acceptodds
Desk rejected at ICLR 2027

PL-JEPA: Joint Predictive and Generative Learning for Structure-Based Drug Discovery

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Structure-based drug discovery is typically formulated as two separate tasks. Virtual screening ranks existing molecules for a protein pocket, whereas structure-based molecular generation constructs new molecules conditioned on that pocket. Although both tasks are governed by the molecular constraints imposed by the same binding pocket, they are commonly modeled separately. We introduce PL-JEPA, a protein-ligand joint embedding predictive architecture that represents these shared constraints as a binding embedding in ligand representation space. The binding embedding provides a common interface for molecular recognition and realization. It can be directly matched with ligand representations for virtual screening or used to condition an equivariant diffusion decoder for 3D molecular generation. PL-JEPA progressively learns this shared interface through predictive alignment, representation-conditioned generation, and joint predictive and generative training, making the binding embedding both searchable and realizable. A single checkpoint supports virtual screening, 3D molecular generation, and target fishing. On DUD-E and LIT-PCBA, PL-JEPA improves BEDROC over DrugCLIP by 36.2% and 84.3%, reaching 68.80 and 11.48, respectively. On CrossDocked2020, it achieves a median Vina Dock score of -8.23 kcal/mol and a diversity of 0.78. On CASF-2016, it improves target fishing Acc@5 from 62.59 to 65.96. Ablation studies show that the three-stage training progressively establishes and integrates the searchable and realizable properties of the binding embedding.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Protein–ligand representation learning, joint-embedding predictive architecture, structure-based drug design, virtual screening, molecular generation, equivariant diffusion models, target fishing

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_B43NOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a framework named PL-JEPA that learns to align the embeddings of ligands and pockets and then uses these aligned embeddings for three tasks within drug discovery pipelines, including virtual screening, molecular generation, and target fishing. The paper shows improved results across tasks and metrics.

Critical strengths
  1. Learning informative molecular representations for search and conditioning generative models sounds technically and is critical in structure-based drug discovery.
  2. The writing is clear and easy to follow.
Critical weaknesses
  1. I don't think the JEPA framing appropriate here. The paper merely aligns ligand and pocket embeddings via a contrative objective on paired data. There are no discussions about JEPA-related techniques, such as context masking. Importantly, JEPA is a non-contrastive approach; the encoders are trained to predict the embeddings of masked target regions from surrounding contexts within the same samples, and therefore it does not require paired data. This clarification is important because ligand-pocket paired data is very limited, and training models via contrastive objectives on such small datasets, such as DUD-E, does not seem to generalize well.
  2. While the overall framework is plausible, I find the proposed approach is incremental, as it combines several previous works, such as contrastive learning, equivariant diffusion, without any critical changes or insightful observations.
Questions for authors
  1. Could you elaborate on the latent calibration objective?
  2. Did you try aligning atom-level representations, instead of global representations via [CLS] tokens?
  3. How sensitive are the docking results?
5 comments

Thank you for your thoughtful review. Part 1 addresses the JEPA framing and contribution; Part 2 addresses the methodological questions.

W1: Multimodal JEPA framing and training data

Our JEPA framing follows recent multimodal developments, particularly VL-JEPA, rather than the masked-region formulation of I-JEPA[1]. VL-JEPA[2] predicts text embeddings from visual inputs using paired data and explicitly adopts a bidirectional InfoNCE objective. Thus, neither within-sample context masking nor a strictly non-contrastive loss is a universal requirement for multimodal JEPA.

PL-JEPA follows the same latent-space prediction principle: a predictor maps the pocket representation to a target representation in ligand space, rather than merely aligning the two raw embeddings (§3.2.2). This formulation is appropriate because a pocket constrains compatible molecular characteristics without specifying a unique ligand. We further extend this architecture through joint predictive and generative training, making the predicted representation useful for both retrieval and molecular generation—not contrastive alignment alone.

