π οΈ Method & tooling: FM-os β 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 Γ longevity: pick a falsifiable question β analyze open aging data β score on a public verifier β write it up honestly β share β let the artifact recruit people, feedback, and funding β 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) β 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 β no claim.
Each turn is falsifiable and ends in a shared, verifiable artifact. No evidence β 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) β cheaply, reproducibly. |
| 4 | VERIFY | Score on a public verifier β the Biolearn leaderboard or a held-out benchmark. No evidence β no claim. |
| 5 | WRITE-UP | Honest report: result AND the null/failure, threats-to-validity, a reproduce command. |
| 6 | SHARE | Build in public β repo + thread + explainer; route to the human hubs (VitaDAO, LBF, Foresight). |
| 7 | COMPOUND | The artifact recruits feedback, collaborators, and funding β they unlock the next, harder question. |
β¦ COMPOUND feeds back into β β 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 | 𧩠scaffolded |
| turn-02-biofm-finetune | Does fine-tuning an open single-cell FM beat a frozen-embedding linear probe at predicting age? | MODEL | 𧩠scaffolded |
The frontier groundbreakers' most recent deep works β verified, with a link + a real quote. Refreshed weekly by scripts/track.py (arXiv + GitHub).
- Alex Zhavoronkov β 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." β Future: Interim report from Insilico's Multi-Modal AI Gym for Science (MMAI) β foundation models for drug discovery + aging.
- Morgan Levine β 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, immuneβ¦) from one blood draw, beating global clocks at system-relevant disease prediction.
"most epigenetic clocks provide a single age estimate, overlooking within-person variation." β Future: System-specific clocks usable clinically to track how interventions shift aging in individual organ systems.
- Tony Wyss-Coray β 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." β Future: Cell-type-resolved plasma proteomic clocks as clinical biomarkers from a single blood test.
- Vadim Gladyshev β 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 β a proposed new hallmark.
"aging is characterized by systematic deterioration of the fidelity of RNA splicing, here termed splicing degeneration" β Future: Splicing degeneration as 'a promising target for aging interventions acting to reverse' it.
- JoΓ£o Pedro de MagalhΓ£es β 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 instabilityβ¦) 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" β Future: Standardize hallmark measurement to overcome methodological heterogeneity and translate into human clinical studies.
- Peter Fedichev β 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" β Future: A three-level intervention roadmap: target dynamic hallmarks, reduce physiological noise, slow/reverse entropic damage.
- Jacob Kimmel β 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" β Future: Run the model lab-in-the-loop to design reprogramming interventions far faster than pure experiments.
- George Church β 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" β Future: Develop replacement-based interventions (cell/tissue/organ) alongside reprogramming and gene therapy.
- Steve Horvath β 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" β Future: Interpretable clocks built at cell-type and single-cell resolution to make epigenetic age causally + clinically meaningful.
- Matt Kaeberlein β 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 β and may have slightly blunted β 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" β Future: Test alternative rapamycin dosing/timing (e.g. mTORC1 cycling) before combining with exercise in older adults.
- Marinka Zitnik β ATHENA-R1: An AI agent for treatment reasoning over a biomedical tool universe (2026)
RL-trained agent (fine-tuned Qwen3-8B) that reasons over 200+ biomedical tools via ToolUniverse; reported to beat GPT-5 on drug- and treatment-prediction benchmarks. A research artifact, explicitly not a medical device.
"ATHENA-R1 is an AI agent for treatment reasoning, trained through reinforcement learning over a universe of biomedical tools." β Future: A shared, open biomedical tool library (ToolUniverse) as the reusable substrate other agents β including a longevity loop β can reason over.
πΈοΈ Congregational view: the field as a spatiotemporal knowledge graph (modeled after getzep/graphiti) β open the field graph β.
Reflections β 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 β 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 β 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-evidenceβNo discipline the field needs.
- Correlation β 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 MagalhΓ£es' toolkit) β 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 β an unclaimed synthesis.
Three tracks every week β full weekly plan in docs/ROADMAP.md:
- π§ Knowledge β ramp on aging biology fast; verify every claim against a primary paper (TRUE = evidenced).
- π οΈ Tooling β each week ships runnable, gated code (the loop); cheap fine-tunes on Tinker/Modal.
- π€ Connections β build in public; reach the hubs + people with an artifact in hand, never empty-handed. In Act II this is the bridge: the credential network (Genentech, Yale/Levine, Georgia Tech) turns computation into bench access, then a raise.
