acceptodds
Under review as a conference paper at ICLR 2027

Confidence-Gated Mechanistic Bridge Retrieval for Biomedical Knowledge Graph Reasoning

Abstract

Biomedical knowledge-graph reasoning requires recovering intermediate mechanistic evidence rather than only connecting a source and target. Semantic guidance can improve graph retrieval, but its value is conditional: useful candidates may be missing, semantically plausible candidates may provide limited downstream retrieval value, and the guidance strength may vary across queries. We introduce Confidence Gated Mechanistic Bridge Retrieval, a graph-constrained framework that separates these challenges into candidate coverage, retrieval utility, and query-specific influence. The method constructs graph-linked Bridges from complementary biological views, estimates retrieval-aware reliability from downstream supervision, and adaptively controls their influence during graph valid search. On DrugMechDB, the full system improves Path Success from 27.02% to 39.52% over graph-only retrieval. More importantly, controlled analyses show that the improvement does not arise from a single adaptive component: candidate coverage, downstream Bridge utility, and query-specific influence constitute distinct failure modes that require separate interventions. These results support a failure-oriented view of semantic graph guidance, in which availability, downstream value, and query-specific influence should be modeled separately.An anonymous repository for code and artifacts is available at https://anonymous.4open.science/r/iclr2027-bridge/

Primary area:
learning on graphs and other geometries & topologies
Keywords:
knowledge graph reasoning, biomedical knowledge graphs, retrieval-augmented generation, graph retrieval, mechanistic reasoning, reliability-aware retrieval, adaptive retrieval

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_kKLGOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies the retrieval of intermediate mechanistic evidence in biomedical knowledge graphs when source and target are known. The paper proposed 3 stages: select candidate bridges from different biological views, estimating the utility probability of the bridges and select a subset, then control the bridge influence through a query specific gate. Lightweight supervised models are trained to estiamte the bridge utility probability and query-level reliability. The searching scenario is a beam search combining graph and bridge scores under a fixed a search budget.

On DrugMechDB, Table 1 reports Path Success increasing from 27.02% to 39.52% and Edge Recall from 33.97% to 59.46% over Graph Only. Further experiments cover candidate coverage, usefulness scoring, single-Bridge selection, and constant or shuffled gate controls.

Critical strengths

S1. The Stage 2 experiment provides useful evidence about candidate utility. In Table 2, the retrieval-aware scorer distinguishes useful Bridges within the clean subset, with AUROC/AUPRC of 0.815/0.655 compared with 0.515/0.257 for lexical similarity. The separate single-Bridge intervention also improves Edge Recall from 38.19% to 39.92% under matched retrieval conditions. These results support the value of learning downstream usefulness within the controlled setting.

S2. The main retrieval computation can be checked using the released artifacts. New graph searches reproduced the main DrugMechDB Clean results. Refitting the confidence and gate estimators from the released training tables also recovered the main retrieval outcomes.

Critical weaknesses

W1. The main experiment does not establish the complete candidate-construction and selection procedure.

W1.1 — The main results start from reference-aligned candidates, leaving the actual bridge selection process insufficiently validated. In the released clean test inputs, all 692 candidate bridges are nodes of their corresponding reference mechanisms, and every one of the 248 queries already has its intermediate answer entity among the candidates. These inputs allow a controlled study of how candidates guide search. However, without the reference-aligned clean candidates, whether the stage 1 can automatically select effective candidate is questionable.

W1.2 — The main comparison does not test the proposed stage 2 Bridge-selection rule. Section 4.3 describes greedy selection using c_h − μ Red(h,H) + β TypeBonus(h,H). However, every released validation/test candidate pool contains at most 3 distinct nodes, matching the Bridge budget of 3. The released main retrieval code keeps all of them, so Stage 2 does not filter the saved candidate sets.

The package's audit/DRUGMECHDB_SELECTOR_PROVENANCE.md explicitly states that the original DrugMechDB redundancy function, coefficient, and complete selector were not recovered.

Semantic+Fixed assigns each candidate confidence 1, while Utility+Fixed uses learned confidence. both use gate 0.3. This comparison tests confidence weighting, but the selected sets do not change.

