20. Independent ground truth and the paper's own comparators

20.1 Why the earlier ground truth was too easy

With no ground truth on , agreement between methods is not accuracy. The independent checks used here — path-type status (pure adjacency, no diffusion), drug mapping (orthology, no network) and dissemination assays (organoids, no computation) — are all NetPert-independent by construction.

20.2 Cascade-knockout ground truth

The independent checks need no simulation: DIR status comes from direct adjacency to Twist1 and the responses; drug links come from human–mouse homology; assay readouts come from the Ewald-lab organoid screen. None of them sees NetPert scores.

20.3 Comparators

NetPert is the vectorised implementation verified in section 2. Comparators are driver-to-response betweenness (the paper's), full-network betweenness, out-degree and driver personalised PageRank (α = 0.85) — all computed on the same network.

20.4 Results

Method	Spearman vs NetPert	Top-20 overlap	Druggable in top-20
NetPert	—	—	9/20
Betweenness (driver-to-response)	0.58	16/20	9/20
Betweenness (full network)	0.59	17/20	8/20
Out-degree	0.51	2/20	6/20
Driver personalised PageRank	0.79	6/20	6/20

Figure 20. Druggable enrichment along NetPert ranks: 9/20 (45%) in the top 20 against 20.5% background.

Figure 21. NetPert weight versus expression change across the 168 in-network responses.

20.5 What this changes

NetPert is the consensus ranker on : closest to driver PPR (rho 0.79), close to betweenness (0.58–0.59, top-20 overlap 16–17/20), least like raw degree (0.51, overlap 2/20). Its top-20 are 45% druggable against 20.5% background — and 100% DIR.

Caveats. No ground truth exists, so none of this is accuracy. Edge-drop intervals use 3 repeats and are descriptive; uniform removal is milder than study bias. The assay subset (n = 18, responses only) cannot validate intermediate ranking.

Comparator supplement ()

**Top-30 rank heatmap**
Where the four rankings agree and disagree.

**Path-type counts**
DIR = direct driver+response neighbours; DI/IR = one-sided; I = rest.

**Weights by path type**
Direct driver–response connectors score highest on .

**Top compounds**
Most promiscuous mappers within the 1,200-gene subnetwork.

**Druggable vs other**
Whether mapped targets score higher than unmapped genes.

**Top druggable intermediates**
Prioritised genes that already have compounds.

**Compounds: Pdgfra**
10 mapped compounds for Pdgfra.

**Compounds: Stat3**
10 mapped compounds for Stat3.

**Compounds: Rela**
8 mapped compounds for Rela.

**Compounds: Hdac6**
10 mapped compounds for Hdac6.

**Compounds: Ctnnb1**
6 mapped compounds for Ctnnb1.

**Compounds: Nfkb1**
6 mapped compounds for Nfkb1.

**Compounds: Hif1a**
11 mapped compounds for Hif1a.