo). Dilutive — the premier founder on-ramp; apply with a public track record.
- Impetus Longevity Grants (Norn Group) 🔒 grant — Fast $10K-$500K aging-science grants (~3-4 wk decisions) — PI/lab-gated; partner into a lab to access.
- Hevolution Foundation 🔒 grant — Large geroscience funder ($300-500K/yr); institutional/PI-gated — a funder to partner toward.
- XPRIZE Healthspan 🔒 grant — $101M team competition (restore function 10-20 yrs). Team/clinical — highest-leverage as a rallying point + network.
- NIA (NIH) 🔒 grant — Largest US non-dilutive aging funder; SBIR/STTR is the realistic founder path once incorporated.
- Astera Institute (Rejuvenome) 🔒 grant — Runs the open ~$70M combinatorial mouse-lifespan dataset — build on it as open data; grants are relationship-driven.
- Vitalist Bay / Vitalism ✅ community — Longevity pop-up city — the highest-density in-person gathering of the frontier/independent crowd.
- Aging Research & Drug Discovery (ARDD) ✅ conference — The top translational-geroscience conference (Copenhagen) — where the serious science+investor crowd is.
- Longevity Summit Dublin ✅ conference — Rejuvenation-biotech heavy (de Grey/O'Dea); researchers + advocates + investors.
- Longevity Marketcap (Nathan Cheng) ✅ media — Most-read industry newsletter; Cheng is a hub node (LBF, Vitalism, Healthspan Capital) — engage to map who's who.
- Lifespan.io ✅ media — Advocacy + news non-profit — a place to get build-in-public work amplified to an engaged audience.
- Longevity.Technology ✅ media — Daily industry/investment newsletter — track funding rounds + deal-flow signals.
This repo IS the artifact: every turn commits data, a result, or a connection. Follow the commits, open an issue with a paper/dataset/collaborator, or PR an entry to data/*.yml.
Generated from data/*.yml by scripts/build.py — do not edit by hand. A sibling of FM-os.