Finally, DUD-E is used only for evaluation, not training. Our training corpus comprises about one million pairwise data. We acknowledge that paired-data availability remains a limitation, but generalization is evaluated beyond DUD-E, including LIT-PCBA, which was designed to reduce ligand-based benchmark biases (§4.2). We will clarify both the multimodal JEPA motivation and the training–evaluation distinction.

[1] Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture

[2] VL-JEPA: Joint Embedding Predictive Architecture for Vision-language

W2: Contribution of the shared binding representation

Our contribution is a shared binding representation through which recognition and generation benefit each other, rather than simply placing two existing models in one pipeline.

Technical value. Table 3 demonstrates this mutual benefit: compared with separately trained alignment and generation, our full framework improves screening BEDROC from 58.81 to 68.80, median Vina Dock from −6.03 to −8.23, and diversity from 0.73 to 0.78. Thus, coupling the tasks improves both without sacrificing molecular diversity. Importantly, a searchable representation is not automatically suitable for generation: joint training without decoder pretraining yields diversity of only 0.52, whereas establishing the representation-to-molecule mapping first raises it to 0.77. Latent calibration further reconciles normalized retrieval alignment with unnormalized generative conditioning, improving both tasks.

Practical value. The paper shows where these gains matter. Figure 7 illustrates difficult active–decoy pairs with similar global shapes and favorable docking scores for both, where PL-JEPA corrects DrugCLIP’s ranking. Figure 4 shows simultaneous improvement in docking scores and diversity, rather than trading one for the other; Figure 8 demonstrates structurally distinct candidates for the same pocket.

We additionally evaluate docking-based binding specificity, rather than merely favoring molecules that dock well to many pockets. Target specificity is important for avoiding unintended interactions in drug discovery. Following Gao et al. [3], we evaluate Delta Score, which measures preference for the intended pocket over non-target pockets (using Glide, different from the main paper’s Vina scores). PL-JEPA leads the evaluated generative baselines across all five metrics.

MethodsAbsolute mean ↑Delta mean ↑Delta median ↑Ratio mean ↑Ratio median ↑
Trainset5.7270.044-0.0730.1590.053
Reference (Original)6.6321.1581.029--
AR5.7370.3930.2000.1500.039
Pocket2Mol5.9460.4370.1060.1700.050
DrugGPS6.5540.4900.3870.2350.134
TargetDiff6.6650.3350.1020.2590.129
DecompDiff7.1020.3540.2200.2740.133
SBE-Diff6.8150.5520.2500.3000.216
MolCRAFT7.5230.5830.3510.3550.316
PL-JEPA8.4210.6120.3970.3880.353

This shows that the framework improves preference for the intended target, beyond absolute docking performance alone.

[3] Rethinking Specificity in SBDD: Leveraging Delta Score and Energy-Guided Diffusion

This is Part 2 of our response, addressing Q1–Q3.

Q1: Latent calibration

The latent calibration objective is motivated by the dual-task nature of PL-JEPA, which unifies virtual screening and molecular generation through a shared binding embedding.

For virtual screening, contrastive alignment maximizes cosine similarity between predicted binding embeddings and ligand embeddings. However, cosine similarity primarily constrains their directions rather than their actual values. While sufficient for similarity-based retrieval, this may be inadequate for molecular generation, where the decoder directly uses unnormalized embeddings as conditioning inputs.

Specifically, our decoder is pretrained using ligand embeddings (Stage II), but ultimately conditioned on pocket-predicted embeddings (Stage III). We therefore introduce a scaled smooth-L1 calibration loss, with stop-gradient applied to the ligand embedding target, to minimize their element-wise discrepancies and reduce the conditioning mismatch.

As shown in Table 3, calibration improves DUD-E BEDROC from 65.34 to 68.80 and median Vina Dock from −7.97 to −8.23, demonstrating its benefits for both retrieval and generation.

