acceptodds
Under review as a conference paper at ICLR 2027

Learning a PRECISE language for small-molecule binding

Abstract

Querying a given protein target against a large set of candidate small-molecule binders is a crucial first step in the drug-discovery pipeline. With billion-compound libraries now available, accurate and fast virtual screening is critical for identifying potential drugs. Sequence-based drug-target interaction (DTI) prediction methods are fast and highly scalable, but leave important challenges in localizing binding sites, handling targets beyond single-chain proteins, and selecting candidates for expensive downstream assessment. We introduce PRECISE, which learns a structure-based protein representation and its compatibility with a quantized codebook of small-molecule space. Specifically, we build on the molecular codebook introduced by CoNCISE and formulate DTI prediction as querying quantized drug representations against a geometric neural network-based representation of the protein surface enriched with electrostatic and geometric features. PRECISE addresses key limitations of previous approaches while remaining highly accurate and scalable. It also generalizes automatically to protein complexes and metal-binding proteins without specialized training. High-throughput virtual screening often fails to align well with downstream workflows. To fix this, we introduce PRECISE-MCTS, a tree search algorithm that uses feedback from a downstream scoring oracle to refine PRECISE hits under a limited time or evaluation budget. Across Vina, Boltz, and Protenix, PRECISE-MCTS acts as a force multiplier for downstream methods. Under matched evaluation budgets, it improves top-10 mean oracle scores in 29 of 30 target–oracle comparisons.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
drug-target interaction prediction, high-throughput screening, quantized codebook, protein complex

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_2ja3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a hierarchical method for virtual screening. The newly trained model first identifies the high-probability ligand cluster (ligand code), then uses the oracle function to perform an MCTS across all candidates. The model is trained with ligand-protein surface pairing with augmented negative samples so that the model could be used for both known pocket screening and pocket-unknown screening. Evaluation is highly biased by in silico docking score and some experimental settings might be questionable.

Critical strengths
  • PRECISE outperforms both ground truth pocket baselines and sequence-level models.
  • Rank according to ligand code - search with oracle functions is an innovative 2-stage framework.
Critical weaknesses
  • Most of the method componentss are dependent on previous works. e.g. the ligand code design is depended on CoNCISE. The MCT is using ConPLex embedding
  • The method section of the paper needs more polish and structured writing to make it easier to follow the whole framework.
  • Unaligned training data might be a problem. None of the baseline methods is trained on the same PLINDER training split. This also raises - the question of ablation studies on different components of the model.
  • Analysis on the quantized ligand space should be provided here.
Questions for authors
  • How would the method perform if we use ConPLEX embedding to train the ligand-surface scoring framework in the beginning so that we do not need a 2-stage procedure
  • How would the method perform compared with other baseline methods on DUD-E benchmark (EF metric)?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 2ja3

Unaligned training data might be a problem. None of the baseline methods is trained on the same PLINDER training split. This also raises - the question of ablation studies on different components of the model

The new Tables 1 and 2 above in the response to Reviewer GvH2 , show the performance of PRECISE in all the DrugClip related experiments, on the same training/testing/validation datasets as in the original DrugClip paper, including using their exact version of DUD-E where they made sure there was no leakage.

Analysis on the quantized ligand space should be provided here.

Can the reviewer clarify what they are looking for here please? We are not sure we understand what is being asked for.

Q1 How would the method perform if we use ConPLEX embedding to train the ligand-surface scoring framework in the beginning so that we do not need a 2-stage procedure

This is very similar to the PRECISE-CODE versus PRECISE-NOCODE study that we did (see Appendix, section C2, table), We find that the codes can improve the embedding, so the ConPLEX embedding would not perform as well.

Q2 How would the method perform compared with other baseline methods on DUD-E benchmark (EF metric)?

As remarked to both other reviewers, we now have an extensive comparison (see Tables 1 and 2 in the response to Reviewer GvH2) and we are performing better, however, as also remarked we don’t think doing better in EF metric on DUD-E is the main contribution of PRECISE. Rather, the main contribution of PRECISE is that we are the only method that can handle multichain and metalloproteins; plus we can screen all of Enamine in about 30 minutes, whereas competitors will be roughly 50 times slower.

