acceptodds
Under review as a conference paper at ICLR 2027

HYPE-DTI: VALIDATION-GUARDED KNOWLEDGE EXCHANGE FOR LOW-LABEL DRUG–TARGET INTERACTION PREDICTION

Abstract

Predicting drug–target interactions (DTIs) from few labeled examples requires models to combine molecular, sequence, and relational evidence without amplifying errors during cross-view exchange. Existing dual-view methods integrate these sources, but often couple information transfer with model updating and final fusion, making it difficult to reject unreliable transfers. We introduce Hype-DTI, a dual-expert framework that validates each stage of collaboration. A relation expert provides selected pseudo-labels to a hybrid self expert, the reverse-distillation update is retained only after validation, and the learned gate is deployed only when it outperforms the selected expert. Across three five-fold low-label benchmarks, with each fold containing 10 positive and 10 negative training pairs, Hype-DTI improves mean AUPR over MoseDTI by 6.10 points on AGO, 0.49 on Blocker, and 4.60 on E-. It also outperforms MoseDTI in every matched semi-inductive fold involving unseen drugs, targets, or both. On DrugCentral, Hype-DTI reaches AUPR, the highest score among evaluated methods that do not use additional task-related DTI labels. ConPLex reaches with labeled DUD-E pretraining, illustrating the benefit of stronger external supervision. These results support validation-guarded exchange in the evaluated low-label and semi-inductive DTI settings, while highlighting the effect of external supervision on cross-dataset performance.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Drug–target interaction prediction, Low-label learning, Multi-view learning, Knowledge graph embedding, Knowledge distillation, Mixture of experts

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_3tssOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Hype-DTI, a dual-expert framework for validation-guarded knowledge exchange in low-label drug–target interaction prediction. A relation expert provides selected pseudo-labels to a hybrid self expert, while validation controls whether reverse-distillation updates are retained and the learned gate is deployed. Experiments show improvements over MoseDTI on low-label and semi-inductive benchmarks, alongside competitive performance on DrugCentral.

Critical strengths
  1. Practical safeguards against negative transfer. The framework separates pseudo-label selection, distillation acceptance, and fusion deployment, allowing unreliable transfers or harmful updates to be rejected.
  2. Strong semi-inductive cold-start results. On AGO, Hype-DTI outperforms MoseDTI in all five matched folds for unseen drugs, unseen targets, and both unseen endpoints. Mean AUPR gains are 8.76, 6.02, and 9.38 percentage points, respectively. These results support its effectiveness when labeled training interactions involving test entities are unavailable.
  3. Consistent performance across evaluation regimes. The method improves mean AUPR on all three low-label benchmarks and performs competitively on DrugCentral, providing evidence beyond a single dataset or training regime.
  4. Transparent reporting of evaluation boundaries. The paper distinguishes semi-inductive evaluation from fully inductive evaluation, separates additional DUD-E supervision, and acknowledges fold dependence and audit limitations. The fixed-expert deployment ablation isolates one specific decision.
Critical weaknesses
  1. The ablation in Table 4 isolates only the gate-deployment decision. It does not separately assess forward pseudo-label filtering, reverse distillation, validation-based rollback, or the additional pretrained representations. Consequently, it remains unclear how much of the improvement comes from the proposed exchange safeguards, stronger representations, or simple validation-based expert selection.
  2. Each low-label fold contains only 40 validation pairs, reused for early stopping, distillation adaptation, checkpoint acceptance, anchor selection, and deployment authorization. Sensitivity to validation size and thresholds is not reported. The statistical justification for the binomial-style standard-error proxy used to weight AUPR and AUROC gains is also unclear.
  3. Knowledge-graph noise robustness is not established. The paper recognizes that graph context can be unreliable, but does not systematically examine incorrect triples, inconsistent relations, or missing edges. Prediction filtering and validation safeguards may limit error propagation, but their effectiveness against confidently incorrect graph-derived predictions remains untested.
Questions for authors
  1. Could you supplement Table 4 with separate ablations of forward pseudo-label filtering, reverse distillation, validation-based rollback, and the additional pretrained representations, while holding other components fixed? Please also compare with simply deploying the better validation-selected expert.
  2. Could you provide the exact graph and training manifest used for the pretrained KGE checkpoint, and clarify whether evaluation DTI labels were excluded? How does the pretraining graph correspond to the graph inspected in Appendix E?
Reviewer_LR7TOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Hype-DTI combines a biomedical-graph relation expert with a molecular/protein “self” expert for drug–target interaction prediction with few labeled pairs. It makes three decisions explicit: which relation predictions become pseudo-labels for the self expert, whether reverse distillation into the relation expert is retained, and whether a learned gate is deployed instead of a validation-selected single expert (Sections 3.4–3.6). The principal experiments use five folds for each of AGO, Blocker, and E-, with 10 positive and 10 negative training pairs and 20 of each class for validation per fold. Mean test AUPR exceeds MoseDTI by 6.10, 0.49, and 4.60 percentage points, respectively (Table 1). Hype-DTI also improves AUPR in every matched AGO semi-inductive fold involving unseen drugs, targets, or both (Table 2; Appendix Table 6), and reaches 86.78% AUPR on DrugCentral, versus 84.79% for MoseDTI under the paper's matched-supervision grouping (Table 3). The only controlled component ablation holds the experts fixed and compares validation-authorized deployment with always deploying the gate (Table 4).

