Structure Knows: The Monoamine Hypothesis Is Its Own Best Evidence

What structure reveals that content cannot

The monoamine hypothesis — the idea that depression is caused by insufficient serotonin and norepinephrine — has been the dominant theoretical framework for antidepressant drug development for over 60 years. It underlies the development of SSRIs, SNRIs, and the way clinical trials are designed, the endpoints that are measured, and the way pharmaceutical approvals are justified.

The question of whether this hypothesis is true or adequately supported is a matter of ongoing scientific debate. IRDME does not take a position on that debate.

But IRDME can ask a structural question: what does the topology of the evidence chain look like?

The experiment

M_MED1 was the first IRDME experiment in the medicine and philosophy-of-science domain. The dataset: 8 epistemic claims and methodological decisions in antidepressant research, modeled as a multilayer graph with three layers:
• d1 — justifies: which claims serve as theoretical justification for others (8 edges)
• d2 — selects_endpoints: which methodological decisions determine which outcomes get measured (4 edges)
• d3 — cites_as_support: which claims are actually cited as empirical evidence for the research program (7 edges)

Nodes: monoamine_hypothesis, ssri_mechanism_claim, hamd_endpoint, rct_efficacy, fda_approval, clinical_guidelines, chemical_imbalance_narrative, serotonin_biomarker_research.

Pre-registered before any analysis. Hash bfb7bbc1, timestamp 2026-05-25T11:17:10 UTC.

Four hypotheses

h1 — FPL directional inequality: r(justifies ↔ selects_endpoints) > r(justifies ↔ cites_as_support). Claims that justify each other should also select the same endpoints more than they cite each other as evidence. PARTIAL — direction correct (r=0.4082 > 0.3162) but not significant at n=8.

h2 — monoamine_hypothesis is rank #1 hub in justifies: CONFIRMED. monoamine_hypothesis is the top-degree, top-betweenness node in the justification layer. It is the structural anchor of the entire theoretical justification chain.

h3 — rct_efficacy is rank #1 hub in cites_as_support (hub shadow): This was the expected result — the prediction was that the justification layer and the citation layer would have different top hubs. The clinical trial results (rct_efficacy) would lead the evidence layer while the theoretical assumption led the justification layer. DENIED.

h4 — r(selects_endpoints ↔ cites_as_support) > 0.30: PARTIAL — the endpoint-selection layer has only 4 edges for 8 nodes, making it structurally degenerate. Correlation is undefined under degeneracy.

The finding that wasn't predicted

h3 DENIED is the most important result. The prediction was for a hub shadow — the same pattern as the photon in the Standard Model, where a node is dominant in one layer and absent in another. The prediction was that monoamine_hypothesis would lead the justification layer while rct_efficacy would lead the evidence layer, suggesting that clinical trial results had become an independent empirical anchor.

What IRDME found instead was hub dominance: monoamine_hypothesis is rank #1 in BOTH justifies AND cites_as_support.

Layer rankings:
• justifies: monoamine_hypothesis #1, ssri_mechanism_claim #2
• cites_as_support: monoamine_hypothesis #1, ssri_mechanism_claim #2, clinical_guidelines #3

The founding assumption does not just justify the research agenda. It is simultaneously the most-cited empirical support for that same agenda.

This is the IRDME structural definition of a self-referential evidence loop. It does not matter what the content of the papers says. The structure of the citation and justification graph — which node is most connected to which, in which relational layer — reveals that there is no independent empirical node outside the loop that serves as a primary anchor.

Hub dominance vs hub shadow

These are structural opposites. In the Standard Model (M_PHYSICS_1), the photon is rank #3 in force_coupling and degree=0 in decay_channel — a hub in one layer, invisible in another. This is hub shadow: the structural signature of a node that organizes interactions but does not participate in outcomes.

In the antidepressant evidence chain, monoamine_hypothesis is rank #1 in the justification layer AND rank #1 in the citation layer. This is hub dominance: the structural signature of a node that organizes both the theoretical apparatus AND the empirical record. The assumption and the evidence are the same entity at the network's center.

If the evidence structure were healthy (structurally independent), the top hub in justifies would be different from the top hub in cites_as_support. You would see clinical trial results dominate the evidence layer while the theoretical mechanism dominated the justification layer — different nodes for different epistemic functions.

The HAM-D finding

There is a second structural result that deserves attention. hamd_endpoint (the Hamilton Depression Rating Scale — the primary efficacy measurement tool in most antidepressant trials) is the rank #1 hub in selects_endpoints and rank #5 in justifies.

The structural interpretation: HAM-D is the most methodologically constraining decision in the chain. It shapes what gets measured in nearly every trial. But it is not well-justified within the theoretical framework — it ranks low in the justification layer.

A well-justified methodology would have high rank in both layers. HAM-D's structural dissociation between methodological dominance and theoretical justification raises a structural question: why is the measurement tool that shapes everything the least-justified element in the theoretical chain?

What this does not say

This analysis does not say that SSRIs are ineffective. It does not say antidepressants are unsafe. It does not adjudicate the ongoing scientific debate about whether the monoamine hypothesis is true.

It says: the graph topology of the evidence chain, modeled from 8 nodes and the relational structure between them, has the structural signature of circularity. That is a methodological observation about the architecture of the evidence, not a claim about drug efficacy.

The n=8 dataset is a representative model, not an exhaustive literature graph. The FPL partial (h1) and endpoint degeneracy (h4) are both consistent with a small, underpowered graph. A v2 run with 20+ nodes would allow proper significance testing.

Running the experiment

The dataset (antidepressant_evidence_chain.json, 8 nodes, 3 layers) is available on irdme.com/datasets. The pre-registration record is public: github.com/vladi160/preregistrations.

The next step is M_MED2: a vaccine evidence chain (expected to be structurally healthy, different top hubs in different layers) as a comparison case. If the structural pattern differs between vaccine approval and antidepressant approval evidence chains, that would be a more direct structural finding.

Structure knows things content cannot.