Disagreement-Informed Arbitration for Gene Regulatory Network Inference: A Score-Level Meta-Classifier and a...

Gene regulatory network inference methods routinely disagree about individual edges, and practitioners resolve those conflicts by choosing one method or averaging them all. We ask whether the conflict can instead be arbitrated per edge. A gradient-boosted classifier is trained on the raw scores that ten inference methods-correlation-based, information-theoretic, sparse-regression and tree-ensemble, including GENIE3, GRNBoost2, CLR and ARACNe-assign to each candidate regulator-target pair, so that the weight given to each method varies from edge to edge. Across six single-cell perturbation screens spanning four cell types, arbitration improves on mean ensembling by +0.056 AUROC on Adamson and +0.083 on Shifrut under target-grouped cross-validation. The evaluation protocol turns out to matter more than the model. Edge-level cross-validation, standard in this literature, inflates apparent gains by 0.060 AUROC through target-gene leakage-comparable to the entire honest improvement. The effect is far larger for methods that represent genes implicitly: a supervised graph-attention link predictor trained on identical folds scores AUROC 0.930 under edge-level cross-validation, better than anything else we evaluate, and 0.533 once target genes are held out. Any method that parameterises genes is exposed, which covers most graph- and embedding-based approaches. A five-category typology of inter-method conflict localises where arbitration pays off, with the largest gains on edges where the methods disagree and the smallest where they already agree, while adding nothing as model input; we therefore report it as a diagnostic instrument rather than a modelling contribution. We also characterise what the ground truth measures: most perturbed genes in widely used screens are not transcription factors, and a mediation screen bounds how much of the perturbation response can be direct.