Q2: Atom-level versus global alignment

Yes, we additionally investigated atom-level representation alignment and evaluated its performance on both DUD-E and LIT-PCBA.

The atom-level variant achieved an AUROC of 92.68 and BEDROC of 69.73 on DUD-E, and an AUROC of 61.97 and BEDROC of 11.51 on LIT-PCBA, showing only marginal improvements over our [CLS]-based approach.

Importantly, computational efficiency is a key consideration for virtual screening, which often involves searching millions of candidate molecules. While atom-level alignment can capture finer-grained interactions, it introduces substantial computational overhead. Specifically, we observed an approximately 70% reduction in predictor-stage throughput compared with the lightweight [CLS]-based predictor, significantly increasing the computational cost of large-scale screening.

Given the limited performance gains and substantially increased computational cost, we adopt global [CLS]-based alignment with a lightweight predictor, providing a more favorable trade-off between screening accuracy and efficiency.

Q3: Sensitivity of docking results

Following the standard CrossDocked2020 evaluation protocol, Vina Dock scores are obtained by docking generated molecules into their corresponding protein pockets using AutoDock Vina, which searches for favorable binding poses and reports the predicted binding affinity (kcal/mol). Lower scores indicate more favorable predicted binding.

For evaluation, we generate 100 molecular samples for each of the 100 test pockets, resulting in 10,000 samples per run. To assess sensitivity to random sampling, we additionally conducted experiments with three independent random seeds. The resulting mean Vina Dock score was −8.12 ± 0.06 kcal/mol (mean ± standard deviation), indicating low variability across independent runs and supporting the robustness of our reported docking performance.

Thank you again for your detailed review. We appreciate your assessment that “the overall framework is plausible,” while taking seriously your concern that the approach may be incremental. Your questions about JEPA framing, latent calibration, atom-level alignment, and docking sensitivity helped us identify where clearer explanations and additional evidence were needed.

Our response explains the multimodal latent-prediction motivation and why the shared binding representation enables recognition and generation to benefit each other, rather than merely combining two independently trained models. The revised manuscript now incorporates the additional experiments and their corresponding discussion: Section 4.3 and Table 3 evaluate docking-based specificity, Appendix E.4 examines explicit generated–conditioning alignment, Appendix E.5 quantifies the accuracy–throughput trade-off of atom-level alignment, and Appendix E.6 reports docking variability across three independent seeds. These analyses complement the training-component ablations in Table 4 and help separate the effects of task coupling, alignment choices, and sampling variability.

We have selectively highlighted important additions in blue. We hope our response and the revised evidence address your concerns about the contribution and methodological choices. We would appreciate your reconsideration of the paper in this light, and would be glad to clarify any remaining point.

Reviewer_wuu3OpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the problem of drug discovery and presents a method that integrates the two dominant ML approaches, virtual screening and generative molecule models. The authors propose a method called PL-JEPA that is based on a binding embedding, a latent representation of the molecular characteristics o a binding pocket. The binding embedding provides a stable dimensional interface across pockets and ligands and preserves both the predictive information and the generative information. PL-JEPA predicts the binding embedding in ligand representation space and iteratively refines it through predictive alignment, representation-conditioned generation, and joint training. An experimental evaluation was performed on the CrossDocked2020 benchmark, comparing PL-JEPA against four virtual screening methods and four generative methods.

Critical strengths

Significance of the problem

  • Drug discovery is a ML application of great practical significance.
  • While there has been a lot of progress with both virtual screening and with generative molecule models, there are still various directions for future research.

