acceptodds
Under review as a conference paper at ICLR 2027

GRASP: Adaptive Context with Regulatory Partner Supervision for GRN Inference

Abstract

Supervised gene regulatory network inference from single cell transcriptomic data seeks to rank unobserved transcription factor-target links by combining expression profiles with partial interaction references. Graph-based approaches exploit such references to provide relational context for prediction. However, local message passing remains bounded by incomplete reference topology. Moreover, same gene contrastive alignment does not explicitly exploit training positive TF-target partners when defining positive comparisons. To address these issues, we propose a framework named GRASP, which couples adaptive context construction with partner-based contrastive supervision for GRN inference. Technically, GRASP uses expression guided edge retention to adapt local propagation to the measured cellular context, and further constructs a teacher induced higher-order view to provide context beyond recorded graph neighborhoods. To exploit relational supervision already available in training interactions, neighborhood contrastive learning augments same gene correspondence with fixed training partners as positives within and across the two views, while a role-specific decoder produces ordered TF-target scores. We additionally employ pair-disjoint evaluation to prevent held out gene-pair connectivity from entering the training graph. Across 42 evaluation settings, GRASP achieves the highest average AUPRC among eight evaluated baselines. Code is available at https://anonymous.4open.science/r/GRASP-302C/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Gene regulatory network inference, Single-cell RNA sequencing, Graph neural networks, Graph contrastive learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_s7TrOpenReview ↗
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
  1. The proposed components are not sufficiently compared with simpler alternative designs, making their specific methodological contributions less clear.
  2. The performance improvements vary across different reference networks and cell types, with relatively small gains or inferior results in some settings.
  3. 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
  1. 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?
  2. Could the authors explain why GRASP achieves relatively small improvements on the Specific reference networks and underperforms some baselines in certain STRING settings?
  3. 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

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 ZhZ_h 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):

Higher-order context (STRING/TF+1000)Mean test AUPRCDrop vs GRASP
Learned latent hypergraph (GRASP)0.1668-
Global mean of teacher representations0.1504−9.8%
No higher-order view (Table 2)0.1433−14.1%

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:

SettingsReason
hESC, hHEP, mDC at TF+1000All nine methods score 0.01-0.05 AUPRC, close to the 0.02-0.03 chance level (Table E2)
mHSC-L at both scopesSparsest STRING reference (values are for TF+500, with TF+1000 in parentheses): 137 (154) links over 70 (81) targets, about 77 (86) training positives per split (Table E1); the label-free GENIE3 and GRNBoost2 lead, and GRASP is first or second among methods trained on reference labels

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.)

(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).

ReferenceScopeProtocol gap (AUPRC)Composition ΔAUPRCInjected (% of test positives)Connectivity ΔAUPRCConnectivity ΔAUROC
NSTF+500+0.075−0.01944-28+0.0110+0.0550
NSTF+1000+0.052−0.00154-25+0.0177+0.0365
STRINGTF+500+0.230+0.011022-49+0.1268+0.1787
STRINGTF+1000+0.271−0.024513-42+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.

Reviewer_JbdLOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GRASP, a graph-based framework for supervised gene regulatory network inference from single-cell expression data and partial regulatory interaction references. GRASP integrates three components: expression-guided edge retention for adaptive local graph construction, a momentum teacher-based higher-order representation to capture dependencies beyond observed graph neighborhoods, and neighborhood contrastive learning that incorporates known TF-target relationships as positive pairs. A role-specific decoder predicts directed regulatory interactions. The method is evaluated across 42 settings involving seven cell types, three reference networks, and two gene set sizes. GRASP achieves the highest average AUPRC among eight baselines, with factorial ablations and additional experiments supporting its effectiveness and robustness.

Critical strengths
  1. Task-aligned integration of higher-order context and relational contrastive supervision.

