AANet: Virtual Screening under Structural Uncertainty via Alignment and Aggregation

W1: No standard deviations or error bars.

We did not include standard deviations initially because most prior studies do not report results or offer pre-trained models from multiple runs, making direct comparison of variability infeasible. However, to address this concern, we have conducted four parallel runs of our method on the DUD-E benchmark to estimate variability. The resulting standard deviations are now included in Table 1.

Table 1. Performance (mean ± std) on the DUD-E benchmark.

We hope this addition improves the clarity and reproducibility of our reported results.

W2: Broader method comparison and related work discussion.

We thank the reviewer for the suggestion. Since we evaluate our model on newly curated benchmark settings, all baselines had to be re-implemented or re-run, which limited the scope of direct comparisons. Some previously published methods used in related studies either lack publicly available inference code or are outdated. We include this comparison with previously reported results in Supplementary Section C.2.

We have further expanded our comparisons to include the following two categories:

• Several representative ligand-based virtual screening (LBVS) methods (ECFP, USR, PhaseShape from the Schrödinger suite):

[1] J Comput Chem. 2007, 28, 10, 1711-23. [2] J. Chem. Inf. Model. 2011, 51, 10, 2455–2466. [3] J. Chem. Inf. Model. 2010, 50, 5, 742–754.

• A flexible docking method (FlowDock) based on flow matching.

[1] FlowDock: Geometric Flow Matching for Generative Protein-Ligand Docking and Affinity Prediction, ISMB 2025.

Due to computational constraints, FlowDock was evaluated on a subset of DUD-E compounds, which may underestimate its EF values (roughly halved). We acknowledge that flexible docking frameworks like FlowDock are not primarily designed for large-scale virtual screening tasks. Instead, they are better suited for post-screening re-ranking or refinement stages, where a small number of candidate actives are already identified. We will further enrich the related work section to reflect these comparisons and position our method more clearly in the context of existing approaches.

Q1: Pocket selection consistency during training?

We thank the reviewer for this question. During training, we randomly sample one candidate cavity on the fly from either the set of positive (overlapping with holo pocket) or negative (non-overlapping) Fpocket-detected pockets on the same protein structure. If no positive pocket is available, a negative one is used instead.

Importantly, all sampling is controlled by a random seed, ensuring reproducibility: for a given dataset, seed, and training setup, the same pocket will be selected for the same sample at the same epoch across different runs. This design balances stochastic training dynamics with full reproducibility.

Q2: Could the method benefit from more augmentations (e.g., randomized positions, alternative predicted structures)?

We thank the reviewer for this valuable suggestion. Due to computational constraints, we were unable to explore such augmentations in this work. However, we agree that these strategies are promising, especially for future engineering-scale models.

1. Disturbed pocket positions may serve as an effective regularization strategy. However, fully random centers could undermine the spatial rationale defined by pocket detection tools, which often capture protein-intrinsic cavity features.
2. Using diverse predicted structures is a highly promising augmentation. Recent studies (e.g., shallow MSA-based predictions [1], AF-Cluster [2]) show that AlphaFold can generate alternative conformations representing distinct functional states (e.g., DFG-in/out in kinases; transporters and G-protein-coupled receptors). Such sampling could help the model learn flexible pocket features. Similarly, including structures from alternatives such as ESMFold or Boltz1 may improve generalization by exposing the model to distributional variability in predicted structures.

We believe these augmentations are valuable actions for improving robustness under structural uncertainty.

[1] Elife, 2022, 11, e75751. [2] Nature, 2024, 625(7996), 832–839.

Q3: Clarification on training with pocket-agnostic data: staged vs. joint optimization? How is λ chosen?

We thank the reviewer for pointing this out and will improve the clarity in the final version.

Our training follows a two-stage design:

1. Stage 1 (pocket-aware): We train a dual-tower model with contrastive loss on paired pocket–ligand data.
2. Stage 2 (pocket-agnostic): We freeze the backbone encoders and introduce lightweight adapters—a cross-attention adapter on the pocket side (approximating an argmax match) and an identity-like linear adapter on the ligand side. This preserves the model’s ability to identify binding pockets from the original contrastive alignment, while enabling it to effectively leverage pocket-agnostic data for further learning.

This design allows the model to reuse DrugCLIP-style embeddings and enables efficient inference while enhancing learning from broader activity/binding datasets.

As for the λ setting, we use a simple value of 1.0. At this scale, the contrastive loss is naturally two orders of magnitude larger than the KL loss, providing sufficient regularization without dominating training. A larger λ overly constrains the model (resulting in little learning), while a smaller λ reduces its regularizing effect.

Q4: Did the authors account for structural similarity when removing overlapping entities with DUD-E or LIT-PCBA to prevent data leakage?

We thank the reviewer for this important question. We acknowledge that ID-based filtering does not eliminate structurally similar entries. We followed this approach primarily to align with prior work and ensure fair comparisons. Notably, some baselines did not apply such filtering at all (cf. Supplementary Section B.7).

