BRIDGE: Bottleneck-Aware Regulator-Set Inference and Diagnosis for Cooperative Gene Regulatory Recovery

Most gene regulatory network (GRN) inference methods focus on ranking individual regulator-target edges. However, biological regulation often involves groups of regulators acting cooperatively. This paper asks whether current methods fail to recover these cooperative sets because they cannot learn the underlying biological structure, or because of identifiable bottlenecks in the retrieval, scoring, and decoding stages of the inference pipeline.

The authors introduce BRIDGE (Bottleneck-Aware Regulator-Set Inference and Diagnosis), a framework that decomposes the recovery of regulator sets into four distinct stages: candidate retrieval, set-level scoring, combinatorial decoding, and outcome evaluation. To diagnose failures, they developed TRACE (Targeted Recovery Attribution for Cooperative Evaluation), a diagnostic suite that uses metrics like coverage, score gaps, and true-set ranks to pinpoint where a model fails. They also introduce a 'leak-free' stress test using random nonlinear mechanisms to evaluate whether models can identify cooperative sets without relying on handcrafted product-correlation features.

The study reveals that edge-ranking performance is a poor proxy for the recovery of complete, functional regulator sets. Using the TRACE diagnostic suite, the authors show that even when the correct regulators are present in the candidate pool (high coverage), models often fail to rank the true set above incorrect alternatives. By implementing Residual HOS2—a model that adds a learned non-additive correction to pairwise scores—they achieved consistent improvements in Jaccard similarity and recall across controlled benchmarks. Furthermore, they demonstrate that for complex networks, a two-stage approach (proposing candidates via pairwise scores followed by HOS2 reranking) significantly reduces computational search costs while preserving recovery accuracy.