W2. The experimental evidence does not sufficiently establish the contribution of the learned components in the main setting.

W2.1 — The Full results in Tables 1 and 4 are not sufficiently reconciled. Table 1 reports Full Edge Recall of 59.46%, whereas Table 4 reports 38.05% on the same held-out split. The released Clean Full row in final_continuous_by_condition.csv has mean gate 0.36035, the Full row in stage3_controls_summary.csv has mean gate 0.10917. Section 5.5 states that it fixes the selected Bridges to isolate semantic influence, but this does not explain the change in Full.

W2.2 — Most of the headline improvement is already obtained with fixed guidance from the supplied candidates. In Table 1, Semantic+Fixed reaches 39.11% Path Success and 58.56% Edge Recall, compared with 39.52% and 59.46% for Full.

W2.3 — Search-order robustness remains unclear. The released retriever consumes its 20-edge expansion budget in adjacency-list order, so changing the order of otherwise identical outgoing edges can change which paths are explored.

Questions for authors

Q1. Candidate construction and reference access (W1.1). What exact procedure produces the clean candidates used in Table 1, and which steps, if any, use reference mechanisms or intermediate answers? An evaluation on candidates generated from the query, known endpoints, and graph would help establish the contribution of Stage 1.

Q2. General selection and the saved candidate sets (W1.2). Are the saved at most three nodes the initial pools H0 or already selected sets H*? How do they relate to Figure 3? Please provide the original candidate pools, selection outputs, redundancy definition, coefficients, and executable selector. Which experiment tests the full selection rule against confidence-only selection when the pool exceeds the same Bridge budget?

Q3. Full configurations and Stage 3 attribution (W2.1–W2.2). Which candidate sets, confidence scores, query gates, controller parameters, and search settings produced Tables 1 and 4? Please explain the difference in Full performance and mean gate, provide the per-query records and generator for Table 4, and state how its configuration was selected. Do the constant, mean-matched, and shuffled-gate controls support the same conclusion when applied to the actual Table 1 Full configuration?

Q4. Search-order robustness (W2.3). What determines outgoing-edge order in the reported experiments? With models and predictions fixed, do the incremental gains persist across a small set of alternative orders or search budgets?

3 comments

Thank you for recognizing our Stage-II usefulness evidence and verifying the main retrieval results. Our answers follow.

Q1 / W1.1 — Candidates and reference access

Table 1 uses supplied reference-aligned Clean Bridges drawn from reference mechanisms, including intermediate answers, to isolate subsequent retrieval decisions under controlled candidate access. They are not automatic-proposer outputs; our framework also supports automatic generation.

The proposer explores two forward rounds from the source, then two reverse rounds from the target, sharing a visited set. It excludes endpoints, groups candidates into molecular/cellular/system views by entity type, and deduplicates canonical IDs. A fixed rule combines lexical relevance, endpoint adjacency, anti-hub weighting and relation quality. Inputs are known endpoints/names, query intent and the fixed graph, excluding held-out intermediate answers and query-specific reference membership.

Before budget truncation, useful-candidate coverage rises from 40.32% for the strongest view to 61.29% for the union. Automatic Top3 with equal confidence and g=0.3 improves PS/ER over Graph Only by 8.06/4.76 percentage points (paired 95% CIs [4.40,12.11]/[2.88,6.84]). Sections 4.2 and 5.1–5.3 clarify both settings.

Q2 / W1.2 — Saved sets and greedy selection

Saved ≤3-node sets are supplied retriever inputs, not complete-greedy outputs. Table 1 Semantic+Fixed versus Utility+Fixed tests confidence weighting; Figure 3 describes larger automatic pools.

Parts of the original selector implementation could not be recovered. We reimplemented the stated rule with explicit operational definitions and evaluated it on automatic pools; this is not recovery of the historical implementation.

Starting empty, greedy maximizes Gain=c−μRed+βTypeBonus until K=3 or pool exhaustion. Red is maximum one-hop undirected-neighbor Jaccard with selected nodes (zero when empty); TypeBonus indicates a new entity type; ties use ascending node ID. Validation fixes C=0.1, μ=0.5, β=0.1. Both strategies share confidence scores, K=3, g=0.3 and search budgets. Of 248 pools, 239 exceed K; greedy changes 101 subsets versus confidence-only.

