Abstract
<title>Abstract</title> <p>Background: A benchmark rank is conditional on choices about which predictions count as candidates. In gene regulatory network (GRN) evaluation, candidate edges may include all gene pairs, only edges from transcription factors, or only edges from transcription factors to nominated target genes. Whether this restriction changes the cross-tissue portability of method rankings has not been isolated empirically. Methods: We conducted a transparent secondary analysis of 54 published median area-under-the-precision–recall-curve (AUPR) values: six inference methods evaluated in immune, kidney, and lung data under three nested candidate spaces. For each candidate space, we compared all 15 method pairs across all three tissue pairs. The primary outcome was the pairwise ranking-reversal rate. Secondary outcomes were Kendall rank concordance, Spearman rank correlation, and agreement on the top method. Uncertainty was assessed by resampling method-pair clusters; an exact 215 sign-flip test compared the most restricted space with the all-pairs space. Results: Cross-tissue reversals increased from 2/45 (4.4%; cluster-bootstrap 95% interval 0.0–13.3%) with all pairs to 10/45 (22.2%; 8.9–40.0%) with transcription-factor sources and 14/45 (31.1%; 13.3–48.9%) with transcription-factor sources and targets. The paired increase from the all-pairs to the most restricted space was 26.7 percentage points (95% interval 8.9–44.4; exact sign-flip p = 0.03125). Mean Kendall τ fell from 0.911 to 0.378, and agreement on the top method fell from 3/3 to 1/3 tissue comparisons. The ordered increase remained in all six leave-one-method-out analyses. Conclusions: In this small benchmark, biologically narrower candidate spaces were associated with less transportable cross-tissue rankings. The result does not establish a universal effect of candidate restriction, but it shows that candidate-space choice can alter comparative conclusions even when methods, metric, and score table are held fixed. GRN benchmarks should report rankings across candidate spaces rather than treating one candidate definition as neutral.</p>