Stand up the loop, ramp on aging biology, ship the first VERIFIABLE result.
GATE 1 β public repo live Β· β₯1 leaderboard submission Β· first grant application in Β· 5 researcher touches.
A genuine finding on OPEN data + an adopted open tool + a real collaborator.
GATE 2 β a reproducible finding write-up Β· an open tool with β₯1 external user Β· 1 named collaborator Β· 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 β a preprint OR top-decile leaderboard Β· non-dilutive funding Β· a named collaborator Β· a lab partnership lined up for Act II.
Sequenced bridge: convert the computational finding + your credential network (Genentech, Yale/Levine orbit, Georgia Tech) into a lab-validated result, then an incorporated, fundable company. Full-time, raise-ready.
GATE 4 β a lab-validated (or cleanly null) finding Β· a co-authored preprint Β· an incorporated entity + a wet-lab co-founder Β· age1 acceptance OR a pre-seed raise underway.
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 lab partnership (dry+wet): your computation earns bench access + co-authorship
- A wet-lab-validated (or cleanly null) finding with the partner lab
- An incorporated company + a wet-lab co-founder from the network
- age1 acceptance or a pre-seed raise β the sequenced fork taken: build the company
16/27 done (59%). Each execution is a checkbox with a beforeβafter eval; a box only ticks with a real result (done requires a non-pending before AND after β no evidence β not done, gated in CI).
- E1 (P1 Β· Foundation) β Stand up the AI-native loop (repo + self-audit + gated
make check) - E2 (P1 Β· Foundation) β Map the field into a verified knowledge base
- E3 (P1 Β· Foundation) β Ship the build-in-public surface (README + live dashboard)
- E4 (P1 Β· Foundation) β Follow the frontier: radar + live tracker + field graph
- E5 (P1 Β· Foundation) β Scaffold the first two loop turns (runnable, honest, PROOF-gated)
- E6 (P1 Β· First Signal) β Run Turn 01 β Systems-Age heterogeneity test on the Biomarkers-of-Aging data
- E7 (P1 Β· First Signal) β First public leaderboard submission (Biolearn / Biomarkers Challenge)
- E8 (P2 Β· Momentum) β Run Turn 02 β fine-tune an open bio-FM vs a frozen probe (age prediction)
- E9 (P2 Β· Momentum) β Ship an open tool a lab actually adopts
- E10 (P2 Β· Momentum) β Secure 1 named academic/industry collaborator
- E11 (P3 Β· Leverage) β Close a non-dilutive micro-grant (VitaDAO / Foresight)
- E12 (P3 Β· Leverage) β Preprint OR top-decile leaderboard finish
- E13 (P4 Β· Wet-lab bridge) β Secure a lab partnership β computation-for-bench (Yale/Levine orbit or Genentech-alum PI)
- E14 (P4 Β· Wet-lab bridge) β First real wet-lab validation (or clean null) of a computational hypothesis, run with the partner lab
- E15 (P4 Β· Spin-out) β Incorporate the entity + recruit a wet-lab co-founder from the network
- E16 (P4 Β· Spin-out) β age1 acceptance OR a pre-seed raise underway (dilutive, full-time)
- E17 (P1 Β· Foundation) β Bridge gaps-analysis G1 (measurement): a standardized, reproducible cross-clock disagreement benchmark
- E18 (P1 Β· Foundation) β Bridge gaps-analysis G4 (hype outruns evidence): an open honest-nulls registry
- E19 (P1 Β· Foundation) β Bridge gaps-analysis G2 (reproducibility / FAIR): a datasheet + FAIR scorecard for open aging datasets
- E20 (P1 Β· Foundation) β Learn from mims-harvard/ATHENA; integrate the frontier signal into the field map
- E21 (P1 Β· Foundation) β Problem map: top aging problems as a relationship graph + the one target + low-hanging fruit
- E22 (P1 Β· Foundation) β Publish the tracker.json data contract for the interactive portfolio instance (data-first slice)
- E23 (P1 Β· Foundation) β Hub architecture (knowledgeΒ·toolingΒ·experts per cited repo) + the per-repo generation-recipe layer
- E24 (P1 Β· Foundation) β Hub Phase-2: JIT deep artifacts for one repo (pyaging) β knowledge graph + gated skill, SHA-stamped
- E25 (P1 Β· Foundation) β Make /longevity-loop an invocable skill that uses every hub toolset (backbone + progressive disclosure)
