Add descriptor-non-redundancy vs task-non-additivity distinction to halt-head section
Browse files
README.md
CHANGED
|
@@ -191,6 +191,14 @@ per-perturbation ordering of E[N] (best |ρ| = 0.23, below a 0.3 novelty ceiling
|
|
| 191 |
The signature is reproducible *within* a cell type but does **not** port across cell
|
| 192 |
lines (cross-line ρ = 0.14) — it is a cell-type-specific property.
|
| 193 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 194 |
---
|
| 195 |
|
| 196 |
## Provenance & caveats
|
|
|
|
| 191 |
The signature is reproducible *within* a cell type but does **not** port across cell
|
| 192 |
lines (cross-line ρ = 0.14) — it is a cell-type-specific property.
|
| 193 |
|
| 194 |
+
Two claims must be held apart. As a **descriptor**, E[N] is *non-redundant* with the
|
| 195 |
+
network (above): it captures per-perturbation structure the graph statistics miss. For
|
| 196 |
+
one downstream **target-class ranking task**, however, it is *non-additive* with a
|
| 197 |
+
STRING baseline built from the same functional modules — adding |ΔE[N]| gives no AUC
|
| 198 |
+
lift over response + STRING in three of four lines. Non-redundancy as a descriptor and
|
| 199 |
+
non-additivity for that one classifier are different statements; only the latter is
|
| 200 |
+
negative, and it does not diminish the signature's novelty.
|
| 201 |
+
|
| 202 |
---
|
| 203 |
|
| 204 |
## Provenance & caveats
|