SelectionPS (%)ER (%)
Confidence-only32.6640.87
Complete greedy33.0641.77

Both means improve numerically; utility differs by +0.00554 (paired 95% CI [−0.01420,0.02451]), without establishing stable superiority. Complete-selector results (§5.4, Table 4, Appendix B) use this implementation on previously inspected data. Its code, automatic pools, outputs and traces are released in the anonymous repository.

Q3 / W2.1–W2.2 — Configurations and attribution

Table 1 evaluates the overall framework; original Table 4 intended to diagnose guidance control, but its Full configuration (mean gate 0.10917) was not aligned with Table 1. Revised Table 5 corrects this.

Controls now share Table 1 Full's Clean candidates, confidences, query features, frozen predictions and search. All 248 Full outcomes match: PS/ER=39.52/59.46%, mean g=0.36035.

The gate is g=0.5√p_rel. A 42-setting grid over five validation candidate conditions selected g_max=0.5, γ=0.5 using utility, with nearest-0.05 cached gates for selection and continuous test gates. Clean validation utility selects constant 0.45; validation gates give mean constant 0.35877. Five hundred global shuffles preserve the test-gate distribution. Search fixes beam=2, hops≤8, nodes≤14, edge considerations≤20 and original adjacency order.

GuidancePS (%)ER (%)Utility
Validation-selected constant39.1159.740.4530
Validation-mean constant39.1159.540.4524
Shuffled (500-run mean)39.2459.060.4518
Full39.5259.460.4550

Full has highest mean utility, without stable independent gate superiority. Full-minus-validation-mean paired 95% CIs are ER [−0.2490,0.0000] points and utility [−0.00072,0.00861].

For W2.2, Bridges supply most improvement; learned components refine their use. Full-minus-Semantic+Fixed ER is +0.90 points [0.23,1.69], a combined weighting/control increment, not gate alone. Results are in §5.2/5.5, Table 5 and Appendix B; code, generator, validation configurations, per-query records and assignments are in the anonymous repository.

Q4 / W2.3 — Search-order robustness

Outgoing edges follow graph insertion order; duplicate (neighbor,relation) pairs retain first occurrence. The 20-edge budget follows this order. With models, predictions and budgets fixed, mean gains remain positive across original, sorted and ten seeded random orders:

ComparisonPS gain (pp)ER gain (pp)
Full–Graph Only9.68–13.7122.91–25.49
Full–Semantic+Fixed0.40–0.810.054–0.901

We retain the original order for main results (§5.5, Appendix B).

Thank you again for your feedback.

Thank you for the detailed response and additional experiments. Revised Table 5 addresses my concern about aligning the gate controls with Table 1. However, I still have concerns as follows.

W1. Table 1 relies on supplied Clean candidates that already contain the intermediate answers, and their original construction remains incompletely documented.

The response confirms that these candidates were drawn from reference mechanisms, and Section 5.1 and Appendix B state that the intermediate answer is present in every test query's candidate pool. Their information source is therefore clear. Table 1 supplies reference-answer information to the retriever, so its large gains remain conditional on that access and do not establish the effectiveness of automatic candidate construction.

Appendix B also states that the historical candidate writer has not been authenticated. Please specify the original rule used to form each saved candidate set of at most three nodes, or identify which parts of that construction cannot now be verified. This matters for understanding what the main experiment actually evaluates.

W2. The reported DrugMechDB experiments do not establish the advantage of the complete three-stage combination under a matched comparison.

Section 4.2 describes multi-view candidate construction and specifies an automatic proposer using the query endpoints and graph. Sections 4.3–4.4 describe the complete greedy rule and query-specific influence control. My comparison of that design and the reported configurations is below. All table numbers refer to the revised manuscript.