- E26 (P1 Β· Foundation) β Fan the hub out to Biolearn (North-Star toolset) β auto-wires into /longevity-loop
- E27 (P1 Β· Foundation) β External validation β register the loop's skills in HKUDS/OpenSpace, run 3 real tasks, read the quality records
| Execution | Metric | Before | After |
|---|---|---|---|
| β E1 | loop self-audit /100 | n/a (no repo) | 100 (9 principles, gated in CI) |
| β E2 | curated, URL-verified entries | 0 | ~75 (16 researchers, 16 startups, 26 tools, 17 ecosystem) |
| β E3 | public live artifacts | none | public repo + GitHub Pages dashboard + generated README |
| β E4 | frontier signal | 0 tracked | 10 verified recent works + weekly arXiv/GitHub tracker + 60-node bi-temporal graph |
| β E5 | runnable loop turns | 0 | 2 scaffolded (Turn 01 Biolearn/Systems-Age; Turn 02 bio-FM fine-tune) |
| β¬ E6 | outcome AUROC Ξ (best single clock vs +cross-clock heterogeneity), 5 seeds | pending | pending |
| β¬ E7 | public leaderboard rank / percentile | not entered | pending |
| β¬ E8 | held-out age MAE, frozen probe vs fine-tuned | pending | pending |
| β¬ E9 | external users / adopters | 0 | pending |
| β¬ E10 | named collaborators | 0 | pending |
| β¬ E11 | non-dilutive funding secured | $0 | pending |
| β¬ E12 | preprint posted / leaderboard percentile | none | pending |
| β¬ E13 | lab partnerships with bench access | 0 | pending |
| β¬ E14 | dryβwet validated (or null) findings | 0 | pending |
| β¬ E15 | incorporated entity + wet-lab co-founder | none | pending |
| β¬ E16 | accelerator / pre-seed capital | $0 | pending |
| β E17 | a shared, reproducible way to quantify how much a clock panel disagrees | none β clocks contradict each other, no consensus metric (triangulated across all 3 research windows) | clockbench.py β deterministic Spearman-agreement benchmark + CI selftest; demo panel headline 0.315, outlier auto-detected; real-data seam via --input |
| β E18 | a shared, cited registry of longevity nulls/failures (so the field stops re-learning them) | none β graveyards recur every window (resveratrol, NAD, young blood, monkey glands) but are unrecorded | data/nulls.yml (6 cited entries) β generated docs/NULLS.md + README section, schema-gated in validate.py |
| β E19 | FAIR/reproducibility scored + gated for the open aging datasets the loop uses | none β no FAIR/metadata standard; datasets picked ad hoc, reproducibility implicit | data/datasets.yml (7 datasets) β scripts/fair.py β docs/FAIR.md scorecard; panel 86/100; gate fails on any unassessed dim; CI-wired |
| β E20 | external frontier work studied β integrated (cited) into the loop | ATHENA / ToolUniverse / Zitnik not tracked; no study on file | research/athena-study.md (cited analysis) + Zitnik (people), ATHENA-R1 (frontier), ToolUniverse (stack) added; finding: ATHENA's citation-allow-list = our existing no-evidence-no-claim; RAG/multi-agent/ToolUniverse overkill for a solo code-only loop |
| β E21 | a clear, visual answer to 'what are we shooting, and what's the quick win' | gaps ranked (gaps-analysis.md) but relationships + the single target + quick wins were implicit | docs/PROBLEMS.md β Mermaid relationship graph (G1 measurement = root bottleneck we target) + top-problems table + how/when/why + 3 low-hanging fruit with done-when; README pointer |
| β E22 | a stable, derived JSON contract the agentic-portfolio instance can consume (no drift) | no machine-readable mission/progress feed for an external interactive surface | site/tracker.json β mission + 48% progress + 21 milestones + roadmap + low-hanging + frontier + 12 hubs + links, generated by build_site.py, deployed on Pages |
| β E23 | a fresh, cheap, future-proof way to offer a KG + agentic tooling for each cited repo | cited repos listed in stack.yml but no KG/tooling per repo; no design for how to do it well | docs/HUB_ARCHITECTURE.md (index+JIT design; reuse graphify/understand/reverse; MCP; SHA-cached) + scripts/repos.py β docs/REPOS.md (22 repos, per-repo KG + reverse recipes), gated in make check |