Originality of the approach

  • The approach integrating virtual screening and molecule generation is original. I am unaware of related work adopting such an approach.
Critical weaknesses

Serious concerns regarding the design of the method

  • The retrieval measures the similarity between the binding embedding and embeddings produced by a ligand encoder. But why should the embedding of the ligand be similar to that of the binding (pocket)?
  • Generation decodes the binding embedding into a novel 3D ligand via an equivariant diffusion decoder. It seems that that way the diversity of molecules generated for one pocket is quite limited.
  • The generative approach considers only binding affinity and ignores other important factors for drug candidates including toxicity and synthesizability that are considered by SOTA RL-based approaches.
 Torren-Peraire, Paula, et al. "Multi-objective optimization in the context of generative chemistry." Nature Communications 17.1 (2026): 9405.
 Papidocha, Sven M., et al. "The elephant in the lab: synthesizability in generative small-molecule design." Current Opinion in Chemical Engineering 51 (2026): 101217.

Unconvincing experimental results for molecule generation

  • Recent SOTA methods such as those listed under my questions need to be compared. Methods such as DiffSBDD and CGFlow report lower Vina scores (< -9.0) and greater diversity (> 0.85).
  • Synthesizability of generated molecules is a key factor for the deployment of generative molecule models. The AiZynth success rate is a metric that is more realistic than SA and should be measured. More generally, the synthesizability of the resulting molecules needs to be analyzed much more carefully.
Questions for authors
  1. The retrieval measures the similarity between the binding embedding and embeddings produced by a ligand encoder. But why should the embedding of the ligand be similar to that of the binding (pocket)?

  2. How do your results for molecule generation reported in table 2 compare to those of more recent methods such as the following ones:
Shen, Tony, et al. "Compositional flows for 3D molecule and synthesis pathway co-design." arXiv preprint arXiv:2504.08051 (2025).
Nakamura, Shogo, Nobuaki Yasuo, and Masakazu Sekijima. "A Transformer for Reaction-Aware Compound Explorations with GFlowNet in QSAR-Guided Molecular Design." Journal of Chemical Information and Modeling 66.9 (2026): 5101.
Lo, Alston, Connor Coley, and Wojciech Matusik. "A genetic algorithm for navigating synthesizable molecular spaces." International Conference on Learning Representations. Vol. 2026. 2026.

  3. How do the methods listed in table 2 compare in terms of AiZynth success rate?

5 comments

Thank you for your detailed review. Part 1 addresses the representation design, diversity, and practical relevance; Part 2 provides the generator and synthesizability comparisons.

W1 / Q1: Why should the embeddings be similar?

Our retrieval follows the CLIP-style cross-modal alignment paradigm: paired supervision aligns representations from different modalities, making embedding similarity meaningful without requiring resemblance in the original input spaces[1].

Importantly, we do not directly compare raw pocket and ligand embeddings. Our predictor extracts and transforms ligand-relevant information from the pocket representation into the ligand representation space (§3.2.2). The resulting binding embedding represents molecular characteristics implied by the pocket, while the contrastive objective explicitly aligns compatible pocket–ligand pairs (Eq. 4). Thus, similarity reflects learned binding compatibility, not structural resemblance.

[1] Learning Transferable Visual Models From Natural Language Supervision, ICML21

W2: Does a fixed embedding limit diversity?

A fixed condition does not imply a fixed output. Conditional diffusion learns a distribution, not a deterministic condition-to-output mapping—as exemplified by text-conditioned image generation [2] and pocket-conditioned molecular generation [3]. Likewise, our decoder models pθ(M∣S^y)p_\theta(\mathcal M\mid\hat S_y): the binding embedding remains fixed, while different noise realizations produce different molecular samples.

We also evaluate diversity empirically rather than relying on stochasticity alone. Our metric measures pairwise Tanimoto distance among molecules generated for the same pocket (§4.3). PL-JEPA achieves 0.78, compared with 0.69–0.73 for the evaluated baselines (Table 2). Figure 8 further illustrates structurally distinct molecules generated for fixed pockets, directly addressing the proposed diversity limitation.

[2] High-Resolution Image Synthesis with Latent Diffusion Models, CVPR2022

[3] 3D Equivariant Diffusion for Target-Aware Molecule Generation and Affinity Prediction,ICLR23

W3: Practical relevance, specificity, and synthesis-aware generation

We agree that affinity alone is insufficient for practical drug discovery, and appreciate the suggested perspectives. Our original evaluation follows the widely-used SBDD metrics. To better assess practical relevance, we additionally examined binding specificity and synthesizability.

