Alzheimer's Amyloid Network: Circular Evidence, a Citation Anomaly, and a Structural Contrast with Schizophrenia

Background: the M_MED series

IRDME's pharmaceutical evidence topology series models medical evidence chains as multilayer graphs and analyzes their structural properties using pre-registered hypotheses. Each drug program is encoded as three epistemic layers:
• justifies -- claim A provides the theoretical or mechanistic basis for claim B
• selects_endpoints -- A determines what outcome measures are used when evaluating B
• cites_as_support -- A is cited as empirical evidence in support of B

Previous results across four experiments:
• M_MED2 (vaccines, healthy control): r = +0.75. Founding mechanism and citation hub are structurally distinct -- healthy differentiation.
• M_MED1 (antidepressants): r = +0.41. Monoamine hypothesis dominates all layers -- mild circularity.
• M_MED3 (opioids): r = +0.14. Pain_undertreated_claim dominates both layers. Porter and Jick citation anomaly detected by topology alone.
• M_MED4 (schizophrenia): r = -0.13. PANSS endpoint structurally decoupled from theory. Dopamine and NMDA research communities have zero structural contact across all three layers (paradigm islands -- graph_connected = False).

M_MED5 tests the Alzheimer's amyloid cascade hypothesis.

Pre-registration

All four hypotheses committed before analysis. Hash: 20286f0b. Timestamp: 2026-06-03T18:48:04Z. Records: github.com/vladi160/preregistrations, experiment M_MED5_v1.

Results: 3 confirmed, 1 partial

h2 CONFIRMED -- amyloid_cascade_hypothesis is rank #1 hub in the justifies layer (degree = 6). It directly justifies APP processing, the amyloid biomarker endpoint, and -- critically -- predicts tau pathology as a downstream consequence. This last edge makes the Alzheimer's network structurally different from schizophrenia: it is a cross-paradigm justification edge connecting the amyloid and tau research streams.

h3 CONFIRMED -- amyloid_cascade_hypothesis is rank #1 hub in the cites_as_support layer (degree = 6). Genetic evidence, ADNI longitudinal data, even failed drug trials, and treatment guidelines all cite the cascade as their empirical or theoretical foundation. The circular structure (guidelines cite the cascade; the cascade cites the guidelines) replicates the pattern found in antidepressants, opioids, and schizophrenia. This is now four out of four tested circular chains.

h1 PARTIAL -- r(justifies, selects_endpoints) = +0.089, Spearman = +0.426, p = 0.699. Direction is positive (unlike schizophrenia's -0.128) but below the pre-registered threshold of 0.20. The Alzheimer's amyloid evidence chain is in a hybrid regime: the endpoint (amyloid PET, CSF Abeta42) is theoretically motivated -- the cascade specifies Abeta42 as the mechanism, making measuring it intrinsically sensible -- but also institutionally amplified beyond that coupling by FDA accelerated approval and guideline embedding. Pearson is near-zero because the biomarker endpoint has degree = 4 in selects_endpoints (four nodes point to it via institutional channels) but only degree = 1 in justifies (only the cascade directly justifies it). The rank-order coupling is real (Spearman = +0.43), but the magnitude coupling is diluted.

h4 CONFIRMED -- lesne_2006_oligomers is rank #5 in cites_as_support. It has degree = 0 in justifies -- zero edges, last place.

The Lesne 2006 citation anomaly

Lesne et al. (2006) Nature reported a 56-kilodalton amyloid oligomer species (Abeta-star-56) that impaired memory in aged rats. The paper was widely cited as evidence that soluble oligomers, not plaques, are the toxic amyloid species in Alzheimer's disease -- refining the cascade hypothesis. It received over 2,300 citations.

In July 2022, Charles Piller published an investigative report in Science raising concerns about image manipulation in multiple figures of the Lesne paper and related lab work. The University of Minnesota investigated. The paper was retracted from Nature in November 2024.

In this dataset, lesne_2006_oligomers has zero edges in the justifies layer by design -- its justificatory standing has been removed. It retains citation edges because citations in the literature persist regardless of the paper's status.

IRDME produced: rank #12 (last) in justifies, rank #5 in cites_as_support. A rank gap of 7 positions.

This is the same structural pattern as Porter and Jick in M_MED3: high citation centrality, low justification centrality. Two independent cases, two unrelated medical domains, same topology.

What makes this notable: IRDME detected the anomaly from graph structure alone, without reading the paper. The tool correctly identified a paper that domain experts and investigative reporters later confirmed had problems -- using only the structural position of the citation in the evidence network.

The M_MED4 vs M_MED5 structural contrast

Before M_MED5, the paradigm islands finding in schizophrenia (graph_connected = False; zero cross-paradigm edges between dopamine and NMDA branches) could have been a construction artifact of how that specific dataset was built.

After M_MED5: the Alzheimer's graph is connected (graph_connected = True). The amyloid cascade explicitly predicts tau tangles as a downstream consequence, creating a justification edge between the two paradigm branches. ADNI data is cited as evidence by both the amyloid and tau communities.

This makes the connected/disconnected distinction a comparative observation between two real domains, not a methodological choice. Schizophrenia and Alzheimer's research are structurally organized differently in their paradigm engagement:
• Schizophrenia: dopamine and NMDA theories coexist without formal structural contact
• Alzheimer's: amyloid and tau theories are integrated -- the dominant theory predicts the competing one as downstream

That is a structural difference in how each field has organized its evidence, detectable from topology.

The five-position structural spectrum

Five experiments, five structural positions on a measurable spectrum:
• Healthy differentiation (vaccines): r = +0.75
• Mild circularity (antidepressants): r = +0.41
• Strong circularity (opioids): r = +0.14
• Hybrid / partial endpoint lock-in (Alzheimer's amyloid): Pearson = +0.09, Spearman = +0.43
• Measurement-layer dominance (schizophrenia): r = -0.13

What IRDME cannot say

This analysis reads graph structure, not scientific content. It cannot:
• Determine whether the amyloid hypothesis is correct
• Evaluate whether lecanemab or donanemab provide clinically meaningful benefit
• Say whether the Lesne paper's data integrity concerns are valid (that is documented elsewhere)
• Determine the right treatment for Alzheimer's disease

What it can say: the structural topology of the published amyloid evidence network shows circular hub dominance, a partially institutionalized endpoint, and a high-citation paper with zero justification edges -- patterns also found in other contested pharmaceutical evidence chains.