Source and configurationStage I / candidate sourceStage IIStage III
Proposed method — Sections 4.2–4.4Multi-view candidate constructionRetrieval-aware Bridge reliability and the complete greedy ruleQuery-specific influence control
Table 1 — Full (Section 5.2, p. 7)Supplied reference-aligned Clean candidatesLearned confidences; all candidates retained without subset selectionContinuous gate
Appendix B — Full on fixed automatic proposal Top3 (pp. 15–16)Automatic poolsTop3 selected by the proposer and assigned learned confidences; no complete greedy ruleContinuous gate
Table 4 — Complete greedy (Section 5.4, p. 9)Automatic poolsLearned confidences and the reconstructed complete greedy ruleFixed gate = 0.3

Table 4 is titled “Automatic-pool selection (K = 3, g = 0.3)”. The automatic-proposal Full result is reported separately in the Appendix B paragraph “Matched learned-path baseline”; it is not a row of Table 4.

Table 1's pools contain at most three distinct nodes, matching the selection budget, so this experiment also does not test nontrivial subset selection.

The automatic-proposal experiment in Section 5.3, also reported as Proposal Top3 in Table 4, does show a real benefit: guidance with every Bridge confidence set to 1 and g fixed at 0.3 improves PS/ER from 27.02/33.97% to 35.08/38.73%. This supports the usefulness of automatically proposed Bridges. However, the automatic-proposal Full result and the Complete greedy result use the two different configurations shown above. A matched evaluation establishing the advantage of automatic proposal + complete greedy selection + query-specific influence control over simpler alternatives remains missing from the evidence discussed here.

W3. The complete Stage-II selector and Stage-III control have not demonstrated stable incremental gains over the matched simpler alternatives.

  • Stage II — Table 4, Complete greedy versus Confidence-only: utility improves by 0.00554, with a paired 95% CI of [−0.01420, 0.02451]. This does not establish a stable advantage of the complete selection rule.
  • Stage III — Table 5, Full versus Validation-mean constant: utility improves by 0.00258, with a paired 95% CI of [−0.00072, 0.00861]. This likewise does not establish a stable incremental advantage of query-specific influence control. These controls use the actual Table 1 Full configuration.

These intervals do not prove that the components have zero effect; they leave their incremental advantage over these simpler controls uncertain.

I recognize the positive results for retrieval-aware Bridge reliability in Table 2 and the small combined gain from learned confidences and the continuous gate in the Clean setting: Table 1 Full exceeds Semantic+Fixed by 0.90 ER percentage points, with a 95% CI of [0.23, 1.69] (Section 5.2). Nevertheless, most of the main-table improvement is already obtained with fixed guidance from the supplied candidates, and the additional experiments do not convincingly establish the value of the complete proposed combination.

I find their methodological significance and the demonstrated benefit of the full framework insufficient for acceptance. I therefore maintain my recommendation of clear rejection.

Thank you for your continued engagement. We address W1 first, then W2 and W3 together: the framework's design and the evidence for its individual decisions.

W1. Clean construction and verification scope

Table 1 supplies reference-aligned node identities, including the intermediate answer, but no connecting path or relation sequence. Its gains are conditional on these cues; comparisons sharing them examine their use. Automatic construction and nontrivial subset selection are evaluated separately.

The executable Clean rule builds a global entity-type map in reference-file order, retaining the last label for each canonical node ID. For each query, traverse its mechanism's nodes in stored order, exclude known endpoints and deduplicate IDs. Insert the reference answer first, preferentially add unrepresented types, then fill in record order to three base nodes. Append the first eligible node of the answer's type if absent.

This reproduces all 1,354 TRAIN, 239 VAL and 248 TEST lists, including order, in both saved tables: 3,682 exact matches. Five TRAIN lists have four nodes; all VAL/TEST lists have at most three. The saved inputs are exactly reproducible; historical writer identity and execution traces remain unauthenticated. The automatic proposer uses the query, known endpoints and graph, without query-specific answers or reference membership.

W2-W3. Research questions, matched evidence and contribution

We distinguish two questions: what each retrieval decision does under controlled conditions, and whether the complete combination stably outperforms simpler alternatives. Our controlled design addresses the former; the latter is a stronger claim assessed separately. An unestablished overall advantage does not replace evaluation of the individual research questions.