The framework addresses two relevant limitations of existing graph-based GRN inference methods: restricted propagation over incomplete reference graphs and contrastive objectives that overlook known regulatory partners. Combining teacher-induced higher-order representations with partner-based contrastive supervision is well motivated for predicting missing regulatory interactions. The factorial ablation study further demonstrates complementary benefits from these components, supporting the effectiveness of the overall design.

  1. Extensive evaluation with explicit consideration of information leakage.

The paper evaluates GRASP across diverse cell types, reference networks, and gene scopes, with additional analyses of reference corruption, cross-cell transfer, and external annotation support. Particularly valuable is the pair-disjoint evaluation protocol, which prevents reverse orientations of held-out regulatory pairs from entering the training graph. The authors demonstrate that conventional edge splitting can substantially inflate performance, making this evaluation design an important contribution to the reliability of supervised GRN inference.

Critical weaknesses
  1. Limited evidence that the inferred regulatory relationships reflect biological regulation rather than reference-network completion.

The evaluation primarily measures recovery of held-out interactions from existing reference networks, including STRING functional interactions that do not necessarily represent direct transcriptional regulation. Although the authors provide external annotation support, agreement with curated interaction databases does not establish whether newly predicted TF-target relationships are functionally active in the corresponding cellular contexts. Validation using perturbation-derived regulatory effects or other independent functional evidence would be important for demonstrating that GRASP learns biologically meaningful regulatory relationships beyond reference-network completion.

  1. Limited validation of the biological relevance of the learned higher-order context.

The teacher-induced higher-order representation enables information exchange between genes without direct edges in the reference network. However, these latent connections are learned primarily through the existing interaction topology and contrastive supervision, without explicit evidence that they correspond to meaningful higher-order regulatory relationships. While the ablation results demonstrate predictive benefits, they do not establish whether the learned context recovers missing regulatory dependencies or primarily provides additional representation smoothing. Analyzing the inferred higher-order gene associations against independent regulatory evidence would strengthen the methodological contribution.

Questions for authors

See the Weakness section

1 comment

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. Agreement with reference networks does not show that predicted TF-target links are functionally active; validation with perturbation or other independent functional evidence is needed.

R1. Thanks for your valuable suggestion! We added independent TF induction validation to test whether GRASP prioritizes functionally responsive targets outside the benchmark reference. Models are frozen and the induction data are not used for training or selection; candidates exclude every training, validation and reference pair, and the top 20% predicted targets are compared with matched backgrounds.

The results in the two independent studies are (gain over the matched background in percentile points; BH corrected q):

StudyComparisonsSignificant enrichmentGain (pp)q
GSE14559, individual TF induction, 48h32 evaluable TFs (1,900 tests)8 of 32 TFs4.76-9.15≤ 0.0220
GSE306024, GATA3 induction, RNA-seq4 splits × 4 clones × 3 times = 4848 of 484.34-12.09≤ 0.0039

All eight TFs in GSE14559 lack outgoing positive TF-target labels in training, and GATA3 is one of them. In GSE306024, GRASP also exceeds a ranking based on training expression correlation in all 48 comparisons, significantly in 44.

For example, GATA6, PDGFRA, SOX7 and FOXA2, none of them a recorded GATA3 target, rank in the top 20% in every split and increase after induction in every clone and time point. GRASP thus prioritizes targets that respond to perturbation of their predicted regulator beyond the links recorded in the reference, across two independent studies. Please refer to the blue-highlighted content in Section 3.8.2, Table 4 and Appendix K.2 in the revised manuscript.

Q2. Is there evidence that the learned higher-order associations reflect real regulatory dependencies rather than representation smoothing?

R2. Thanks for your valuable suggestion! We added an independent functional test of gene associations read out from the learned incidence representation S (gene-to-hyperedge weights, Eq. 2). Clustering S from frozen mDC models gives 27 gene clusters. A set of 25 genes, denoted M12, has no recorded connectivity: its members share no training positive link, no training neighbor and no reference link in either orientation, so the set is not a restatement of recorded neighborhoods.

