Evidence audit of the Atlas’s mTOR corpus

This page measures the Atlas against itself. Not how many studies it holds — that number says nothing about the field — but what kind of evidence this Atlas’s mTOR pathway map is built on: how much of it reaches a human, which parts of the map have no human study cited behind them, how many links in the map rest on a single paper, and how much of what is claimed carries a recorded boundary.

Every number below is computed from the corpus at build time, never written by hand. The method is at the bottom, so anyone can recompute them or apply the same audit to a different pathway.

What the corpus is made of

Each study carries a code for the kind of claim it can support, not for how good it is. A mechanistic paper is not worse than a trial; it answers a different question. The distribution matters because a pathway described almost entirely by molecular work supports conclusions about molecules, not about people.

43 of 428 studies — 10.0 % — are human evidence. Almost everything else is animal, cellular or theoretical (the 4 off-ladder records are preprints and registered trials, which can include human studies not yet reported). That is the single most important number on this page, and it is a fact about this collection, not a measurement of the field.

Is this corpus’s human share rising over time?

If the pathway were maturing towards clinical use, the share of human work would rise over time. Here is what the corpus shows, by five-year window.

Which parts of the pathway reach a human

The map holds 121 causal links, each of which must cite at least one study in the corpus. Grading every link by the strongest evidence behind it shows where the pathway is understood in people and where it is understood only in cells.

15 of 121 links (12 %) have any human evidence behind them; 84 (69 %) rest on mechanistic or review papers alone. This is the quantified form of what the Atlas elsewhere calls the human endpoint desert.

How much rests on one paper

A link cited by one study is not wrong. It is unreplicated, which is a different thing, and it is the first place to look when a result fails to reproduce.

54 of 121 links (45 %) are carried by a single paper in this corpus. Some of those papers are definitive structures; others are one result nobody has repeated. The Atlas does not currently distinguish the two, and that is itself a gap.

4EBP1-LONGEVITY, 4EBP1-MITO, AKT-TSC, AMPK-MITOPHAGY, AMPK-MTORC1, AMPK-TSC, BISTERIC-MTORC1, CGASSTING-SENESCENCE, CR-MTORC1, DEPTOR-MTOR, ERK-TSC, EVE-BREAST, EVE-IMMUNE, EVE-LAM, EVE-PNET, EVE-RCC, FOXO-LONGEVITY, GATOR1-RAG, GLN-MTORC1-ARF1, GLN-RAG, HYPOXIA-REDD1, ISR-SALR, LARS-RAG, LEU-LARS, LKB1-AMPK, METFORMIN-AMPK, METFORMIN-MTORC1, MITODYS-MTORC1, MTORC1-HIF1A, MTORC1-MAPK, MTORC1-PGC1A, MTORC1-PROSTATE, MTORC1-ROS, MTORC1-SENESCENCE, MTORC2-LIPID, MTORC2-MAM, MTORC2-PROSTATE, MTORC2-SGK1, PDCD4-TRANSL, PI3K-MTORC2, PTEN-PI3K, RAGULATOR-CGASSTING, RAPA-LAM, REDD1-TSC, ROS-MTORC1, S6K1-PDCD4, SALR-MTORC1, SAM-SAMTOR, SAMTOR-GATOR1, SESN2-AGING, TBC1D7-TSC, TEM-MTORC1, TEM-RCC, ULK1-AMPK

How much carries its boundary

A claim without its conditions is a claim that has quietly been generalised. The Atlas records, for each link, the species, tissue, dose, nutrient state or duration the finding was established under — the point where it stops being universal.

71 of 121 links (59 %) carry a recorded boundary condition. The remainder are not thereby unconditional: they are links where the boundary has not yet been written down. Treating a blank as "holds everywhere" is the exact error this layer exists to prevent.

What kind of boundary

The sentence is for a reader; the tag is what makes it answerable. "Female mice only in SEL2009" is clear to anyone and invisible to a query, so each recorded boundary also carries a machine-readable kind. The first group below describes the biology; the second records a curator’s judgement that the link is disputed; the third is computed from the state of the corpus rather than written by hand.

Boundaries on the biology

Set by the curator

Computed from the corpus

This is the query no other resource can answer. Reactome, KEGG, SIGNOR and Open Targets record the interaction; none of them record that it was only ever shown in one sex, or only in cell culture, in a form you can query across a whole pathway. The two time-related kinds have a page of their own: how long the exposure lasted, and what happens when it stops.

What has been tested and failed

Findings that did not replicate, and compounds that did not extend lifespan, are kept with the same visibility as positive results.

The corpus holds 5 studies whose curated category is Negative_result (JIB2026, MAN2021, MIL2011, POU2013, STR2012). The label marks studies whose headline result was null; a trial with a null pre-registered primary endpoint may carry another category (MOE2025, the PEARL trial, is filed as Human). The pathway map also has a dedicated sign for a tested-and-null link (no-effect), and it is currently used 0 times. That gap is real: null relationships are being recorded at the level of the study but not yet at the level of the map.

Who has checked this

Each link in the map has a curation state. Proposed means it was drafted and cited but has not been signed off by a second reader; Contested means the Atlas records the direction, mechanism or scope of the link as disputed, which is a statement about the literature, not about review; Confirmed means a reviewer has checked the claim against the cited papers.

What this audit does not measure

• It measures the corpus, not the literature. The Atlas is a hand-picked subset. Every proportion here is conditional on what one curator selected, and selection is the largest uncertainty on this page.
• It says nothing about quality. An evidence code is a statement about the kind of claim a study can support. A badly run trial still codes as human evidence; an excellent structure still codes as molecular.
• It does not weight by size or replication. A trial of twenty people and one of a thousand count the same here.
• It cannot see what was never published. Negative results that never reached a journal are invisible to any audit of the published record, including this one.

Method

Figures are recomputed on every build by build_evidence_audit.py from two sources: the study corpus (atlas_data/studies_baked.json) and the pathway links baked into the map. No number on this page is typed by hand, so a change in the data changes the page and nothing else has to be remembered.

Link grading takes the strongest evidence code among the studies citing that link. Human evidence means codes S and H. Period counts use five-year windows by publication year; undated records (1) are excluded from that table only.

The audit is written to be pathway-agnostic: the same script, pointed at a different corpus and map, produces the same measurements for any signalling pathway. The interesting comparison is not this page in isolation but two of them side by side.

Data and code are CC BY 4.0 — see Data & Citation. Corrections are welcome and are logged; see About & Methodology.