Abstract
<title>Abstract</title> <p>Claims that one gene regulatory network (GRN) inference method outperforms another depend on both the data and the evaluation protocol. We examine this dependence in a small, controlled benchmark using the five 100-gene networks from the DREAM4 In Silico Network Challenge. Four transparent estimators—absolute Pearson correlation, shrinkage partial correlation, histogram mutual information, and target-wise Extra Trees—were fitted to 25, 50, or 100 multifactorial expression profiles. The same predictions were evaluated over either all 9,900 directed non-self edges or an oracle candidate set restricted to genes with at least one true outgoing edge. Across ten reproducible subsampling runs, performance at 100 profiles remained near the random-prevalence baseline: median ratios of area under the precision– recall curve (AUPRC) to edge prevalence ranged from 0.91 to 1.05 under the all-edge protocol. Restricting candidates increased raw AUPRC because prevalence rose, but did not improve prevalence-normalized AUPRC. More importantly, the restriction reversed 95 of 900 matched pairwise method comparisons (10.6%); pair-specific reversal fractions ranged from 3.3% to 18.0%. Changing sample size reversed 26.3–32.3% of aggregated pairwise comparisons. These results do not rank modern GRN methods. They show, within a deliberately narrow low-signal setting, that relative order is fragile and that candidate-universe and prevalence choices must be reported alongside benchmark scores.</p>