Critical strengths
  1. Consistent gains in the harder evaluated settings. The AGO and E- low-label improvements over MoseDTI are 6.10 and 4.60 AUPR points, and all 15 matched semi-inductive AGO fold comparisons favor Hype-DTI (Tables 1–2; Appendix Table 6). This supports the practical value of the overall system beyond a single split. The Blocker gain is much smaller and should be interpreted separately.
  2. A useful, directly tested deployment safeguard. With the same archived expert checkpoints, validation-authorized deployment improves mean test AUPR over always using the gate by 1.77 points on AGO and 4.25 points on E-, while tying on Blocker; the gate is used in 9 of 15 low-label folds (Table 4; Appendix Table 5). The authors also disclose high cross-fold test overlap and the limits of their leakage audit (Appendices D–E), which helps delimit the evidence.
Critical weaknesses
  1. The main mechanism is not isolated. Hype-DTI changes the self representation, pseudo-label selection, reverse-distillation/rollback procedure, and gate training/deployment, but Table 4 isolates only the last deployment choice. There are no matched ablations of forward transfer, reverse transfer, or rollback using the same experts, splits, and training budget. Consequently, the observed gains establish the performance of the full pipeline but do not show that the proposed sequence of validation guards, rather than another changed component, causes those gains. This directly limits the paper's central methodological claim.
  2. Model selection and provenance leave important uncertainty. Each nominal “10-shot” fold also uses 40 labeled validation pairs for early stopping and several adaptive decisions (Section 4.1; Sections 3.5–3.6 and Appendix B). Reusing this small set across many rules may make the selected policy unstable; no separate selection-set or multi-seed sensitivity analysis is reported. The test sets overlap heavily across folds (Appendix D), and the Blocker improvement is only 0.49 AUPR points. Finally, Appendix E says the pretrained KGE checkpoint lacks a verified training manifest; the authors' post hoc audit found no leakage in inspected artifacts, but cannot establish the checkpoint's training lineage. These points reduce confidence in the breadth and robustness of the empirical conclusion without themselves demonstrating leakage.
Questions for authors
  1. Mechanism (could change my rating): Can you provide matched, stage-level ablations that keep the expert architectures and training budget fixed while removing (a) selective relation-to-self transfer, (b) reverse distillation with rollback, and (c) deployment authorization? At minimum, please report the selected expert/gate AUPR before and after each stage on AGO and E-, so the claimed contribution of each guard can be assessed.
  2. Selection robustness (could change my rating): How were the numerous thresholds and acceptance rules chosen, and were any benchmark test outcomes consulted during their design? Can you show sensitivity to validation-set resampling or independent seeds, especially for the 40-example validation sets and the small Blocker gain?
  3. Graph provenance (could change my rating): Can you supply a verifiable manifest for the TransE checkpoint's training corpus and commands, or another check that directly rules out task-related DTI labels in that pretraining corpus? Appendix E explicitly states that the current audit cannot establish this lineage.
Reviewer_PRnCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Hype-DTI investigates how to combine molecular structure, protein sequence, and knowledge graph information when drug–target interaction labels are extremely scarce, while avoiding the propagation of erroneous information between experts. The framework consists of two experts: a relation expert that predicts interactions using entity embeddings pretrained on DRKG, and a hybrid self expert that combines molecular graphs, protein sequences, and pretrained representations from ChemBERTa and ESM models. Its central design introduces three decisions into the collaboration pipeline. First, it selects high-confidence, entity-diverse pseudo-labels for relation-to-self transfer. It then performs reverse distillation, using validation performance to determine whether to retain the update or roll back to the previous checkpoint. Finally, it trains a gate and uses validation performance to decide whether to deploy the learned gate or fall back to a single expert.

