acceptodds
Under review as a conference paper at ICLR 2027

Generative Retrieval for Reverse Target Fishing with a Proteome-Scale Benchmark

Abstract

Reverse target fishing identifies protein targets for a query molecule and is central to drug discovery and repurposing. Discriminative approaches adapted from virtual screening face sparse, long-tailed associations and missing-not-at-random measurements: treating unrecorded pairs as negatives can suppress genuine targets. We reformulate the task as generative retrieval and introduce GenFish. Residual vector quantization converts frozen protein language model features into shared semantic identifiers. Given a molecular fingerprint, GenFish autoregressively models codes conditioned on the query and preceding code, ranking proteins by identifier likelihood without explicitly labeling unrecorded pairs as negatives. We introduce ProTargetBench, the first proteome-scale benchmark for reverse target fishing to jointly support measured-panel discrimination and full-catalogue recovery of experimentally recorded positives. Covering 20,431 UniProt-reviewed human proteins, it integrates curated bioactivity evidence and distinguishes experimental inactivity from unmeasured associations. On 198 held-out molecules, GenFish achieves higher mean performance across all three full-catalogue retrieval metrics than four discriminative neural baselines and three non-neural baselines, including ligand-based methods. A controlled positive-hiding experiment shows that targeted removal of false-negative competition improves contrastive retrieval beyond matched random masking; GenFish remains stronger under the same incomplete supervision. Pfam analysis shows that the semantic identifiers retain biologically meaningful structure. These results establish generative retrieval as a principled paradigm for reverse target fishing under incomplete bioactivity records.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Target fishing, Drug–target interaction prediction, Generative retrieval, Semantic identifiers, Residual vector quantization, Proteome-scale benchmark, Partially observed labels, Protein language models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_mkFeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes GenFish, a framework for reverse target fishing: given a small molecule, identify the protein targets that could potentially bind it. This work avoids using false-negative pairs, which may affect retrieval learning in common contrastive learning. Additionally, this paper introduces a dataset, ProTargetBench, to support this task. This paper conducts comprehensive experiments to prove the effectiveness of GenFish.

Critical strengths
  1. This paper focuses on a meaningful research topic. Reverse target fishing is relevant to drug discovery, and the issue of incomplete bioactivity labels and false-negative pairs is worth investigating.

  2. This paper introduces a dataset related to this task. ProTargetBench integrates multiple bioactivity databases and distinguishes measured negatives from unmeasured pairs, providing a useful setting for evaluating proteome-scale retrieval.

  3. GenFish achieves the best average performance across the retrieval metrics, particularly R@50. The authors also investigate false-negative competition, molecular similarity, and the biological relevance of semantic IDs.

Critical weaknesses
  1. The overall novelty contribution is limited. The core framework of GenFish is quite similar to TIGER (Recommender Systems with Generative Retrieval, NeurIPS 2023). Both use pretrained representations, residual quantization to construct semantic IDs, an additional token to resolve ID collisions, and ID prediction for retrieval. GenFish mainly adapts this framework to molecule-to-protein retrieval. Although it introduces a molecule-conditioned interaction module, the methodological novelty is still limited, and the contribution is not clearly validated through ablation studies.

  2. In Table 1, even the simple weighted k-NN almost outperforms all deep learning baselines, and some of these DL methods such as DrugCLIP are well-known. It is noted that these baselines have been substantially modified from their original implementations, this raises concerns about whether they are properly adapted and optimized, and whether the comparisons are convincing.

  3. Some improvements are marginal. In Table 1, GenFish only slightly improves R@10 over k-NN (35.27% vs. 34.83%). In Table 2, its Overall AP is only marginally better than LigUnity (58.58% vs. 57.75%), while its Enzyme AP is slightly worse (81.22% vs. 81.97%). Overall, the gains appear limited in many cases.

  4. In Figure 4, Oracle Masking only improves the standard contrastive R@50 from 31.30% to 33.40%, while GenFish achieves 50.59%. Although this confirms that false-negative competition affects contrastive learning, it does not adequately explain the much larger improvement from GenFish. The connection between the proposed motivation and the performance gains remains unclear.

  5. The writing and presentation need substantial improvement. Some figures are too small to read. There are also noticeable typos and formatting inconsistencies in many figures and the main text. The related work discussion is also quite limited, with relatively few references.

Questions for authors

See weaknesses 2, 3, 4, more experiments or stronger justification are needed.

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

This paper proposes GenFish, which formulates reverse target fishing as generative retrieval over protein semantic identifiers learned through residual vector quantization. It also introduces ProTargetBench, a benchmark covering 20,431 human proteins with separate evaluations for full-catalogue retrieval and measured-panel discrimination.