PL-JEPA improves docking-based specificity, rather than merely favoring molecules that dock well to many pockets. Target specificity is important for avoiding unintended interactions in drug discovery. Following Gao et al. [4], we evaluate Delta Score, which measures preference for the intended pocket over non-target pockets (using Glide, different from the main paper’s Vina scores). PL-JEPA leads the evaluated generative baselines across all five metrics.

MethodsAbsolute mean ↑Delta mean ↑Delta median ↑Ratio mean ↑Ratio median ↑
Trainset5.7270.044-0.0730.1590.053
Reference (Original)6.6321.1581.029--
AR5.7370.3930.2000.1500.039
Pocket2Mol5.9460.4370.1060.1700.050
DrugGPS6.5540.4900.3870.2350.134
TargetDiff6.6650.3350.1020.2590.129
DecompDiff7.1020.3540.2200.2740.133
SBE-Diff6.8150.5520.2500.3000.216
MolCRAFT7.5230.5830.3510.3550.316
PL-JEPA8.4210.6120.3970.3880.353

This advantage is consistent with our training design: contrastive alignment distinguishes matched from mismatched pocket–ligand pairs, while joint generative training incorporates this compatibility information into molecular generation.

Crucially, synthesizability is not a limitation inherent to PL-JEPA: our framework can incorporate—and improve upon—a synthesis-aware generator. To directly address this concern, we replaced the UniGEM-based decoder with pretrained CGFlow, which jointly designs molecules and synthetic pathways.

The generator and AiZynth comparisons are provided in Part 2 below.

We find the binding-aware learning improves docking performance and diversity without sacrificing retrosynthetic accessibility. PL-JEPA therefore complements advances in synthesis-aware generation, providing a framework in which molecular recognition and practical generation constraints can benefit from joint learning.

[4] Rethinking Specificity in SBDD: Leveraging Delta Score and Energy-Guided Diffusion

This is Part 2 of our response to W3–W5 and Q2–Q3.

W3 / W4 / Q2: Comparison with recent generative methods

PL-JEPA is a general framework that can incorporate stronger generative models. Using pretrained CGFlow as the decoder, PL-JEPA improves Vina dock and diversity, while preserving retrosynthetic accessibility. This demonstrates that advances in molecular generation can be integrated into PL-JEPA and further enhanced through binding-aware joint learning. We will add this result in our main paper.

MethodVina Mean ↓Vina Median ↓SA ↑Diversity ↑AiZynth Avg. ↑AiZynth Med. ↑
Pocket2Mol-7.15-6.790.740.6929.1%22.0%
TargetDiff-7.80-7.910.580.729.9%3.2%
DecompDiff-R-7.03-7.160.660.730.9%0.0%
MolCRAFT-7.92-8.010.690.7216.5%9.1%
CGFlow (3DSynthFlow, high β)-9.42-9.61—0.6936.1%36.0%
PL-JEPA (CGFlow Decoder high β)-9.57-9.650.750.7337.2%36.7%

W5 / Q3: AiZynth success rates

The AiZynth results for the methods in the generation comparison are listed separately below.

MethodAiZynth Avg. ↑AiZynth Med. ↑
Pocket2Mol29.1%22.0%
TargetDiff9.9%3.2%
DecompDiff-R0.9%0.0%
MolCRAFT16.5%9.1%
DiffSBDD2.9%2.0%
PL-JEPA (Ours)10.4%6.3%
CGFlow (3DSynthFlow, high β)36.1%36.0%
PL-JEPA (CGFlow Decoder high β)37.2%36.7%

Thank you again for your detailed review, particularly your recommendations to assess practical relevance beyond docking affinity and to compare with synthesis-aware generative methods. We agree that target selectivity and synthesizability deserve explicit evaluation.