I really appreciate the author's additional results on EF and de-leaked datasets. However, the results in LIT-PCBA benchmark is quite discouraging. I will maintain my rating regarding limited novelty based on related works.

Reviewer_GvH2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PRECISE for large-scale structure-based virtual screening. PRECISE represents the protein using its molecular surface and scores local surface regions against the quantized ligand codes learned by CoNCISE. This allows the model to perform pocket-directed screening while avoiding scoring every molecule in a large library individually. The same model can also be used for whole-protein DTI prediction and ligand-conditioned pocket localization. The paper also introduces PRECISE-MCTS. It starts from compounds in high-scoring codes and uses feedback from a downstream scoring method to search the candidate space under a limited evaluation budget. The experiments include PLINDER, DUD-E, pocket localization, composite protein surfaces, and large compound libraries.

Critical strengths

1.The scalability aspect is interesting. Using the CoNCISE codebook makes it possible to screen very large libraries without running the full model on every molecule. This is a practical problem for structure-based screening, and the paper gives a reasonably complete system around it.

2.I found the pocket localization and limited-budget screening experiments useful, since they test the method beyond the whole-protein DTI setting. The MCTS experiments also show that oracle feedback can improve over random selection from the same high-scoring codes.

Critical weaknesses

1.I am not fully convinced by the methodological contribution. The main components are largely inherited from prior work: the ligand codes come from CoNCISE, the protein side uses an existing geometric architecture over surface features, and the downstream search uses ConPLex embeddings together with clustering and oracle feedback. The resulting system is useful, but it is not clear to me what new modeling principle is introduced beyond integrating these components for large-scale screening.

2.The virtual screening evaluation feels limited relative to the paper’s main claims. DUD-E is the main standard screening benchmark, where PRECISE performs similarly to, but slightly below, DrugCLIP. I would have liked to see evaluation on a more realistic benchmark such as LIT-PCBA. Most of the later screening experiments measure docking or model-based oracle scores rather than experimentally defined active enrichment, so it remains unclear whether the gains in scalability and oracle optimization translate into better virtual screening performance.

Questions for authors

1.Could the authors provide results on LIT-PCBA? I would also be interested in EF1% or BEDROC in addition to AUROC and EF5%. Strong performance on these metrics would increase my confidence in the virtual screening claims.

2.PRECISE-MCTS uses random descendant sampling. How stable are the results across different random seeds? Also, Appendix F.1 describes naive clustering baselines, but these do not seem to appear in the main 65-target comparison. Could the authors report a direct comparison between these baselines and PRECISE-MCTS? This would help clarify how much of the gain comes from adaptive search rather than the clustering structure itself.

1 comment
AuthorsOpenReview ↗

Response to reviewer GvH2

W1. I am not fully convinced by the methodological contribution…The resulting system is useful, but it is not clear to me what new modeling principle is introduced beyond integrating these components for large-scale screening.

PRECISE is able to screen multichain proteins and metalloproteins, two classes of targets that cannot be handled by prior methods. This is a significant breakthrough. It is surprising that the codes can carry over to the surface representation and give us this result.

W2.The virtual screening evaluation feels limited relative to the paper’s main claims. DUD-E is the main standard screening benchmark, where PRECISE performs similarly to, but slightly below, DrugCLIP. I would have liked to see evaluation on a more realistic benchmark such as LIT-PCBA. Most of the later screening experiments measure docking or model-based oracle scores rather than experimentally defined active enrichment, so it remains unclear whether the gains in scalability and oracle optimization translate into better virtual screening performance.