Critical strengths
  1. The paper addresses an important challenge in drug–target discovery, particularly incomplete and selectively observed molecule–protein associations.
  2. ProTargetBench provides a potentially valuable evaluation resource by distinguishing experimentally inactive pairs from unmeasured associations.
Critical weaknesses
  1. The proposed generative objective does not completely eliminate the negative competition underlying its motivation. The central motivation is that discriminative training can suppress unmeasured positive associations by treating them as negatives. However, GenFish optimizes cross-entropy over normalized code vocabularies, where increasing the likelihood of an observed target identifier implicitly decreases the probability assigned to alternative code tokens. These alternatives may correspond to unmeasured but genuinely relevant proteins. Although semantic-code sharing can mitigate this effect, the method essentially moves competition from protein identities to code distributions rather than eliminating it. The paper does not sufficiently characterize when such implicit competition remains harmful or why this formulation is preferable to other positive-unlabeled learning strategies.

  2. The comparisons with existing molecular retrieval methods do not adequately isolate the benefits of generative retrieval. Several neural baselines are substantially modified to accommodate the benchmark. For example, pocket-based representations are replaced by sequence-derived features, and important components of the original HypSeek and LigUnity objectives are omitted. These adaptations are disclosed but make it difficult to interpret performance differences as advantages over the original approaches. More importantly, a strong discriminative method explicitly designed for positive-only or positive-unlabeled supervision under the same protein representations is missing. Such a comparison is needed to distinguish the benefit of generative identifiers from the benefit of avoiding conventional negative sampling.

  3. The false-negative experiment does not sufficiently explain the magnitude of GenFish's improvement. In Figure 4, oracle removal of hidden-positive competition improves the contrastive baseline's held-out R@50 from 31.30% to 33.40%. However, GenFish achieves 50.59% under the same visible positive associations, substantially exceeding the oracle-masked baseline. Thus, the controlled experiment demonstrates that false-negative competition affects contrastive retrieval, but does not explain most of the observed performance gap. The authors should further disentangle the effects of positive-only supervision, semantic identifier sharing, and the autoregressive prediction architecture.

  4. Proteome-scale generalization is not sufficiently evaluated despite the large candidate catalogue. Although ProTargetBench contains 20,431 human proteins, Appendix A reports that 17,564 have no positive training associations. The main evaluation uses only 198 held-out molecules, and the reported metrics are not sufficiently stratified by target training support, protein family, or target rarity. Consequently, it remains unclear whether GenFish can recover genuinely underrepresented or previously unsupported targets, rather than primarily retrieving proteins that are well represented in the training data. Furthermore, the full-catalogue metrics evaluate recovery of recorded positives, not experimental correctness of newly proposed interactions. These distinctions are important for interpreting the practical value of proteome-scale target fishing.

Questions for authors

See weekness.

Reviewer_tyM6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents GenFish, a generative retrieval framework for reverse target fishing. GenFish represents proteins with discrete semantic identifiers, and predicts the probabilities of the identifier codes when given a molecule. Retrieval is then performed by ranking proteins by their identifier probabilities. Furthermore, the authors build the ProTargetBench benchmark to evaluate full-catalogue target retrieval and activity discrimination within measured panels. On ProTargetBench, GenFish outperforms the baselines on the three full-catalogue retrieval metrics.

Critical strengths
  1. ProTargetBench provides a clear and useful evaluation protocol for the reverse target fishing task.

  2. GenFish provides a simple and well-motivated way to learn from sparse and long-tailed associations.

  3. Experimental results are good. GenFish outperforms the baselines, and the analytical experiments are helpful for understanding the proposed method.

Critical weaknesses
  1. The technical contribution of GenFish over existing semantic-ID retrieval methods is moderate. The core idea and mechanism of GenFish are very close to those of TIGER. However, the experiments do not include a matched semantic-ID baseline using standard full-prefix AR prediction. Using the same molecular inputs, protein identifiers and training data can help isolate the contribution of GenFish's adjacent-code, low-rank conditioning.

  2. It remains unclear whether the performance gain mainly comes from avoiding explicitly treating unmeasured pairs as negatives. Although the positive-hiding experiment shows that false-negative competition lowers the performance of the adapted contrastive model, it does not show how much of GenFish’s advantage comes from mitigating this issue; after all, the performance gain might well come from its representation and architecture.

  3. The design choice of code prediction order is not fully justified. GenFish follows the residual-quantization order, but this construction order does not necessarily imply that the same direction should be used for conditional prediction. It remains unclear how sensitive the retrieval performance is to this choice.

Questions for authors

See "weaknesses"

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.