Spokes.wiki Search About

Synthesis — Life Sciences

Where this spoke came from

Three sources sat in the hub _inbox under bio-ai because no spoke owned them, and each park record said why: psychology-wiki owns minds and behaviour, machine-learning-wiki owns models as models, ai-governance-wiki owns policy — and a proteome, a hippocampal epigenome and a biology tokeniser are none of those. The curator named the spoke life-sciences on 2026-09-17.

What the three say together

Two are results and one is infrastructure, and they sit at different depths of the same stack. hippocampus-aging-epigenome is a dataset-scale result: single-cell measurement of one tissue across ages, and the finding is a change in which cells are there — resident microglia replaced by blood-derived, more inflammatory ones between about 50 and 75. autism-protein-interaction-map is a map-scale result: 1,800 interactions rewired by disease mutations, with a finished AI model (AlphaFold) used to choose what to test. israel-ai-bio-consortium is the layer under both: a representation ("Bio Tokens") and a platform ("Factory Model") so that the next result comes from a model trained over data like the first, rather than from an instrument applied to the second. The through-line is biological-foundation-model as a thing being built toward.

Since 2026-09-22 the spoke holds one, and the first thing it shows is that the claim depends on the use. geneformer was trained on 29.9 million single cells and made a prediction that held up in engineered heart tissue — after fine-tuning on labelled data. Tested by an outside group with no fine-tuning, its embedding lost to a simple gene filter (single-cell-foundation-model-zero-shot). At the far end of the pipeline, rentosertib-phase-2a carried an AI-found target and an AI-designed drug through a 71-patient placebo-controlled trial: safe enough to continue, with a lung-function signal at the top dose. The company that made the drug also wrote the report. So each step from model to medicine now has an example in this spoke, and none has been checked end to end by anyone independent.

A second thread: where brain regions come from

two-brain-progenitor-lineages (added 2026-09-27) sits apart from the aging/disease/AI-pipeline thread above — it's developmental biology, not an AI-driven result, and it doesn't close any growth edge below. It belongs in the spoke on domain (molecular neuroscience, the organism's own biology) and is recorded as its own thread rather than forced into the existing one. The finding: the forebrain and midbrain come from one progenitor lineage (marked by Otx2), the hindbrain from a separate one (Gbx2), fixed apart by chromatin state before any brain structure forms, and conserved across roughly 550 million years of animal evolution. The practical payoff is a working recipe for human hindbrain motor neurons from stem cells — a cell type that had resisted directed differentiation.

That payoff is the one place this thread touches the other: rentosertib-phase-2a and israel-ai-bio-consortium both gesture at AI-assisted drug discovery reaching the clinic, and hindbrain motor neurons are exactly the disease-modeling substrate spinal muscular atrophy and ALS research would need. Nobody has yet run that pipeline on cells derived this way — worth watching, not yet a finding.

Open questions

Founding weakness

Two of three primaries are unread. Science returns 403 to this hub; CNN geo-blocks. The pages say so line by line. Half closed 2026-10-08 (quality cycle): the hippocampus study's own 2024 bioRxiv preprint (PMC11507755) is open and was read via pypdf — same study, same first authors, not the final published text. It supplies the N (40 donors, 4 age groups) and an astrocyte effect size (p=0.0002) the release never gave, and it contradicts the release's lineage framing: the preprint describes microglia as two states of the same resident population (Micro1→Micro2, an epigenetic switch), with no mention anywhere of yolk-sac origin, bone-marrow lineage, or physical replacement by an infiltrating population — the "replacement" claim this spoke has carried since 2026-09-17 is not in this text. Recorded as a live discrepancy between the preprint and secondary reporting on the published paper, not resolved — see hippocampus-aging-epigenome. The autism-map primary (CNN geo-block) is still unread.

Growth edges

Ranked; each names the kind of source that would close it (see ../QUALITY.md → Growth edges).

  1. A biological foundation model, described. Its data, its training, one confirmed prediction. — needs: a T1 paper or a T2 technical report of a named model (single-cell, protein, or multimodal). Closed 2026-09-22 by geneformer (T1, Nature 2023), with an outside zero-shot test beside it (single-cell-foundation-model-zero-shot, T1).
  2. The two Science primaries. Read them; replace the release-level numbers with the paper's N, ages and effect sizes. — needs: the PDFs. Half done 2026-09-22: the hippocampus abstract is read (PubMed 42490474) and changed one claim ("resembling" monocyte-derived, not traced from bone marrow). Advanced 2026-10-08 (quality cycle): the hippocampus study's 2024 bioRxiv preprint (same authors, PMC11507755) is full-text open and was read. It supplies N=40 and an astrocyte effect size, and it surfaces a bigger problem than a missing number — the preprint's own microglia finding is a state-switch (Micro1→Micro2) inside the resident population, not a lineage replacement, which contradicts the framing this spoke has carried since founding. Still needs: the published Science text itself, to see whether the final paper added a genuine lineage analysis (as secondary reporting on it claims) or whether "replacement" was always a release-level overstatement of a state-transition finding. The autism-map CNN primary is still unread.
  3. Longitudinal confirmation of midlife microglial turnover. — needs: a T1 animal or human time-series study.
  4. Drug discovery as a pipeline, not a press release. The consortium and the autism map both promise therapies; nothing here follows one. — needs: a T1/T2 account of an AI-derived candidate reaching a trial. Half close 2026-09-22: rentosertib-phase-2a is an AI-derived target and drug through a registered phase 2a trial, but it is the sponsor's own report. The residue is edge 5.
  5. The same candidate, followed from outside. An independent account of an AI-derived drug in the clinic — a phase 2b/3 result, a regulator's review, or a trial not run by the discovering company. — needs: T1/T2 from a party other than the sponsor. (new 2026-09-22, residue of 4) Status 2026-10-11: the phase 3 exists (GENESIS-IPF-3, first dose 2026-09-09) but is sponsor-run and not due to read out before late 2029. A result from it would still not close this edge on its own.
  6. Pretraining with no labels. single-cell-foundation-model-zero-shot found Geneformer's zero-shot embedding weaker than a gene filter. Has any biological foundation model since shown zero-shot value on data its builders did not choose? — needs: a T1 benchmark by a third party. (new 2026-09-22)

Contradictions / tensions

Cross-spoke adjacency

Index — Life Sciences Wiki

Catalog of every page, grouped by schema.org @type. Spine: synthesis (thesis), log.md (history), this file (catalog). Spun out of the hub _inbox on 2026-09-17 (3 sources). Two of the three founding sources are held through secondaries; the Science primaries are unread (one abstract read 2026-09-22).

DefinedTerm (concepts / methods)

ScholarlyArticle (sources)

NewsArticle (sources)

Organization (entities)

Person (entities)

Synthesis

Related pages

Links to

Linked from