acceptodds
Under review as a conference paper at ICLR 2027

CoPocket: Pocket-Conditioned Molecular Generation via Multi-Modal Feature Constraints and Interaction-Aware Contrastive Learning

16 views

Abstract

Targeted molecular generation aims to generate molecules against protein pockets, which is a challenging task in drug discovery. Most existing methods neglect the multiscale hierarchical representation of protein pockets, and fail to sufficiently exploit critical interaction within protein-ligand binding. This paper proposes a drug design framework, namely CoPocket, focusing on different modal features of binding pockets to guide high quality molecule generation. First, we construct an atomic-level spatial interaction graph to capture the binding patterns between protein pockets and ligands. Guided by these patterns, we develop a cross-modal contrastive learning module to optimize the representation of pocket sequences. We also capture higher-order structural relations and atomic-level structural features through hypergraphs and atomic graphs of protein pockets, respectively. A cross-modal hierarchical fusion module fuses multi-modal embeddings from pocket sequences, pocket hypergraph structure, and pocket atomic graph structure. In addition, a pocket-conditioned molecule generator employs fused features as constraints for generating molecules possessing target characteristics. Experiments demonstrate that CoPocket outperforms existing methods across multiple key metrics, validating its effectiveness for molecular generation tasks.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Targeted Molecule Generation, Cross-Modal, Hierarchical Representation

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_m4deOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CoPocket, a pocket-conditioned generation framework inegrating protein sequence embeddings, pocket hypergraphs, atomic graphs, and interaction-aware contrastive learning. A hierarchical fusion module combines multimodel pocket representations to guide molecular generation. Experimentsnon CrossDocked2020 report competitve results in docking socres, molecular properties, and diversity.

Critical strengths
  • The paper addresses pocket-conditioned molecular generation, an important problem in structure-based drug discovery, and inegrates sequence, structural, and interaction infromation.

  • CoPocket achieves favorable Vina scores, synthentic accessibility, and molecular diversity compared with the selected baselines.

Critical weaknesses
  • The proposed framework primarily combines established techniques, including contrastive learning, hypergraph neural networks, cross-modal feature fusion, and attention-based generation. Prior work such as DeepICL and FGMOL has already explored interaction-aware ligand generation. The authors do not sufficiently clarify what fundamentally distinguishes CoPocket from these existing approaches or demonstrate a substantial new modeling capability.

  • The evaluation omits several relevant recent methods, including MolCRAFT, MolPilot, and FGMol. Moreover, only 52 protein–ligand pairs are retained for testing, raising concerns about generalization. The improvement over TSMGen in Vina score is marginal (−7.693 vs. −7.642). Additional benchmarks, stronger baselines, structural-validity evaluations, and statistical significance analyses are necessary to substantiate the performance claims.

  • Table 2 does not consistently demonstrate the effectiveness of the proposed components. Removing the atomic graph improves the Vina score from −7.693 to −7.708, while removing both fusion modules further improves it to −7.730. Several variants also achieve nearly identical diversity scores. These findings weaken the justification for the architectural complexity and the claimed contributions of individual modules.

Questions for authors
  1. What are the key methodological differences between CoPocket and prior interaction-aware molecular generation methods such as DeepICL?

  2. Can the authors include stronger recent baselines, such as MolCRAFT and MolPilot, and additional benchmarks (PoseBusters and CBGBench) to demonstrate generalizability?

  3. Why were only 52 test protien-ligand pairs retained? Were all baselines evaluated on the same filtered test set?

  4. Why does removing certain proposed modules improve the Vina score? Can the authors provide stronger evidence of their individual contributions?

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

CoPocket combines pocket sequence embeddings, residue-level hypergraphs, and atom-level graphs for conditional molecular generation. Besides, it uses contrastive learning to incorporate pocket–ligand interaction information and then fuses the different representations. Overall, the experiments include baseline comparisons, component ablations, and case studies.

