Architecture

Nine layers from biological truth to a deployable decision. Four of them — 2, 4, 7 and 9 — are additions to the obvious five-layer model, and four of the nine failure modes live in exactly those four.

#LayerFailure if unguardedGuard
1REALITY
Does modulating target T actually cause endpoint E?
On-target biology conflated with drug pharmacology.unresolvable; declared in the model card
2SELECTIONadded
Which truths ever get the chance to become evidence?
Degree bias and confounding by indication are born here, not in the model.audit/degree.py · data/attribution.py
3EVIDENCE
What have we actually observed?
Undated edges admitted, making temporal claims unverifiable.data/ingest.py
4ATTRIBUTIONadded
Drug-level observation to target-level claim. Which target did it?
Uniform propagation makes the label a function of target degree.data/attribution.py — noisy-OR + Mantel-Haenszel
5LABELING RULE
How do we interpret the evidence?
Unlabelled silently treated as negative; tiers pooled.data/labels.py — three states, tiers never pooled
6DATASET LABEL
positive / earned-negative / unlabelled, per tier
Multiaxiality corrupts the label matrix.data/ontology.py · models/hierarchy.py
7SPLITadded
Which labels does the model see, and which is it tested on?
An interpolation split behind an extrapolation claim.data/splits.py · audit/leakage.py
8MODEL
What can the GNN learn from it?
A popularity detector with an excellent AUROC.models/hetero_gnn.py — degree offset, partial rho
9VERDICTadded
Promote or reject? At what operating point, on what evidence?
A score shipped where a decision was required.eval/metrics.py · pipeline.py

The three nested questions

Each is stated in full rather than by label: the first draft of the PDF named Q2 and Q3 in a panel and defined them nowhere.

Q1 Is there any signal beyond node popularity?

Contrast: GNN vs a degree-only null: logistic regression on log-degree, no biology, no relations.

If the contrast ties: The knowledge graph encodes study attention, not safety biology.

Q2 Does the graph help beyond non-graph target features?

Contrast: GNN vs a tabular model on target annotations only: expression breadth, pLI and LOEUF, protein family.

If the contrast ties: Graph structure adds nothing. Ship the cheaper tabular model.

Q3 Do the relation TYPES carry the signal?

Contrast: Relation-aware (R-GCN, CompGCN, HGT) vs relation-agnostic (GraphSAGE, GAT on a collapsed adjacency).

If the contrast ties: Connectivity matters; relationships per se do not. This is the literal research question.