In the independent bone marrow derived dendritic cell (BMDC) IRF4 conditional deletion data GSE52767 (three conditions, five control and five deletion samples each; background matched on training expression mean, variance, graph degree and TF identity), M12 is downregulated in all three conditions (Table 5, Appendix L):

ConditionAdjusted effect (log₂ ratio)Welch qExact assignment qSigned contrast (pp)Signed contrast q
Unstimulated−0.64920.01820.0353−25.1270.0035
LPS 6h−0.52570.00840.0353−21.5960.0070
LPS 24h−0.42690.02040.0353−23.1890.0022

All q values are corrected across the 2,974 tested gene sets. The effect survives removing the five MHC class II members. IRF4 is a transcription factor, so the joint downregulation of these genes under its deletion shows that the latent hyperedges group genes with a shared regulatory dependence that the reference does not connect.

The predictive gain is also not reproduced by uniform smoothing: replacing the learned hyperedges with a global mean recovers only part of the gain (mean test AUPRC on STRING/TF+1000; Section 3.3, Table 2):

Higher-order contextMean test AUPRC
Learned latent hyperedges (GRASP)0.1668
Global mean of teacher representations0.1504
No higher-order view0.1433

Please refer to the blue-highlighted content in Section 3.8.3, Table 5 and Appendix L in the revised manuscript.

Thanks again for your careful review and constructive suggestions. Please let us know if you have further questions.

Reviewer_VLghOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

GRASP is a supervised GRN link predictor combining (i) expression-guided edge dropout for the student GAT, (ii) a momentum-teacher hypergraph view that pools genes into K latent hyperedges to provide context beyond the recorded graph, and (iii) neighborhood contrastive learning that treats training TF-target partners as positives within and across views. A role-specific decoder gives directed scores. The paper also introduces a pair-disjoint split that keeps both orientations of a gene pair in the same partition, and shows that the legacy directed split leaks heavily under symmetrized message passing (STRING AUPRC drops from ~0.43 to ~0.18). Across 42 BEELINE-style settings GRASP has the best average AUPRC (0.35/0.34 vs 0.31 for GRANet).

Critical strengths
  1. Section 3.4 is the most useful part of the paper. Showing that the standard directed edge split in GENELink/GCLink-style evaluation leaks reverse orientations into the training graph, and that this inflates STRING results by 0.2+ AUPRC, is something the community needs to know, independent of the method.
  2. The evaluation is broad and the ablation is a full 2³ factorial on one reference, which is more careful than usual in this literature.
Critical weaknesses
  1. The method is an assembly of known parts (GCLink’s contrastive link prediction, BGRL-style momentum teacher, hypergraph pooling, NCLA-style neighbor positives). Theorem 1 and Proposition 1 are essentially bookkeeping. The gain on the Specific reference, where absolute performance is meaningful, is +0.005/+0.007 over GRANet. The large gains are on Non-Specific and STRING TF+1000, where most methods sit at 0.02–0.05 AUPRC and the ordering is close to noise.
  2. Some results are unexplained and make me worry about the protocol. GENIE3, an unsupervised tree method, beats every supervised model on STRING mHSC-L by a wide margin (0.55 vs 0.13) and wins or nearly wins several other STRING cells. If the supervised methods are losing to a method that never sees labels on those settings, something about those settings deserves discussion.
Questions for authors

-Were all eight baselines retrained under the pair-disjoint protocol, or are their numbers taken from the legacy protocol? Table 3 only shows GRASP under both.

-Why does GENIE3 dominate on STRING mHSC-L? Is this a label-sparsity artifact (137 positives)?

-The directionality analysis shows only 62–72% forward preference. How should a user interpret a directed score when roughly a third of annotated edges are scored backwards?

-Hyperparameter sweeps change validation AUPRC by ≤0.47% across eight parameters. Is that measured at the reference configuration only, or does the finding hold for the baselines too under the same sweep?