Critical strengths
  1. Using pocket–ligand interactions to guide representation learning is well motivated, and the combination of sequence and structural information is also a reasonable choice for this task.
  2. The ablations cover the main representations and fusion modules, as the full model achieves the highest QED, SA, and Lipinski scores among the variants in Table 2. Thus, the result provides some evidence that the components in the framework are useful.
Critical weaknesses
  1. The paper has several reporting issues, including “represention” on page 4 and “of of” on page 8. The full model’s SA is reported as 0.762 in Table 1 but 0.760 in Table 2. Section 3.5 also claims the best results on QED, SA, Lipinski, and Time, although Table 2 has no Time column.
  2. PAMol already combines pocket hypergraphs with sequence features, while TSMGen uses both residue-level and atom-level representations. Although CoPocket introduces contrastive learning and additional fusion modules, it does not clearly explain how these changes address limitations of the earlier methods.
  3. The Vina improvement over TSMGen is only 0.051, with no reported statistical uncertainty to show how stable this difference is. Removing both BCF and AGMF also improves Vina, so the benefits of these modules need to be discussed more carefully across the different metrics.
  4. Section 2.2 constructs hyperedges using the minimum heavy-atom distance between residues, while Appendix A.2 uses Cα distances. These definitions can produce different graphs, leaving an important part of the method unclear.
  5. Appendix A.5.2 claims structural novelty based on the t-SNE plot and selected examples in Figure 6. However, a quantitative assessment against the training and pretraining data is missing, so the extent of this novelty remains unclear.
Questions for authors
  1. What are the main methodological differences from PAMol and TSMGen? Please clarify what the contrastive learning and fusion modules contribute beyond these methods.
  2. How consistent are the reported improvements across runs or test pockets? Please provide uncertainty estimates for the main comparisons and ablations.
  3. Which distance definition was used to construct the hypergraphs in the experiments: minimum heavy-atom distance or Cα distance? Please reconcile the descriptions in Section 2.2 and Appendix A.2.
Reviewer_TfYGOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CoPocket combines three descriptions of a protein pocket: a ProteinBERT sequence representation, a residue hypergraph, and an atom graph. During training, the sequence representation is aligned with a pocket–ligand interaction-graph representation using a bidirectional contrastive loss. Outer-product fusion and a learned gate combine the pocket features, which condition an autoregressive SMILES generator pretrained on ChEMBL. The evaluation reports docking and molecular-property metrics on a filtered CrossDocked subset, with additional computational case studies on PARP13.

The idea of transferring interaction information into a pocket encoder is reasonable. However, the current evidence does not establish a reliable improvement over existing pocket-conditioned generators, and the mathematical description does not yet make the claimed interaction mechanisms fully convincing.

Critical strengths
  1. The problem is useful. Efficient generation of chemically reasonable molecules for a specified pocket is an important component of computational screening.
  2. The model attempts to combine complementary levels of description. Sequence context, residue neighborhoods, and atom-level structure can contain different information. Using ligand interactions as training supervision for a pocket-only representation is a sensible design choice.
  3. The evaluation considers more than docking. QED, normalized SA, diversity, and a generation-time comparison help expose practical trade-offs. CoPocket is competitive on several reported chemical-property metrics.
  4. The paper supplies multiple ablations and substantial implementation detail. Removing individual modalities, fusion modules, and contrastive learning is a useful start. The appendix also explicitly describes autoregressive generation from the pocket and the generated prefix, which helps clarify the intended inference setting.
Critical weaknesses

1. The benchmark subset changes substantially, and comparability is unresolved

Section 3.1 states that preprocessing reduces the usual 100,000 training pairs to 14,770, and the 100 test pairs to 52, because some PDB files cannot be converted into hypergraphs. This removes 85.23% of training examples and 48% of test examples. Such extensive filtering can materially change both the training problem and the difficulty of the test set.

Were all baseline results in Table 1 recomputed on these exact 52 pockets using the same evaluation pipeline, or were some taken from published results on a different subset? The manuscript does not make this clear. Until it does, the reported ranking cannot be interpreted confidently. Please release the retained identifiers, exclusion reasons, and evaluation outputs, and compare the retained and excluded pockets in size, target family, and ligand properties. If practical, repair the parser and report performance on the full standard test set.