The additional experiments and corresponding discussion from our response are now included in the revised manuscript. Section 4.3 now includes the Glide-based specificity results, with Table 3 reporting the full comparison. PL-JEPA leads the evaluated generative methods across all five benchmark metrics, supporting improved docking-based pocket selectivity rather than affinity alone. Appendix E.2 provides the CGFlow-decoder comparison and AiZynth results, explicitly distinguishing the original PL-JEPA decoder from the synthesis-aware variant. The latter improves docking and diversity relative to CGFlow while preserving retrosynthetic accessibility. Our earlier response also explains why learned cross-modal compatibility is meaningful and why a fixed conditioning embedding can support diverse stochastic outputs.

Important additions are selectively marked in blue. We hope the new evidence addresses your concerns about diversity and practical relevance. Would you consider re-evaluating the paper in light of these results and clarifications? We welcome any remaining questions or concerns and would be happy to continue the discussion.

Reviewer_urH2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors adapt a Joint Embedding Predicive Architecture (JEPA) to protein ligand complexes to connect virtual screening and molecule generation in one model. It uses separate pocket and ligand encoders based on the Uni-Mol architecture, to learn a shared 512-dimensional “binding embedding”, on which a equivariant diffusion decoder is conditioned on. They train the model in 3 stages, to a) align the prediction of both pocket and ligand encoder, similar to DrugCLIP, b) pretrain the decoder on a frozen embedding, reconstructing the ligands coordinates and atom types, and c) jointly update all modules. The model therefore supports DrugCLIP-style ligand screening consistently reported improvements in AUROC, BEDROC and EF1% on VS benchmarks, while also supporting molecular generation, demonstrating favorable docking scores and diversity on CrossDocked.

Critical strengths

The paper presents a coherent integration of molecular retrieval and generation, supported by strong reported screening gains and informative ablations.

  • Applying JEPA to molecule generation and screening is a useful addition to the current frameworks, leading to a shared “binding embedding” that supports screening, generation, and reverse retrieval of targets.
  • The reported VS gains are substantial, outperforminc DrugCLIP consistently.
  • Feasibility of molecular generation is demonstrated in Table 2, as generated molecules have competitive docking scores.
  • The paper ablates decoder pretraining and calibration, supporting that the joint training helps balancing ligand and pocket contribution.
Critical weaknesses
  • It would be useful to re-encode generated molecules and measure their similarity to the conditioning embedding and reference-ligand embedding, to investigate the benefit of the joint embedding. Comparing generated-to-reference fingerprint similarity distributions against a pocket-conditioned baseline such as MolCRAFT would help to check whether joint learning changes the ligand related information expressed in generated samples.
  • The evaluation focusses on absolute properties. A comparison of generated and reference molecule properties, like molecular-size, and selected geometric distributions, would help to evaluate the distribution learning capability of the model.
Questions for authors
  • For molecular generation, a standard CrossDocked2020 split applied only to decoder training would not prevent exposure to test pockets in Stage I retrieval alignment or Stage III joint training. Could the authors document used data splits across all stages?
  • How closely do generated molecular and geometric distributions match the reference data?
3 comments

Thank you for your constructive review. We address W1, W2/Q2, and Q1 below.

W1: Embedding consistency and fingerprint similarity

We conducted additional experiments to analyze embedding consistency, fingerprint similarity, and the effects of explicit latent alignment.

(1) Embedding similarity. After joint training, generated–conditioning and generated–reference cosine similarities changed from 0.3285/0.3220 to 0.2626/0.2877, while conditioning–reference similarity slightly increased from 0.5080 to 0.5135. Importantly, generated–conditioning similarity remained substantially higher than the random-pair baseline (0.0712), indicating that generated molecules retain meaningful conditioning information.