Critical strengths

1. Clear motivation. The paper identifies three distinct risks in dual-expert collaboration: pseudo-labels may be incorrect, reverse updates may degrade the original expert, and fusion may underperform a single expert. Addressing these risks through pseudo-label selection, rollback, and deployment authorization, respectively, gives the method a clear design rationale and provides fallback options when collaboration is unsuccessful.

2. Consistent empirical gains across multiple evaluation settings. Compared with MoseDTI, Hype-DTI improves AUPR by 6.10, 0.49, and 4.60 percentage points on the three low-label datasets, respectively. Across the three semi-inductive settings involving unseen drugs, unseen targets, or both, it outperforms MoseDTI in all 15 matched folds. Improvements are also observed on DrugCentral.

Critical weaknesses

1. The core methodological contribution remains unclear. MoseDTI already explicitly adopts the same four-stage pipeline: training the relation expert, using its pseudo-labels to train the self expert, performing reverse updates to the relation expert, and finally training the gate for fusion. Therefore, the dual-expert architecture, bidirectional knowledge exchange, and staged training are not new contributions of this work. The identifiable changes in Hype-DTI mainly consist of stronger pretrained representations, modified pseudo-label selection and distillation objectives, gate training with both experts frozen, and validation-driven rollback and deployment rules. These modifications are insufficient to substantiate the paper’s central claim. Furthermore, consider the ablation results on E-, the full system improves over MoseDTI by 4.60 percentage points, of which 4.25 percentage points arise from allowing the system not to deploy the gate. Does this merely indicate that the learned gate reduces performance relative to the already-trained experts, while the fallback strategy simply reverses this performance loss?

2. Reliability depends heavily on a small, repeatedly reused validation set, and its stability under low-label conditions remains unverified. Each fold contains only 20 training examples but uses an additional 40 validation examples. The same validation set supports multiple decisions, including early stopping, distillation-weight adjustment, update rollback, reference expert selection, and final gate deployment authorization.

3. The relation expert’s pretraining supervision raises concerns about the fairness of the comparisons. The relation expert relies on frozen knowledge graph embeddings and serves as the initial teacher for the entire pseudo-labeling pipeline. Consequently, the information it was exposed to during pretraining directly affects how its low-label performance should be interpreted. However, Appendix E explicitly acknowledges that no verified KGE training manifest was provided, and inspecting the currently supplied graph does not establish that the pretrained checkpoint was trained on that same corpus. The cold-start and DrugCentral experiments were also not included in the audited reruns. (§3.2; Appendix E) These checks therefore do not establish that the upstream representations were never exposed to the relevant supervision.

Questions for authors

See Weaknesses.

Reviewer_5k32OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper describes a method to rank drug-target pairs for screening using scarce and noisy data on known drug-target interactions and a knowledge graph, drug molecular structure and target protein sequence. The method is an improvement of previous such method called MoseDTI and performs four stages, with the ability to "veto" the passage of information: A knowledge graph ("the relation expert") that trains on the twenty pairs, scores a random 2% of the drug-by-target grid, and proposes only very confident predictions as pseudo-labels; a pair-intrinsic model ("the self expert") that trains on the observed and pseudo labels (the latter weighed differently) with minimal fitting; the self expert teaches the relation expert where the former is confident and the latter is not, with an AUPR threshold; fusion, with specific thresholds, that can fall-back to a single expert.

Critical strengths

The governance of the whole pipeline is notable. The fact that there is a (size scaled) validation threshold that must be met or the entire transfer is rolled back is a useful concept with evidence that it works, and which should be adopted more broadly. Another good thing is the endpoint diversity cap, which limits the reliance on a small and well-connected set of observations for decisions.

Critical weaknesses

The decision machinery here relies on a very small set of samples, namely AUPR over 40 validation pairs. Such small numbers can be very noisy. The authors should show how this behaves with multiple random seeds. ConPLex with DUD-E pretraining (a simpler method), on DrugCentral (plenty of labelled data), beats this method (according to the authors' own reporting), so while this method introduces useful concepts, it is not clear whether they truly contribute to drug discovery. As the authors state, held-out entities may appear in other graph relations so we can't tell whether this method would perform well on truly novel drug-target interactions.

Questions for authors

Can you report results across multiple random seeds, and in particular how often the rollback decision and the gate authorisation decision flip between seeds? If the guard decisions are themselves unstable, the reported gains are harder to interpret.

The advantage over MoseDTI is much larger on the ten-shot benchmarks than on DrugCentral. Does the benefit of validation guarding shrink as labelled data grows, and if so, in what label regime does the added complexity pay for itself?

Is a fully inductive variant feasible, in which held-out drugs and targets are removed from the auxiliary non-DTI relations as well? How much of the cold-start gain would you expect to survive?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.