The experiments have explicit, different purposes. Section 5.3 tests whether automatically proposed Bridges improve retrieval. Table 2 tests whether downstream usefulness is distinguishable from semantic similarity. Table 4 holds the automatic pool, confidence scores, K and guidance strength fixed to test the complete selection rule. Table 5 holds candidates and confidences fixed to examine query-specific control. Table 1 studies weighting/control under supplied Clean access. These configurations isolate decisions rather than treating every row called "Full" as the same complete pipeline. We do not add gains across them to infer a combined advantage.

The supported findings are substantive. As you recognize, automatic Top3 guidance improves PS/ER from 27.02/33.97% to 35.08/38.73%, without query-specific reference access. Table 2's Clean usefulness AUROC is .815 versus lexical .515; its matched single-Bridge evaluation raises ER from 38.19% to 39.92%. This distinguishes recovering a plausible candidate from identifying one useful to budgeted evidence retrieval. Table 1 Full exceeds Semantic+Fixed by .90 ER points, with a paired 95% CI [.23,1.69]; this is a combined weighting/control increment, with its scope stated explicitly.

Selection and control require more specific conclusions. Your cited utility intervals, complete greedy versus confidence-only +.00554 [-.01420,.02451] and Clean Full versus the validation-mean constant +.00258 [-.00072,.00861], do not establish stable independent superiority. We retain that qualification. Within the selector, however, the matched TypeBonus ablation gives +2.02 PS points [.40,4.02] and +.01588 utility [.00375,.03138]. These pointwise, unadjusted intervals support type complementarity in that configuration; they do not establish redundancy's independent benefit or replace the complete-rule comparison. Likewise, shuffled-gate controls and positive/negative cases characterize sensitivity to the query-strength correspondence, while strong constants delimit the independent gate claim.

The complete automatic matched evaluation also separates overall performance from these component findings. It does not establish stable overall superiority, and its query gate matches the validation-selected constant in per-query PS/ER/utility. We therefore make no claim that every proposed decision is independently necessary for the largest reported gain. That limitation is explicit; it does not remove the positive evidence for automatic Bridge guidance, retrieval-aware usefulness estimation, or the value of identifying where additional selection/control helps and where simpler choices suffice.

We ask that the paper be evaluated on its documented contribution: mechanism-evidence retrieval, the separation of availability, utility and influence, and controlled evidence of their roles and limits. Consistent full selector/gate superiority is a separate claim, not a substitute for assessing these findings.

We hope these clarifications address your concerns and make the contribution and its evidence clear. Thank you again for your time and consideration.

Reviewer_V6SnOpenReview ↗
Rating
3
Confidence
1(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how intermediate biomedical entities, called Bridges, can guide path retrieval between known endpoints in a fixed knowledge graph. It separates three decisions: obtaining candidate Bridges from different biological views, estimating their usefulness to the retriever, and choosing how strongly they should influence search for each query. The main experiments on DrugMechDB use controlled interventions to study these decisions. Additional experiments examine reliability estimation on EpiKG and transfer to other retrieval settings.

Critical strengths

The paper makes a useful distinction between an intermediate entity being semantically plausible and being helpful to a constrained retriever. Table 2 tests this within the Clean candidate set, where retrieval-aware reliability reaches an AUROC of 0.815 compared with 0.515 for lexical similarity. This is more informative than a comparison that only separates plausible candidates from obviously corrupted ones. It supports the claim that downstream usefulness deserves its own estimation step.

The query-specific gate experiment is also carefully controlled. Table 4 fixes the selected Bridges and search conditions, then compares the learned gates with a nearly mean-matched constant and shuffled gate assignments. Edge recall is 38.05 for Full, compared with 37.29 for the constant and 37.18 for shuffled gates. The 500 shuffled assignments preserve the gate distribution, making this a useful test of whether matching guidance strength to the query matters. The effect is modest, but it directly addresses the proposed mechanism.