We fixed an issue with updating training in batch mode, and PRECISE is now performing better than DrugCLIP on both DUD-E and LIT-PCBA which we ran on at this reviewer’s request. ConGLuDE performs worse that PRECISE on DUD-E but somewhat better than PRECISE on LIT-PCBA. Results are as follows:

Here are the results on the DUD-E benchmark:

MethodAUROC (%)BEDROC (%)EF 1%
DrugCLIPZS80.9350.5231.89
ConGLUDe82.0450.8032.52
PRECISE88.4060.1038.40

And here are the results on the LIT-PCBA benchmark:

MethodAUROC (%)BEDROC (%)EF 1%
DrugCLIP57.176.235.51
ConGLUDe66.2513.6312.25
PRECISE64.507.906.30

Q1.Could the authors provide results on LIT-PCBA? I would also be interested in EF1% or BEDROC in addition to AUROC and EF5%. Strong performance on these metrics would increase my confidence in the virtual screening claims.

We provide these results as requested (tables above). We also tested ConGLuDE on DUD-E and LIT-PCBA in response to reviewer 1, and results are split: ConGLUDE is somewhat better than PRECISE on LIT-PCBA, but ConGLuDE is worse than PRECISE on DUD-E. Even though our results now look good on these benchmarks, we want to stress that this is not in our opinion the principal contribution of the paper: PRECISE can screen multimeric proteins and metalloprotein complexes. ConGLUDE, DrugCLIP, CoNCISE, and ConPLex cannot do so. Plus PRECISE is much faster than other screening methods. The PRECISE formulation allows it to screen Enamine's billions-large drug library within 30 mins (including the time for MCTS+docking), compared to ~1 day for DrugCLIP or ConGLUDE.

Q2.PRECISE-MCTS uses random descendant sampling. How stable are the results across different random seeds? Also, Appendix F.1 describes naive clustering baselines, but these do not seem to appear in the main 65-target comparison. Could the authors report a direct comparison between these baselines and PRECISE-MCTS? This would help clarify how much of the gain comes from adaptive search rather than the clustering structure itself.

We did not end up having enough time to address this exact question. However the new RAD results (see response to reviewer xz52 above) seem to indicate that having the PRECISE code up front steers the sampling toward the high-scoring portions of the space. .

Reviewer_xz52OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PRECISE targets two problems in sequence-based ultra-high-throughput virtual screening: pocket specificity is limited, and large candidate sets are hard to prioritize when the downstream evaluation budget is small. It combines CoNCISE's pretrained quantized molecular codes with geometric protein-surface representations for local binding prediction, and uses PRECISE-MCTS for oracle-guided hierarchical search. Experiments show improved whole-protein DTI prediction over CoNCISE (AUROC 0.908 vs. 0.892) and better top-10 mean oracle scores in 29 of 30 target-oracle comparisons.

Critical strengths
  1. The figures are visually appealing and easy to understand. The visualizations do a good job of conveying both the method and the results.
  2. The paper shows pocket-specific screening on billion-scale indexed molecular libraries. With a fixed docking budget, the reported median screening time increases from 75 to 135 minutes as the library grows from 250 million to 7 billion compounds.
Critical weaknesses
  1. The research gap and new contributions are not fully clear. Prior work such as ConGLUDe [1] already handles structure-aware virtual screening and ligand-conditioned pocket prediction, while CoNCISE supports billion-scale molecular retrieval. The paper explains why it builds on CoNCISE, but it is not clear which important limitations remain in existing methods and what new benefits PRECISE provides.
  2. The experimental settings are not fully clear. The reported DrugCLIP results on DUD-E are lower than those of the original full model and match an ablation setting. Since this difference is not explained, it is hard to tell whether the comparison is fair or whether the screening performance is really similar.

[1] Schneckenreiter et al. (2026). Contrastive Geometric Learning Unlocks Unified Structure- and Ligand-Based Drug Design.

Questions for authors
  1. Could you clarify which important limitations of existing methods PRECISE addresses? What are its main new contributions beyond prior work such as ConGLUDe [1] and CoNCISE, and which results support them?
  2. What does PRECISE-MCTS add over existing adaptive search methods such as RAD [2]? Does its specific search design offer benefits over simpler feedback-guided strategies under the same candidate pool and oracle budget?
  3. Which DrugCLIP model and settings were used? Please explain why the reported AUROC of 0.771 and EF@5% of 8.69 match the ablation results in Table 5 of the original DrugCLIP paper (training without RDKit conformer replacement and HomoAug), while the full model reports 0.8093 and 10.66, respectively?