(2) Fingerprint similarity. PL-JEPA achieved substantially lower generated-to-reference fingerprint similarity (mean/median: 0.0793/0.0784) than MolCRAFT (0.2258/0.2086), suggesting greater structural novelty and broader chemical space exploration.

(3) Latent alignment ablation. We additionally introduced a loss explicitly encouraging generated–conditioning alignment. Although similarity increased from 0.2626 to 0.3792, diversity decreased from 0.78 to 0.64, Vina slightly worsened from −8.12 to −8.02 kcal/mol, and virtual screening performance dropped by approximately 15%.

These results suggest that higher embedding similarity does not necessarily improve downstream performance. Virtual screening primarily relies on relative ranking, while molecular generation benefits from structural diversity. Thus, PL-JEPA seeks to balance binding-related representation learning and chemical diversity rather than maximizing absolute embedding alignment.

W2 / Q2: Molecular size and bond-length distributions

Molecular size. The number of heavy atoms is sampled from a predefined conditional distribution estimated from the CrossDocked training set, following the same approach used by MolCRAFT and TargetDiff. Thus, molecular size is determined by this sampling distribution rather than predicted by the diffusion decoder. The generated molecules have a mean heavy-atom count of 22.7.

Bond-length distributions. We compare the reference and generated molecules using Jensen–Shannon divergence of the six frequent bond-length distributions (lower is better). “–”, “=”, and “:” represent single, double, and aromatic bonds, respectively.

BondPocket2MolTargetDiffDecompDiff-RPL-JEPA
C–C0.4940.3690.3260.252
C:C0.4140.2630.2380.151
C–O0.4520.4210.3460.274
C–N0.4220.3620.3370.296
C:N0.4840.2350.2350.187
C=O0.5100.4610.3910.319

As discussed in the paper, this benefits from our framework’s ability to pretrain the decoder on small-molecule data, allowing it to learn molecular geometry before pocket-conditioned generation.

Q1: Data splits

We use LigUnity data, which have been deduplicated against DUD-E and LIT-PCBA. We cluster all proteins with MMseqs2 at 50% sequence identity and remove data in the same clusters as CrossDocked from virtual-screening training.

Thank you again for your careful review and encouraging overall assessment. Your questions about re-encoding generated molecules, their relationship to the conditioning and reference embeddings, and molecular geometry helped us make the empirical analysis more informative.

We have now incorporated the additional experiments and their corresponding discussion from our response into the revised manuscript. Appendix E.1 reports molecular-size information and bond-length distributions, while Appendices E.3–E.4 present embedding diagnostics, fingerprint similarity, and the explicit-alignment ablation. Section 4.3 and Section 4.5 point readers to these analyses. They clarify how generated molecules retain conditioning information while allowing structural diversity, and why maximizing generated–conditioning similarity does not necessarily improve downstream performance. Our earlier response also specifies the data-split and filtering procedures relevant to your question.

Important additions are selectively highlighted in blue for ease of review. We hope these analyses and clarifications address your concerns. If any point remains unclear, we would be happy to discuss it further. We would appreciate your re-evaluation of the paper in light of the updated evidence.

Thank you for your the additional experiments. The additional analysis on re-embedding and the distribution analysis addresses my concerns. It also highlights that PL-JEPA can be integrated with other decoders, such as CGFlow. I find the formulation of a shared pocket and ligand representation reasonable, and empirical results demonstrates it’s feasibility, I therefore maintain my recomondation.

Comments on the paper

2 comments

Dear Reviewers,

The authors have now responded to the questions and concerns raised by everyone. Could I kindly ask you to review the authors' responses, take a look at the comments from the other reviewers, and actively participate in the discussion?

Thank you all very much for your valuable time and contributions!

Best regards,

The AC

Dear Reviewer urH2 and B43N,

As the rebuttal period is coming to an end, I kindly remind you to check the authors' responses. Please indicate whether your concerns have been adequately addressed. Many thanks!

Best,

AC

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

Related papers

Discussion (0)

Sign in to comment.