| β E24 | a real, grounded, provenance-stamped KG + skill generated from a cited repo, with a freshness gate | hub had the recipe layer only; no deep artifact proving the JIT layer works | hub/pyaging/ β knowledge-graph.json (20 nodes/edges) + compute-aging-clocks SKILL.md + manifest (SHA c2b3000e3c2c); scripts/hub_gen.py gates integrity (make check) + checks live staleness (FRESH); grounded in pyaging's real API |
| β E25 | an invocable /longevity-loop skill that reaches all per-repo toolsets without bloat | hub toolsets existed (hub/*/SKILL.md) but nothing exposed them through one invocable skill | SKILL.md generated by scripts/skill_gen.py β backbone = the loop's own tools; progressive disclosure = hub toolsets (auto-discovered from hub/, 1 today: compute-aging-clocks) + a JIT recipe; drift-gated in make check; installed at ~/.claude/skills/longevity-loop |
| β E26 | the North-Star repo has a grounded KG + gated skill, auto-disclosed by /longevity-loop | hub had 1 exemplar (pyaging); Biolearn (the leaderboard/loader for E6/E7) had no toolset | hub/biolearn/ β knowledge-graph.json (19 nodes/edges) + biolearn-leaderboard SKILL.md (real API, ties to E6/E7) + manifest (SHA 0d714f5a0c0a, FRESH); skill_gen auto-discovered it β SKILL.md now discloses 2 toolsets, no hand-editing |
| β E27 | the loop's skills register into a third-party quality-first hub and are selected/applied on real tasks (evidence, not vibes) | openspace-trial/ had a runbook + inventory but no results; the loop's skills had never been exercised outside this repo; the runbook's install step (pip install -e .) was unverified | source-installed OpenSpace v2.0.0 (py3.12, non-editable β editable install proven broken); 3 skills registered (trusted); 3 tasks all SUCCESS/completed; biolearn-leaderboard + longevity-loop judged 'skill applied' (2/3); evolution engine correctly NOOP (plans give no executed evidence); sandbox denied bash/write (blast radius contained). Verdict: trial-more |
6 logged. Longevity claims that failed, were refuted, or returned a clean null β recorded so the field (and this loop) stops re-learning the same failures. No evidence β no claim applies to negatives too. Full registry with sources: docs/NULLS.md.
- Resveratrol / sirtuin-activating compounds (Sirtris SRT501) β refuted (2010β2013): In-vitro activation β in-vivo benefit β watch for assay-artifact confounds before scaling.
- Antioxidant supplements for longevity (free-radical theory) β refuted (2007β2015): A compelling mechanism (oxidative damage) does not survive controlled outcome trials.
- Young-blood plasma transfusion (Ambrosia) β failed (2019): Striking mouse parabiosis data + paid 'trials' are not human evidence β regulators intervened.
- NAD+ boosters (NMN / NR) for human longevity β unproven (2016β2025): A moved biomarker (NAD+ up) is not a hard outcome β the exact biomarker-vs-endpoint gap (G3).
7 open aging datasets scored against FAIR (Findable Β· Accessible Β· Interoperable Β· Reusable) β 6 openly accessible. The field lacks shared FAIR/metadata standards for multi-omic aging data (gaps-analysis.md G2), so each dataset carries a datasheet + an honest self-assessment (unknown β no). Full scorecard with load recipes: docs/FAIR.md.
This repo is a hub: for each cited repo it offers a knowledge graph (knowledge) + agentic tooling (skills/plugins/workflows), and links them to the experts who build them. Design (high-quality Β· fast Β· cheap Β· fresh Β· future-proof): docs/HUB_ARCHITECTURE.md Β· per-repo generation recipes: docs/REPOS.md. Fresh index, just-in-time depth β deep artifacts are generated on demand and cached by commit SHA, never committed stale. First deep exemplar (Phase 2): hub/pyaging β a grounded knowledge graph + a gated skill, SHA-stamped, with a live staleness check (make hub / scripts/hub_gen.py --refresh).
The top aging problems, how they relate, the one root bottleneck we target (G1 β measuring aging), and the low-hanging fruit for quick wins β as a graph: docs/PROBLEMS.md. Root bottleneck = our lane; everything downstream is gated on being able to measure aging in a shared, validated way.
π€ = AI-forward Β· π¬ = active in the open community (good first contacts).