The ChEMBL pretraining overlap also needs to be audited. Report exact-compound and scaffold overlap with test ligands, as well as target/pocket similarity across the supervised split. CrossDocked contains multiple poses and related structures; counting pairs does not establish the number of independent proteins or ligands. These are requests for a data audit, not an allegation of leakage.

2. The numerical results do not yet establish a robust improvement

The leading docking comparison is −7.693 versus −7.642, an absolute difference of 0.051 kcal/mol (Table 1). Presenting this as a 0.67% improvement gives an impression of precision without showing uncertainty or practical relevance. No per-pocket confidence intervals or independent-run variability are reported. The full model also has lower QED than TSMGen, 0.606 versus 0.659, and ties its normalized SA at the precision shown in Table 1.

Table 2 further weakens the claim that the proposed components are necessary for better target binding. Removing the atom graph improves docking from −7.693 to −7.708; removing both BCF and AGMF improves it to −7.730. Removing the hypergraph changes docking by only 0.002. The QED improvements of the full model are more consistent, but modest. This supports a possible trade-off, not a clear overall superiority claim.

Please report paired per-pocket differences, uncertainty across training and sampling seeds, and a prespecified multi-metric success criterion. For generalization across targets, the pocket—not each sampled molecule—should be the primary unit of statistical comparison.

The additional results in Table 3 compare the best docking score with different baselines on different targets. These values depend strongly on how many candidates were generated and how they were selected. They cannot substitute for a common all-method/all-target evaluation with a matched sampling budget.

3. Several method definitions are inconsistent or undermine the proposed interpretation

The hypergraph is defined in two incompatible ways. Section 2.2 constructs edges using the minimum heavy-atom distance between residues below 5 Å. Appendix A.2 instead uses Cα–Cα distances below 5 Å. These produce different neighborhoods: side-chain contacts can be close while their Cα atoms are much farther apart. Please state which definition was implemented and update the other description.

The interaction block does not explicitly transmit neighboring atom features as written. In Equations (21)–(23), the key and value for edge (i,j)(i,j) are computed from [hi,eij][h_i,e_{ij}], while Equation (20) defines the edge feature from distance. There is no hjh_j in this message. With these definitions, an update can depend on the central atom and surrounding distances without directly knowing the chemical identity of the atom at the other end of each edge. This matters for distinguishing, for example, different chemical partners at comparable distances. Is Equation (21) meant to contain hjh_j, or is there an omitted edge update or chemical edge feature? If the implementation follows the printed equations, the claim of learning chemically specific interaction patterns needs stronger support.

The “reverse” fusion matrix contains the same information as the forward matrix. Equations (9)–(10) give Mf=ah⊤M_f=a h^\top and Mr=ha⊤=Mf⊤M_r=h a^\top=M_f^\top. Flattening the transpose simply permutes the entries. Separate learned projections can provide extra capacity, but they do not by themselves establish a new directional interaction mechanism. A parameter-matched single-outer-product control is needed to support that interpretation.

The cross-attention description needs tensor-level clarification. The pocket encoders are globally pooled to one vector per example, and the generator uses pocket features as keys and values. If this is one key/value token per pocket, the attention softmax is identically one and that branch cannot select pocket information based on the molecular query. If the vector is expanded into multiple tokens, or unpooled features are used, please specify how. The reverse branch also needs a clear description of causal masking during training.

Some of these issues may be typographical or omitted implementation details. They should be resolved rather than treated as demonstrated implementation bugs. Together, however, they make it difficult to assess the method's central technical claims.

4. The contribution of interaction supervision and fusion is not isolated

The paper combines established ingredients: pretrained sequence embeddings, graph/hypergraph encoders, jumping-knowledge aggregation, contrastive alignment, bilinear feature interactions, gates, and a conditional language model. Such an integration can be useful, but the novelty needs to be located precisely and supported by controlled evidence.

TSMGen is an especially relevant comparator: it is cited as a method based on higher-order structural dependencies and bidirectional fusion, yet the paper does not explain clearly what CoPocket changes relative to it beyond the aggregate performance table. A component-level comparison would be more informative than a list of separately named modules.