Critical weaknesses

The construction of the initial candidates is insufficiently clear for assessing the complete pipeline. Section 3 defines Clean Bridges as reference-aligned, the Stage-I analysis uses these candidates, and the overall results in Table 1 are also tied to the Clean condition. Meanwhile, the actual view-specific proposer is only described abstractly. This is a reasonable diagnostic setting, but it leaves open whether a practical proposer can supply the candidates on which the later stages depend. I therefore read the strongest evidence as conditional on candidate availability.

The additional gains from the learned components are small in the overall retrieval experiment. Fixed semantic guidance already reaches 58.56 edge recall, compared with 59.46 for Full, and the path-success difference is 0.41 percentage points (Table 1). This is consistent with the paper’s emphasis on understanding separate retrieval decisions, but it limits the strength of a practical performance claim. Paired uncertainty for these comparisons would help, especially since the evaluation section mentions bootstrap intervals but the main DrugMechDB tables do not report them.

Questions for authors
  1. Are the candidates in Table 1 constructed using reference mechanisms? Please specify the actual proposer, the information available to it, and which results use reference-aligned candidates.
  2. How are DrugMechDB queries divided into training, validation, and test sets? Are all candidate and corruption variants of the same underlying query kept in the same split?
  3. Please clarify the candidate and selection conditions that differ between Tables 1 and 4. Can you also report a paired confidence interval for Full versus the mean-matched constant in Table 4?
2 comments

Thank you for recognizing the value of distinguishing semantic plausibility from downstream retrieval usefulness, and for your positive assessment of our controlled query-specific gating experiment. We address your questions below.

Q1 / W1 — Candidate construction and experimental settings.

Yes, Table 1 uses reference-aligned Clean Bridges to isolate the effects of subsequent retrieval decisions under controlled candidate access. However, our framework does not require reference-aligned candidates and supports fully automatic Bridge generation. The diagnostic experiments in Section 3 start from reference-aligned Clean Bridges and introduce controlled corruptions; the controlled Stage-I view partition also uses reference-aligned candidates. Table 2 uses supplied Clean candidates and their corruption variants, and the revised Table 5 uses the Table 1 Clean configuration for gate controls. Section 5.3 evaluates automatically generated candidates.

Our automatic proposer performs bidirectional graph exploration (two forward rounds from the source and two reverse rounds from the target), constructs candidates across molecular, cellular, and system-level views, and ranks them using lexical relevance and local graph features. It uses only the known endpoints and their names, query intent, and the fixed knowledge graph, without using held-out intermediate answers or query-specific reference membership.

With automatically generated Top3 Bridges and fixed equal-confidence guidance (g=0.3), PS and ER improve by 8.06 and 4.76 percentage points over Graph Only, respectively. We have clarified the proposer and distinguished the two experimental settings in Sections 4.2 and 5.1–5.3.

Q2 — Data splits and candidate/corruption variants.

No variants derived from the same mechanism record cross the training, validation, or test splits. DrugMechDB is split at the mechanism-record level, ensuring that all task, candidate, and corruption variants of each record remain in the same split.

To verify this, we conducted a data-split audit covering the released Bridge and gate training and prediction tables. The resulting split statistics are as follows:

SplitTask queriesMechanism records
TRAIN1,3541,120
VAL239202
TEST248218

The audit confirms zero split-assignment mismatches. Different mechanism records may share question texts or endpoint pairs, with nine test queries sharing endpoint pairs with TRAIN or VAL. We have clarified the splitting protocol and verification results in Section 5.1 and Appendix B.

Q3 — Experimental settings and paired controller uncertainty.

Table 1 evaluates the overall framework under controlled candidate access, whereas Table 4 was intended to diagnose query-specific guidance control. However, the originally reported Table 4 Full configuration was not aligned with Table 1 Full. We have corrected this mismatch in revised Table 5.

To facilitate a direct comparison and assess the contribution of query-specific control, we have re-evaluated the gate controls using exactly the same candidate set, Bridge confidences, frozen predictions, and search settings as Table 1 Full.

Full minus the validation-mean constant gives an ER difference of −0.0806 percentage points (paired 95% CI [−0.2490, 0.0000]) and a retrieval-utility difference of +0.00258 (paired 95% CI [−0.00072, 0.00861]). These intervals do not establish a stable overall advantage over this constant. The matched results appear in revised Table 5, with paired intervals reported in Section 5.5 and Appendix B.