- JoΓ£o Pedro de MagalhΓ£es π€π¬ β University of Birmingham: HAGR aging databases; computational biogerontology
- Alex Zhavoronkov π€π¬ β Insilico Medicine: deep-learning aging clocks + generative AI drug discovery
- Peter Fedichev π€π¬ β Gero: physics/AI dynamical models of aging
- Jacob Kimmel π€π¬ β NewLimit (fmr Calico): ML-designed reprogramming payloads; bio foundation models
- Morgan Levine π€π¬ β Altos Labs (fmr Yale): epigenetic aging clocks (PhenoAge)
- Tony Wyss-Coray π€π¬ β Stanford: plasma-proteomic organ aging clocks
- Vadim Gladyshev π€ β Harvard Medical School: mouse + single-cell (scAge) aging clocks
- George Church π€π¬ β Harvard / Wyss Institute: gene-therapy longevity; synthetic biology
- Nir Barzilai π¬ β Albert Einstein College of Medicine: centenarian genetics; TAME metformin trial
- Matt Kaeberlein π¬ β Optispan (fmr U. Washington): rapamycin geroscience; Dog Aging Project
- David Sinclair π¬ β Harvard Medical School: information theory of aging; reprogramming; sirtuins
- Eric Verdin π¬ β Buck Institute: geroscience; ketone bodies; immune aging
- Andrew Steele π¬ β Independent (author of 'Ageless'): longevity science communication
- Charles Brenner π¬ β City of Hope: NAD+ metabolism; vocal longevity-hype skeptic (a good reality check)
- Joe Betts-LaCroix π¬ β Retro Biosciences: reprogramming + autophagy longevity company
- Kristen Fortney π€ β BioAge Labs: ML on longitudinal human data for aging drug discovery
- Marinka Zitnik π€π¬ β Harvard Medical School (Zitnik Lab / mims-harvard): AI agents for medicine; therapeutic reasoning over biomedical tools (ATHENA, ToolUniverse); graph ML for therapeutics
π€ = AI-native platform.
- NewLimit π€ β AI-guided epigenetic reprogramming to restore youthful cell function (well-funded-private)
- Retro Biosciences π€ β reverse cellular aging (reprogramming, autophagy) + AI protein design (well-funded-private)
- Gero π€ β physics-based AI modeling of aging + generative molecule design (well-funded-private)
- Shift Bioscience π€ β AI 'virtual cell' for safe single-gene cellular reprogramming (early)
- BioAge Labs π€ β aging-biology multi-omics ML platform (metabolic disease) (public)
- Insilico Medicine π€ β generative AI drug discovery (PandaOmics, Chemistry42) + clinical pipeline (public)
- Altos Labs β cellular rejuvenation via partial reprogramming (+ growing computation arm) (well-funded-private)
- Calico Life Sciences β basic biology of aging β age-related-disease medicines (well-funded-private)
- Recursion π€ β AI drug discovery via cellular imaging (Recursion OS; merged Exscientia) (public)
- Isomorphic Labs π€ β AI drug design built on AlphaFold (well-funded-private)
- Cellarity π€ β AI on single-cell transcriptomics to design cell-state-correcting medicines (well-funded-private)
- Xaira Therapeutics π€ β AI-native drug discovery on David Baker's generative protein models (well-funded-private)
- Rubedo Life Sciences π€ β AI-driven senolytics discovery (ALEMBIC platform) (well-funded-private)
- Gordian Biotechnology β high-throughput in-vivo pooled screening (Mosaic) for diseases of aging (early)
- Loyal β canine longevity β lifespan-extension drugs for dogs (well-funded-private)
- Centenara Labs β hallmarks-of-aging therapeutics portfolio (fmr Rejuveron) (well-funded-private)
- Geneformer β Transformer on ~30-104M single-cell transcriptomes; fine-tune for cell-type / perturbation / aging tasks.
- scGPT β GPT-style single-cell foundation model; fine-tune for annotation, integration, perturbation prediction.
- ESM-2 β Protein LM (8M-650M) that runs on a laptop; embeddings for variant-effect on longevity genes.
- TxGNN β GNN over a drug/disease knowledge graph for zero-shot repurposing; rank candidate longevity interventions.
- AltumAge β Deep pan-tissue methylation clock (MLP over ~20k CpGs) to fine-tune/benchmark vs ElasticNet clocks.
- AlphaFold β Structure prediction; usually query the precomputed AlphaFold DB rather than run locally.
- scanpy β Standard single-cell preprocessing (QC/clustering/DE) before any aging model.
- CZ CELLxGENE Census β API to slice ~33M+ standardized cells by tissue/age/disease in seconds β fastest cohort pull.
