- 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
- Learning informative molecular representations for search and conditioning generative models sounds technically and is critical in structure-based drug discovery.
- The writing is clear and easy to follow.
Critical weaknesses
- 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.
- 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
- Could you elaborate on the latent calibration objective?
- Did you try aligning atom-level representations, instead of global representations via [CLS] tokens?
- How sensitive are the docking results?
5 comments
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.
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.
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