Please include a generator with the same ChEMBL pretraining and training budget but a simple pooled pocket representation, a parameter-matched concatenation/bilinear fusion baseline, and shuffled-pocket or mismatched-pocket controls. The latter are important because favorable QED, SA, and pooled diversity can come from the molecular prior even if the model uses little target information.

The contrastive negatives also need attention. Equations (3)–(5) treat all off-diagonal pocket–interaction pairs as negatives. Repeated pockets with different valid ligands can therefore create contradictory supervision. Please describe grouping, deduplication, and multi-positive handling. Distinguishing correct from perturbed poses of the same pocket–ligand pair would more directly test whether the representation learns interactions rather than pocket identity.

5. The geometry, efficiency, and molecular-quality comparisons lack an end-to-end protocol

Appendix A.4 describes a SMILES generator, but the paper then evaluates bond lengths, all-atom distances, docking poses, and generation speed against methods that generate 3D structures directly. The conversion from generated SMILES to coordinates is not adequately specified. If a conformer generator or force-field optimization determines bond lengths, the agreement in Figure 2 cannot be attributed solely to CoPocket learning geometry.

Please document the full pipeline: SMILES sanitization, stereochemistry, conformer generation, minimization, protonation, receptor preparation, docking settings, and rejected samples. This is particularly important because the preprocessing mentions “isomer removal” without defining whether it means removing stereochemical information or eliminating duplicate stereoisomers.

Similarly, 4.3 seconds for 100 molecules in Figure 7 is encouraging, but only comparable if all methods are timed to the same output: for example, valid molecules with usable 3D conformers, or docked candidates. Separate decoding, conformer generation, and docking time, and report common hardware, batching, and success counts.

The main table omits validity, uniqueness, and explicit training-set novelty, which are important for an autoregressive SMILES model. Although Section 3.3 gives the SA range as 0–1, please provide the normalization formula, the diversity fingerprint/distance, and the precise Lipinski definition. The latter is described as a count of conforming molecules, while the table reports values around 4–5, suggesting a different quantity such as the average number of satisfied criteria. A t-SNE plot cannot establish novelty or broad chemical-space coverage by itself.

6. The case-study interpretation exceeds the evidence and contains inconsistencies

Docked poses and distances of 2.2–3.4 Å do not establish stable hydrogen bonds or robust binding, as claimed in Section 3.6. Hydrogen-bond assignment also depends on donor/acceptor chemistry and geometry. Appendix A.6 assigns multiple hydrogen bonds to aromatic fluorine and a halogen bond to fluorine; these assignments require particular care and should not be inferred from contact proximity alone.

There is a concrete consistency problem in the Mol_1 discussion: Appendix A.6 refers to a “phenyl ring connected to the cyano group,” but the Mol_1 SMILES in Figure 3 contains a trifluoromethyl group and no nitrile. The Mol_1 structure in the Figure 9 interaction diagram also appears different from the structure in Figure 3. Please reconcile molecule identities, poses, interaction diagrams, and prose using a single set of deposited structures and identifiers.

Finally, 7KZH and 7TGQ are structures of the PARP13/ZAP central domain, not independent protein targets. The paper correctly introduces them as PARP13 structures, but the concluding language about validation “across different targets” overstates the breadth of this case study. Docking candidates to these structures is a useful hypothesis-generation exercise; it does not establish target specificity or biological modulation.

Questions for authors
  1. Were all Table 1 methods rerun on the same 52 pockets, with identical docking and filtering? What caused the unusually large number of parsing failures?
  2. Which hypergraph definition and interaction-message equation are actually implemented?
  3. How many key/value tokens enter each generator cross-attention layer, and how is causality maintained in the reverse attention branch?
  4. Does interaction supervision improve a target-specific endpoint over a matched pretrained generator, and does that benefit survive shuffled-pocket controls and independent seeds?
  5. How are the learned loss weights in Equations (19)/(36) constrained? Without a specified parameterization, the normalized weighted auxiliary objective does not establish a well-behaved balancing mechanism. Please report learned values and compare with fixed weights.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.