- pyaging β PyTorch package bundling 50+ aging clocks with one API β score biological age on a laptop.
- Biolearn β Open standardized platform for the Biomarkers of Aging Challenge β the code-only leaderboard to compete on.
- ToolUniverse β Open library of 200+ biomedical tools (drugs/targets/disease APIs) behind one interface β the agent tool layer ATHENA reasons over; a substrate a longevity agent could plug into.
- scAge β Epigenetic age from sparse single-cell methylation β detect cell-level aging + rejuvenation.
- DunedinPACE β Pace-of-aging (rate, not age) from 450k/EPIC β a strong intervention outcome variable.
- Tabula Muris Senis β Mouse aging cell atlas (~500k cells, 18 organs) β the go-to open single-cell aging benchmark.
- GTEx Portal β Human multi-tissue expression with donor age; expression matrices download freely (genotypes gated).
- NCBI GEO β Largest public expression/epigenomics archive; search 'age'-flagged series to build custom aging sets.
- Human Cell Atlas β Open multi-omic reference maps (70M+ cells) for age-stratified tissue baselines.
- HAGR (GenAge / CellAge) β Curated aging/senescence gene sets (GenAge 307, CellAge 866) β priors / feature filters.
- UK Biobank β 500k multi-omic + health cohort β ACCESS-GATED (application/fee/cloud-only), not laptop-downloadable.
- Biomarkers of Aging Challenge β Open competition + curated dataset (methylation/proteomics/outcomes, 500+ people) β the North-Star code-only leaderboard.
- Open Problems in Single-Cell β Community benchmarking harness (tasks/datasets/metrics) to fairly evaluate a model vs baselines.
- LAB-Bench β 2,457-question benchmark for LLMs on biology research tasks (literature, sequences, DBs).
- BixBench β Benchmark for LLM agents on real computational-biology analysis workflows.
β = realistically open to a solo/independent builder.
- Biomarkers of Aging Challenge / Longevity Prize β grant β Open, code-only competition + curated dataset (Biolearn) β the IDEAL first credibility-gated signal, no wet lab.
- VitaDAO / VitaLabs β community β DeSci collective; fast fellowship grants (~$65K), light application, active Discord β most accessible funding+community on-ramp.
- Foresight Institute β Longevity Grants β grant β Monthly-deadline frontier grants (AI-for-science + longevity), unusually open to non-traditional applicants.
- Foresight Fellowship β fellowship β Year-long fellowship; mentorship + intros to funders/senior scientists; global, independent-friendly.
- Longevity Biotech Fellowship (LBF) β fellowship β The main 'how do I get into longevity biotech' front door (ODLB merged in); cohort + retreat + community.
- age1 β accelerator β Laura Deming's longevity accelerator (~$500K, 4-mo). Dilutive β the premier founder on-ramp; apply with a public track record.
- Impetus Longevity Grants (Norn Group) π grant β Fast $10K-$500K aging-science grants (~3-4 wk decisions) β PI/lab-gated; partner into a lab to access.
- Hevolution Foundation π grant β Large geroscience funder ($300-500K/yr); institutional/PI-gated β a funder to partner toward.
- XPRIZE Healthspan π grant β $101M team competition (restore function 10-20 yrs). Team/clinical β highest-leverage as a rallying point + network.
- NIA (NIH) π grant β Largest US non-dilutive aging funder; SBIR/STTR is the realistic founder path once incorporated.
- Astera Institute (Rejuvenome) π grant β Runs the open ~$70M combinatorial mouse-lifespan dataset β build on it as open data; grants are relationship-driven.
- Vitalist Bay / Vitalism β community β Longevity pop-up city β the highest-density in-person gathering of the frontier/independent crowd.
- Aging Research & Drug Discovery (ARDD) β conference β The top translational-geroscience conference (Copenhagen) β where the serious science+investor crowd is.
- Longevity Summit Dublin β conference β Rejuvenation-biotech heavy (de Grey/O'Dea); researchers + advocates + investors.
- Longevity Marketcap (Nathan Cheng) β media β Most-read industry newsletter; Cheng is a hub node (LBF, Vitalism, Healthspan Capital) β engage to map who's who.
- Lifespan.io β media β Advocacy + news non-profit β a place to get build-in-public work amplified to an engaged audience.
- Longevity.Technology β media β Daily industry/investment newsletter β 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 β do not edit by hand. A sibling of FM-os.