The paper addresses a real pain point in single-cell signaling causal discovery—lack of replicability across batches/labs—and proposes an intuitively sensible fix: using intervention metadata to bias the graph search proposal distribution. The claimed empirical pattern (improved edge stability, fewer implausible feedback loops, larger gains when interventions are sparse but reliable) is directionally credible, and comparing to a NOTEARS-style baseline is a reasonable starting point for positioning. If the intervention catalog is genuinely open and the evaluation uses replicated datasets, the work has potential practical value because it targets stability/interpretability rather than merely optimizing fit. The main weakness is that the excerpt leaves the causal identification story and evaluation protocol under-specified, making it hard to judge whether the stability gains reflect better causal recovery or simply stronger regularization toward prior beliefs. Key uncertainties include: how priors are encoded (strength, form, and sensitivity), whether hyperparameters are tuned on test replicates, what “edge stability” means quantitatively (selection frequency under resampling? agreement across batches?), and whether held-out perturbation settings constitute a proper out-of-distribution test with appropriate metrics (e.g., interventional predictive checks or known-target recovery). Without ablations (prior vs no prior; varying prior correctness; robustness to mis-specified interventions) and clearer reproducibility details (code, preprocessing, exact datasets), the conclusion that the method improves replicability is plausible but not yet fully justified as a causal improvement rather than a biasing/regularizing effect.

A careful causal-discovery paper that actually acknowledges the constraints of intervention metadata instead of pretending the graph is uniquely identifiable.