Thank you again for your helpful feedback on the experimental settings and evaluation protocol.

Thank you for your responses, I will maintain my original score.

Reviewer_9P9qOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a graph-constrained retrieval framework for recovering intermediate mechanistic evidence in biomedical knowledge graphs. Its stages are multi-view candidate Mechanistic Bridge construction, retrieval-aware estimation and selection of Bridges, and confidence-gated control of their influence during path search. Semantic guidance prioritizes existing graph paths without modifying connectivity. Experiments primarily use DrugMechDB-based tasks and supplementary knowledge graphs. The full pipeline improves Path Success from 27.02% to 39.52% and Edge Recall from 33.97% to 59.46% against graph-only retrieval. Controlled experiments isolate candidate coverage, Bridge utility and query-specific influence.

Critical strengths

S1. Meaningful mechanistic-evidence problem. The focus on recovering meaningful intermediate evidence rather than simply connecting entities is important. Separating candidate availability, downstream utility and query-specific guidance strength is a useful conceptual decomposition.

S2. Carefully structured diagnostic experiments. The paper separates full-pipeline comparisons from matched component interventions, including Bridge-quality corruption, utility-based selection and shuffled-gate controls.

Critical weaknesses

W1. Novelty and contribution of learned components remain unclear.

The framework combines multi-view generation, supervised relevance estimation, diversity-aware greedy selection and query-specific interpolation of graph and semantic scores (Sections 4.2–4.4). These resemble established ranking and adaptive retrieval ideas. A close biomedical baseline is -- Joy and Su, “Federated Knowledge Retrieval Elevates Large Language Model Performance on Biomedical Benchmarks” (2026), introducing BTE-RAG/DrugMechDB-related tasks;

Another work: -- Asai et al., “Self-RAG: Learning to Retrieve, Generate, and Critique Through Self-Reflection” (ICLR 2024), is relevant to adaptive retrieval.

-- Table 1 (p. 7) shows Semantic+Fixed at 39.11% Path Success and 58.56% Edge Recall versus 39.52% and 59.46% for Full: most of the headline improvement is from introducing semantic Bridges, with modest incremental gains from learned stages.

W2. There is a limited scope of evidence that utility and confidence models improve realistic retrieval.

Bridge confidence is estimated with a logistic regression on TF-IDF and graph features (Section 4.3); the gate uses another logistic model (Section 4.4). In Table 2, selection improves Edge Recall but slightly reduces Path Success relative to lexical similarity. In Table 4, query-specific control exceeds the mean-matched constant by only 0.76 Edge Recall points and shuffled gates by 0.87. The Stage-I controlled study uses held-out “Clean” reference-aligned candidates (Section 4.2, p. 5), useful diagnostically but different from realistic generation. Broader practical value and component complexity remain unresolved.

Questions for authors

Q1 (W1, Section 4.4 Tables 1 and 4). What is technically new in the gate relative to learned weighting and adaptive search, and how does it compare with a learned path scorer trained with matching supervision and budget?

Q2 (W2, Section 5.4, Table 2). Why does retrieval-aware selection improve Edge Recall but not Path Success relative to lexical similarity? Does this vary by Gene, Metabolite and Process task?

Q3 (W2, Section 5.5, Table 4). Could the authors report paired confidence intervals for small controller gains and characterize cases where query-specific gating is useful?

Q4 (W2, Sections 4.2 and 5.3). How much of the result uses reference-aligned Clean Bridges versus automatically generated candidates? How much gain survives imperfect generation?

1 comment

Thank you for recognizing our problem, conceptual decomposition, and diagnostic experiments as strengths. Our answers follow.

Q1 — Contribution

1. Contribution. Our contribution is a mechanism-oriented perspective on scientific graph retrieval: intermediate hypotheses actively guide evidence recovery rather than serving solely as retrieval outcomes. Biomedical retrieval must recover evidence explaining a relationship, beyond connecting entities.