[2] Hall and Keiser (2024). Retrieval Augmented Docking Using Hierarchical Navigable Small Worlds.

1 comment
AuthorsOpenReview ↗

Response to Reviewer xz52

W1: it is not clear which important limitations remain in existing methods and what new benefits PRECISE provides.

Unlike prior methods ConGLUDe and CoNCISE, PRECISE is able to handle multichain protein targets and metalloproteins: no previous method can do this.

W2: The experimental settings are not fully clear. The reported DrugCLIP results on DUD-E are lower than those of the original full model and match an ablation setting. Since this difference is not explained, it is hard to tell whether the comparison is fair or whether the screening performance is really similar.

Thank you for pointing this out. Our results for DrugClip+Pocket in Table 1 and on the DUD-E benchmark reported in Section 3.1 were using the pre-trained weights of DrugCLIP from their public github here: https://github.com/bowen-gao/ While our results match the checkpoints they publish on their github, in the paper, they instead report a more powerful DrugClip model including augmented training with homolog pocket 3D structures. In order now to be as fair as possible to DrugClip, in the revised section B3, on the DUD-E and LitPCBA benchmarks, we now quote best DrugClip performance exactly from their paper (and remove DrugClip+Pocket from Table 1, because we have no way to be sure we are running the best DrugClip model in this setting). Our new PRECISE results (now that we are updating things correctly in our batching, see note above) do better than DrugClip on DUD-E and LitPCBA across the board. However, note that DrugClip needs to be told the pocket, and PRECISE does not, so this is not an apples to apples comparison.

Q1: Could you clarify which important limitations of existing methods PRECISE addresses? What are its main new contributions beyond prior work such as ConGLUDe and CoNCISE, and which results support them?

The most important limitation that we address is we can screen for targets that are impossible for ConGLUDe and CoNCISE, namely multichain proteins and metalloproteins. No existing method other than PRECISE that we are aware of can do this. PRECISE is shown to be competitive to these other methods on the problems they can solve, but fundamentally, the new thing is the generalized set of targets we can screen

Q2: What does PRECISE-MCTS add over existing adaptive search methods such as RAD? Does its specific search design offer benefits over simpler feedback-guided strategies under the same candidate pool and oracle budget?

In response to this reviewer, we did an investigation of RAD. We find that adding PRECISE codes to RAD enriches the strongest hits much earlier with a smaller docking budget. Using their DUDE-Z docking set, we ran RAD only on molecules that scored highly in PRECISE and compared it to using the original RAD algorithm on pure morgan fingerprints. We found that at lower oracle budgets PRECISE codes enrich the top 100 binders by DOCK score earlier than that of the original morgan fingerprint implementation.

Molecules docked% of libraryOriginal RADPRECISE Aided RAD
1000.009%0.00.1
2500.021%0.10.9
5000.043%1.42.4
1,0000.086%4.19.1
2,0000.17%16.721.4
2,9230.25%23.027.3
5,8460.5%40.646.1

Q3: Which DrugCLIP model and settings were used? Please explain why the reported AUROC of 0.771 and EF@5% of 8.69 match the ablation results in Table 5 of the original DrugCLIP paper (training without RDKit conformer replacement and HomoAug), while the full model reports 0.8093 and 10.66, respectively?

We repeat our above answer to this again: Our initial results were using the pre-trained weights of DrugCLIP from their public github here: https://github.com/bowen-gao/DrugCLIP However, the most powerful DrugClip model, as the reviewer notes, includes augmented training with homolog pocket 3D structures. In order to be as fair as possible to DrugClip, we now use the reported DrugClip performance numbers from their paper, and our new PRECISE results (now that we are updating things correctly in our batching, see note above) do better than DrugClip on both DUD-E and LitPCBA across the board.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.