4 comments

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. The method assembles known parts, Theorem 1 and Proposition 1 are bookkeeping, the Specific gain over GRANet is only +0.005/+0.007, and the larger gains are where most methods score 0.02-0.05 AUPRC.

R1. Thanks for your valuable comments!

GRASP is designed around a central difficulty of GRN inference under an incomplete reference: many true regulatory links are missing, so genes need context beyond their recorded neighborhoods. The teacher induced higher-order view supplies this context through latent hyperedges that let genes without a training edge exchange information. Which genes these hyperedges group must in turn come from the task's own supervision. The training TF-target pairs are exactly such positive evidence, and neighborhood contrastive learning makes use of these positives by treating training partners as positives within and across the student and higher-order views. Because the cross-view terms pass through the higher-order representation, partner relations shape the hyperedge associations S (Section 2.3). An independent IRF4 deletion experiment confirms that the hyperedges learned in this way group functionally related genes that the reference does not connect (Section 3.8.3, Table 5).

The parts therefore depend on each other. The higher-order view gains three times as much when partner positives are present, and the gain requires the learned grouping rather than pooling alone (STRING/TF+1000, Section 3.3, Table 2):

Mean test AUPRCWithout partner positivesWith partner positives
Without higher-order view0.13470.1433
With learned higher-order view0.14240.1668
Gain from higher-order view+0.0077+0.0235
Learned hyperedges replaced by global mean-0.1504

Independent TF induction data further show that GRASP's top ranked targets outside the reference respond when the TF is induced (Section 3.8.2, Table 4).

Theorem 1 and Proposition 1 state the two properties this design depends on, which set it apart from local message passing. Theorem 1 shows that the higher-order view stays equivariant to gene ordering while letting a gene's output depend on genes it is not connected to, which propagation along training edges cannot do. Proposition 1 shows that training partners remain positives under every sampled student graph, so the supervision that shapes the hyperedges does not change with the sampled routes. In addition, Section 3.8.3 tests the first property empirically.

Regarding the gains, on Specific the strongest methods already reach 0.83-0.95 AUPRC in five of the seven cell types, which leaves little headroom; GRASP is still the highest or tied-highest method in six of the seven. On Non-Specific and STRING we agree that differences are small where methods score around 0.02-0.05. The average margins over GRANet, however, come mainly from the three mHSC cell types, where scores are well above that range, for example on mHSC-E (AUPRC, TF+500/TF+1000; split-level SDs 0.02-0.05, Table F2):

mHSC-EGRASPGRANet
Non-Specific0.18/0.150.07/0.06
STRING0.19/0.310.09/0.16

(Continued in Part 2/2.)

(Part 2/2, continued.)

Q2. Why does the unsupervised GENIE3 dominate STRING mHSC-L (0.55 vs 0.13)? Is this a label-sparsity artifact?

R2. Thanks for your valuable question! It is largely a label-sparsity effect. STRING mHSC-L is the smallest reference in the benchmark (values are for TF+500, with TF+1000 in parentheses): 137 (154) links across 70 (81) targets (Table E1), which leaves only about 77 (86) training positives in each split. GENIE3 and GRNBoost2 use no reference labels and are unaffected, while all seven label-trained methods score well below them, with GRASP first or second in that group (Table 1). The candidate pool also sets mHSC-L apart: each target is ranked against only 39 (40) TFs using 847 cells, whereas in hESC, hHEP and mDC it is 264-414 TFs using 383-758 cells (Table E1). With fewer competing regulators and more samples for each, ranking from expression alone is favoured.

As for the other STRING cells, GRASP is ahead of GENIE3 in 9 of the 14 STRING settings, by a wide margin in the following (AUPRC):

STRING settingGRASPGENIE3
mESC, TF+500/TF+10000.25/0.200.05/0.05
mHSC-E, TF+10000.310.12
mHSC-GM, TF+10000.440.25

Besides mHSC-L, GENIE3 leads only at TF+1000 on hESC, hHEP and mDC, where all nine methods score 0.01-0.05 AUPRC, close to the 0.02-0.03 chance level (Table E2). All nine methods share the same splits and held-out candidates, so Table 1 remains a like-for-like comparison.

