- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes GRASP, a framework for supervised gene regulatory network (GRN) inference from single-cell RNA sequencing data and partially observed transcription factor (TF)-target interactions. The method combines three components: expression-guided edge retention to adapt local graph propagation, a momentum teacher-based higher-order representation to incorporate context beyond observed graph neighborhoods, and Neighborhood Contrastive Learning (NCL) that treats known training partners as positive pairs within and across views. A role-specific decoder is used to predict directed TF-target interactions. The paper also adopts a pair-disjoint evaluation protocol to prevent reverse-edge connectivity leakage under symmetrized message passing. Experiments across 42 settings, covering seven cell types, three reference networks, and two gene set sizes, show that GRASP achieves the highest overall average AUPRC among eight baseline methods. Additional ablation, robustness, cross-cell transfer, and external annotation analyses are provided to evaluate the framework.
Critical strengths
1. Comprehensive experimental evaluation and competitive prediction performance.
The paper provides an extensive evaluation across 42 settings involving seven cell types, three reference networks, and two gene set sizes. The comparison includes eight representative baselines, covering both traditional expression-based methods and recent graph-based approaches. GRASP achieves the highest overall average AUPRC at both TF+500 and TF+1000, with scores of approximately 0.35 and 0.34, compared with 0.31 and 0.31 for GRANet. The use of multiple data splits and random seeds further strengthens the experimental evaluation. These results provide evidence that the proposed framework is effective across different reference sources and cellular contexts.
2. Well-structured ablation studies and attention to evaluation reliability.
The paper includes a complete factorial ablation study to examine the contributions of expression-guided augmentation, higher-order context, and neighborhood contrastive supervision. The results show that these components provide complementary benefits. Another notable strength is the pair-disjoint evaluation protocol, which prevents held-out gene pairs from remaining directly connected through reverse interactions in a symmetrized training graph. This addresses an important evaluation issue in supervised graph-based GRN inference. The additional robustness and cross-cell transfer experiments also help assess the method beyond the main benchmark.
Critical weaknesses
- The proposed components are not sufficiently compared with simpler alternative designs, making their specific methodological contributions less clear.
- The performance improvements vary across different reference networks and cell types, with relatively small gains or inferior results in some settings.
- The evaluation protocol comparison changes both training graph connectivity and test candidate composition, making it difficult to isolate the impact of reverse-edge leakage.
Questions for authors
- Could the authors provide a controlled comparison between the proposed higher-order context module and a simpler global aggregation method to better demonstrate its effectiveness?
- Could the authors explain why GRASP achieves relatively small improvements on the Specific reference networks and underperforms some baselines in certain STRING settings?
- Could the authors clarify how changes in test candidate composition may influence the observed performance differences between the legacy and pair-disjoint evaluation protocols?
2 comments
(Part 2/2, continued.)
Q3. The protocol comparison changes both training connectivity and test candidate composition; how does composition affect the legacy versus pair-disjoint gap?
R3. Thanks for your valuable question! We examined the two factors separately on Non-Specific (NS) and STRING (the legacy-pair-disjoint AUPRC gap on Specific is below 0.01 at both scopes; Table 3).
Candidate composition. We freeze each legacy model with its training graph and compare its original test candidates with the intersection of the legacy and pair-disjoint test sets, the only pool on which neither protocol's model has trained (both exclude candidates whose reverse is a training positive; class prevalence is standardized).
Connectivity and leakage. Keeping supervised pairs, NCL positive sets and test candidates fixed, we randomly inject test-positive connections into the pair-disjoint propagation graph only, at the rate at which legacy test positives have a reverse training edge.
Mean over seven cell types; protocol gap = legacy − pair-disjoint AUPRC (Table 3).
| Reference | Scope | Protocol gap (AUPRC) | Composition ΔAUPRC | Injected (% of test positives) | Connectivity ΔAUPRC | Connectivity ΔAUROC |
|---|---|---|---|---|---|---|
| NS | TF+500 | +0.075 | −0.0194 | 4-28 | +0.0110 | +0.0550 |
| NS | TF+1000 | +0.052 | −0.0015 | 4-25 | +0.0177 | +0.0365 |
| STRING | TF+500 | +0.230 | +0.0110 | 22-49 | +0.1268 | +0.1787 |
| STRING | TF+1000 | +0.271 | −0.0245 | 13-42 | +0.1211 | +0.2175 |
Composition shifts individual settings in both directions, but its average effect (at most 0.025) is small relative to the protocol gap (0.052-0.271). Connectivity does affect prediction, but the connections at issue are exactly those of held-out pairs: this change is itself part of how pair-disjoint prevents reverse-edge leakage, not an independent effect. The STRING gains with test candidates unchanged support our interpretation in Section 3.4 that reverse-edge visibility can produce overly optimistic evaluation.
Thanks again for your careful review and constructive suggestions. Please let us know if you have further questions.
We are truly grateful for the time you have taken to review our paper and for your insightful comments. We address them below, where Q denotes the comment of the reviewer and R denotes our response.
Q1. Can the authors compare the higher-order context module with a simpler global aggregation method?
R1. Thanks for your valuable suggestion! We ran the requested control on STRING/TF+1000 under the protocol of Table 2. The learned latent hypergraph context Zh is replaced by the parameter-free global mean of the teacher representations, so every gene receives the same shared context vector; everything else is kept unchanged. Mean test AUPRC falls to 0.1504, between the full model and the variant without the higher-order view (Table 2):
A single shared vector recovers only part of the benefit, not the part that depends on which genes are pooled together: the learned incidence assigns different gene subsets to different latent hyperedges, and this gene-specific redistribution is what a global mean cannot provide. The learned grouping is also biologically meaningful. In an independent experiment, a set of 25 genes grouped by the learned incidence, with no recorded link among its members, decreases in expression together when IRF4 is deleted (Section 3.8.3). This shows that the latent hyperedges capture functional gene relationships that the reference does not record.
Please refer to the blue-highlighted content in Section 3.3, Table 2 and Section 3.8.3 in the revised manuscript.
Q2. Why are the gains small on Specific, and why does GRASP trail some baselines in certain STRING settings?
R2. Thanks for your valuable question! Both come down to what the reference supplies for a held-out pair.
Specific. Specific is cell-type-matched, so the recorded network already supplies much of the context a propagation model uses: GRASP and GRANet are within 0.02 AUPRC of each other in every cell, and GRASP is still the highest or tied-highest method in six of seven (0.88 vs 0.89 on mESC). GRASP adds context beyond the recorded neighborhood, so its margin over GRANet grows as the reference becomes less matched to the cell: +0.005/+0.007 on Specific, +0.037/+0.019 on Non-Specific and +0.057/+0.046 on STRING (Appendix F.1).
STRING. GRASP has the highest mean AUPRC of the nine methods at both scopes (0.193/0.167 vs 0.183/0.140 for GENIE3 and 0.137/0.121 for GRANet), and trails another method by more than 0.01 in five of the fourteen settings:
On mHSC-L, each target is ranked against only 39 (40) TFs from 847 cells, rather than 264-414 TFs from 383-758 cells in the three settings above (Table E1); fewer competing regulators and more samples per regulator favour ranking from expression alone.
(Continued in Part 2/2.)