Feature preview release — data and UI are under rapid iteration; published snapshots define the reporting protocol.
Primary verdict
nominal single-module gain, not significant
Not significant does not mean the study lacks value; it only means no detectable independent increment on this decision task of this benchmark.
Key metrics
Per-disease breakdown
| Disease | Single-module Recall@20 | Atlas increment |
|---|---|---|
| Athero | 0.000 | 0.000 |
| Dep | 0.050 | 0.000 |
| Hyper | 0.050 | 0.000 |
| Osteo | 0.000 | +0.050 |
| T2D | 0.000 | +0.050 |
In-family evaluation
Joins family Hallmarks (72 modules)：family-additive
Family increment per disease
Complementary to the 332-module atlas view: the nomination is evaluated as a new member of its closest family (e.g. Hallmarks). The replace test asks — if it replaced the family's most similar existing module, would family performance hold?
Atlas neighborhood
In the atlas, this gene set is closest to NUT:Clove oil (NUT family, r = 0.562)。
Family affinity (mean correlation per module family)
Based on Pearson correlation of z-score columns: which classic modules your gene set sits closest to in the 332-module atlas — high affinity overlaps known atlas regions; low affinity may open new territory.
Most similar modules (redundancy)
Redundancy proxy based on Pearson correlation of z-score columns; high correlation means the concept is already covered by existing atlas modules.
AI-proposed gene set (46)
For concept nominations (Track A), the gene set was expanded by our LLM pipeline and is fully published for audit and reproduction.
steeramed_bench Table 1 (cv_logistic C=0.1, seeds 42/123/456, capped Recall@20)