Q3. Were all baselines retrained under the pair-disjoint protocol, or are their numbers taken from the legacy protocol?

R3. Thanks for your valuable question! All eight baselines are evaluated under the pair-disjoint protocol, and no Table 1 number is carried over from a legacy split. The six supervised baselines were trained on the same splits, candidates and pseudo-negatives as GRASP (Section 3.1; Appendix E.1). GENIE3 and GRNBoost2 use no reference labels and are evaluated on the same held-out candidates. Table 3 compares GRASP with itself under the two protocols.

Q4. How should a user interpret a directed score when the forward preference is only 62-72%?

R4. Thanks for your valuable question! A directed score expresses ranking support for an ordered TF-target candidate. For a pair where both orientations are eligible, users should compare the forward and reverse scores together with their margin and consistency across training repeats. A small margin or an unstable preferred orientation should be reported as unresolved direction. The score itself is not a calibrated probability that the orientation is correct.

The endpoint order evaluation compares singly annotated held out links whose two endpoints both qualify as TFs and targets. Forward preference is the fraction with s(u,v)>s(v,u); ties do not count as forward wins. The current rates are 68.1%/70.0% on hESC Specific-500/1000, with orientation AUROC 0.687/0.718, and 98.24%/97.18% on mESC, with orientation AUROC 0.974/0.941. The 62-72% in the question were our earlier single-split estimates; both hESC rates now average the five fixed splits. These measurements characterize endpoint order within this eligible subset.

The training role comparison identifies where the difference between cell types is concentrated. On Specific-1000, preference for links whose target never appears as a training regulator is 69.31% on hESC and 98.70% on mESC. When both endpoints appear in both training roles, the rates are 78.30% and 82.83%. The splits contain 4,727-4,728 training positives on hESC and 28,634-28,635 on mESC, supplying different amounts of directed supervision for the independent regulator and target projections. Where directed supervision is limited, as on hESC, about 30% of eligible links still score higher in the reverse orientation, so the directed score should be used there as weaker evidence of direction. Please refer to the blue-highlighted content in Section 3.6, Figure 3a and Appendix H in the revised manuscript.

Q5. Is the ≤0.47% hyperparameter sensitivity measured for GRASP only, or does it also hold for the baselines?

R5. Thanks for your valuable question! It is measured for GRASP only; the sweep was not run for the baselines. Each sweep varies one of GRASP's eight hyperparameters around its reference configuration, and the mean validation AUPRC changes by at most 0.47% in aggregate (Appendix I.2), so GRASP's results do not depend on a narrow choice of settings.

Section 3.5 of the revised manuscript now reads "Sweeps of GRASP's eight hyperparameters".

Thanks again for your careful review and constructive suggestions. Please let us know if you have further questions.

Thanks for the detailed response. R3 and R5 fully address my concerns, and the GENIE3 explanation in R2 is convincing — I’d ask that the label-sparsity limitation be stated explicitly in the paper rather than only here. The global-mean ablation and the IRF4/TF-induction checks strengthen the case that the components interact, though I’d still describe the Non-Specific/STRING margins as concentrated in the mHSC cell types rather than uniform across references. I’m keeping my weak accept and have updated the soundness score.

Dear Reviewer VLgh,

Thank you for your feedback and for updating the soundness score! Following your suggestion, the revised manuscript now states this limitation explicitly in Appendix F.1 (blue-highlighted). A supplied reference helps only when it provides enough training links. Where labels are scarce, as on STRING mHSC-L, methods that infer regulators from expression alone rank them better.

Please do not hesitate to reach out to us if you have any further questions.

Best regards,

the Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.