Three linked components address coverage, utility, and influence. Multi-view construction broadens candidate access across molecular, cellular, and system levels. Retrieval-aware reliability uses downstream-benefit supervision; selection accounts for redundancy and type complementarity. Query-specific control regulates their collective influence. Utility estimation cannot recover missing candidates; coverage cannot distinguish useful hypotheses from distractors; useful hypotheses do not justify uniformly strong guidance. BTE-RAG retrieves mechanistic context for answer generation; Self-RAG uses reflection tokens for passage retrieval and evidence/generation evaluation. We guide graph-valid path retrieval.

2. Rationale and comparison. Unlike the matched path scorer, our gate preserves structural and mechanistic scoring functions and learns their relative influence at query level. This introduces a stronger structural inductive bias, potentially reducing overfitting when mechanism-level supervision is limited.

The TRAIN-only logistic path scorer shares the graph, supervision source, utility, frozen Bridge confidences, and search budgets; input representations and supervision granularity differ. On Clean, Full/scorer PS/ER are 39.52/59.46% versus 35.48/56.93%; paired differences are +4.03 [1.57,6.94] and +2.53 [0.62,4.46] percentage points (95% CIs). On automatic Top3, they are 34.68/39.12% versus 33.06/39.82%, with the utility CI spanning zero. Clean results favor our implementation over this numeric linear scorer without isolating its source or establishing independent gate superiority.

3. Table 1. We agree that Bridge guidance accounts for most improvement. Under supplied Clean access, confidence weighting and query-specific control refine an already effective guidance strategy; their modest increments are not the primary gain.

Q2 — ER–PS tradeoff

The tradeoff reflects prioritizing path completion versus recovering mechanistically informative evidence. Connecting endpoints need not recover evidence explaining their relationship. Under fixed budgets, prioritizing such evidence may improve ER without increasing PS. Our utility jointly considers both: M=0.7 PS+0.3 ER. Among 1,313 positive automatic singleton TRAIN labels, 973 improve ER only and 340 improve PS.

The K=1 comparison covers 248 queries under five candidate conditions (1,240 query-condition cases):

TaskPS lexical→reliability (%)ER lexical→reliability (%)
All29.92→29.1938.19→39.92
Gene31.87→31.3836.46→38.82
Metabolite25.71→25.7133.47→36.78
Process28.46→27.3141.17→41.87

Mean ER increases by 0.69–3.31 points, with mean PS changes of −1.15 to 0 points. Metabolite has the largest ER gain with unchanged mean PS. This is consistent with recovering richer mechanistic evidence with small observed path-completion changes (§5.4, Appendix B).

Q3 — Uncertainty and cases

For revised Table 5’s actual Table 1 Full configuration, Full minus the validation-mean constant gives PS +0.40 [0.00,1.24] points, ER −0.0806 [−0.2490,0.0000] points, and utility +0.00258 [−0.00072,0.00861]. These paired 95% CIs use 5,000 mechanism-record bootstrap draws, retaining task variants together; they do not establish stable overall superiority.

As motivated in Q1, Bridge usefulness does not determine guidance strength. With Bridges/confidences fixed, Tolbutamide–type-2-diabetes Gene uses g=0.28795 instead of the validation-mean constant 0.35877, redirecting a potassium-channel-related path toward insulin secretion. PS changes 0→1; ER remains 1/7. This illustrates how separate influence control can improve completion (§5.5, Appendix B).

Q4 — Automatic versus Clean candidates

Gains survive automatic proposal. Automatic Top3 with fixed equal-confidence guidance improves PS/ER by 8.06/4.76 points over Graph Only—approximately 64.5%/18.7% of Clean Full gains, although guidance configurations differ.

Table 1 uses reference-aligned Clean Bridges to isolate guidance under controlled access; §5.3 separately evaluates automatic candidates constructed without held-out answers or query-specific reference membership.

Before budget truncation, useful-candidate coverage increases from 40.32% for the strongest view to 61.29% for the union. Automatic Top3 with equal confidence and g=0.3 reaches PS/ER 35.08/38.73% versus Graph Only 27.02/33.97%; paired gain CIs are [4.40,12.11]/[2.88,6.84] points.

Thank you again.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.