22. Main experiments repeated with the independent truth

Sections 7 and 11 characterised behaviour on simulations. Here the same questions are asked of the network: agreement persists down the ranking (top-10/20/50 overlap with betweenness 8/16/37) while raw degree diverges immediately (2/20) — NetPert is not hub ranking, and the top-100 are all DIR genes. The analogue of false responses is missing ones (14/182 absent; assay subset n = 18, responses only). Twist1 has 199 out-neighbours and the TRRUST ablation (rho 0.945) shows regulatory direction contributes without dominating.

22.1 Number of response genes

Figure 23. Expression change versus dissemination effect (18 assayed genes).

22.2 False responses in the response list

Figure 24. Betweenness versus dissemination: null, as with NetPert (same 18-gene subset).

22.3 Driver degree and directed networks

Figure 25. Twist1 degree context in the interactome.

22.4 Summary of the revised picture

22.4 Summary of the picture. Against every NetPert-independent check available — DIR saturation, drug enrichment, missing-edge stability, relaxation insensitivity — the ranking behaves well. What it lacks is a ground truth: treat it as a prioritisation, not a validation.

Three repeats per edge-drop level means wide intervals, so small differences are suggestive, not established. All results are -network measurements.

Sensitivity slices ()

**NetPert vs assay**
Honest null: rho=+0.31, p=0.21 (responses, not intermediates).

**NetPert vs |log2FC|**
Strong DE genes are not systematically top-ranked.

**Up vs down weights**
No systematic up/down bias in NetPert weights.

**Connectivity vs assay**
Hubness alone does not explain inhibition.