To further mitigate potential leakage, we performed an additional analysis using MMseqs2 to identify and exclude any training proteins (from PDBbind and ChEMBL) that contain chains with ≥90% sequence identity to any DUD-E structure or target sequence. We then compared our method and DrugCLIP under both the original de-redundant and this stricter de-homologous setting. As shown in Table 4, our method maintains strong performance even after removing homologous sequences, demonstrating superior generalization.

Table 4. Performance on DUD-E under de-redundant and de-homologous (90%) settings.

Q5: Could training a model on pockets randomly sampled from multiple detection methods improve generalization?

We thank the reviewer for this suggestion. We investigated this possibility and found the following:

1. The slight performance drop observed with certain pocket detection tools largely stems from individual tool failures on specific targets, especially on AF2-predicted structures. For example, PocketFinder failed to detect the correct pocket for IGF1R, PTPN1, and RENI (IoU < 0.1); SurfNet failed on EGFR, IGF1R, RENI, and RXRA. In contrast, Fpocket consistently detected valid pockets, making it more robust overall.
2. In early-stage experiments (Stage 1 model only, without adapters), we augmented training data using positive pockets from three other detection tools. This yielded no consistent improvement: for example, EF@1% on the AF2-annotated setting dropped slightly (34.4 → 33.7), while AF2-blind improved modestly (29.3 → 30.7). Similar minor improvements were observed across apo settings. Given AANet’s demonstrated robustness across tools, we believe the model has already internalized spatial priors shared across detection algorithms, reducing the benefit of this augmentation strategy.

W1: The paper does not provide information about the content and preprocessing steps of key datasets, such as ChEMBL35, used during training. This limits the transparency of the training process.

We thank the reviewer for pointing this out. We will include the following details in the final version for clarity.

The ChEMBL35 dataset was processed as follows:

Molecule filtering:

• Removed salts or kept only the largest fragment
• Kept molecules with molecular weight between [100, 800]
• Removed those containing atoms other than [H, C, N, O, F, Cl, Br, I, S, P, B, Se]
• Filtered out molecules with unbranched long chains containing 6 atoms or more.

Activity filtering: We retained activity records with:

• Confidence score = 9
• Assay type = functional or binding
• Standard type ∈ ['Ki', 'IC50', 'Kd', 'EC50', 'ED50', 'AC50', 'XC50']
• Values converted to molar units and filtered to lie within the –log10 range [5, 12].

Target mapping: Protein targets were mapped to UniProt IDs and matched with AlphaFold structures. Targets without available predicted structures were discarded.

Contrastive supervision masking: Each known protein–ligand activity pair was recorded using its UniProt ID and InChI-key. During training, we masked non-diagonal entries that correspond to active (positive) pairs to ensure that true positive pairs are never treated as negatives, even if they are not aligned in the current batch. This masking strategy ensures correct supervision and avoids misleading the model during contrastive learning.

W2: In Appendix Table S8, AANet underperforms DrugCLIP on the full LIT-PCBA dataset, despite showing strong results on curated subsets. This raises concerns about its generalization ability and suggests that the model may rely on cavity augmentation.

We thank the reviewer for the observation. We would like to clarify the following:

Marginal difference and metric interpretation: The performance gap is marginal (EF@1%: 5.51 vs 5.13), and our model outperforms DrugCLIP in BEDROC (0.0677 vs. 0.0623), which better reflects early enrichment — a critical criterion for practical screening.

Unconventional binding sites in LIT-PCBA: LIT-PCBA contains atypical targets such as mTORC1 and PKM2, where the ligand binds at oligomeric interfaces (hetero- or homo-dimers). Our method uses monomeric AlphaFold structures for second-stage training, which may reduce performance on these targets. Notably, our model trained only on PDBbind still outperforms DrugCLIP on the full LIT-PCBA dataset (EF@1%: 5.86 vs. 5.51; BEDROC: 0.0699 vs. 0.0623), suggesting the performance is not dependent on cavity augmentation.

Broader generalization: Our method consistently outperforms DrugCLIP on a variety of datasets, including DUD-E, its curated subsets, LIT-PCBA subsets, and the newly added DEKOIS benchmark. This demonstrates strong generalization beyond any specific data augmentation strategy.

Table 1. Performance on DEKOIS (77 targets) under holo, apo-pred-annotated (the cavity detected by FPocket that is closest to the holo pocket), and apo-pred-blind settings.

Q1: Is there any discussion of ligand-based virtual screening methods (LBVS)? For example, traditional LBVS methods (SHAFTS) and recent deep learning-based approaches (S-MolSearch) can bypass the need for protein structural information entirely.

We thank the reviewer for highlighting the relevance of ligand-based virtual screening (LBVS), which can indeed bypass the need for protein structural information and is particularly useful when high-quality ligands are available or when protein structures are unreliable.

We would like to clarify that LBVS and SBVS address different use cases. LBVS is often more straightforward when good query ligands are available (e.g., potent, exogenous ligands), but may be less effective in identifying novel chemotypes or in the absence of relevant prior ligands. In contrast, SBVS leverages the protein’s structural context, which can help discover structurally diverse actives but depends on structure quality.

Regarding the methods mentioned:

• SHAFTS is relatively outdated and unfortunately could not be executed on our current hardware.

• S-MolSearch does not provide pre-trained checkpoints, training data, or inference code, making reproduction infeasible at this time.

However, to address the reviewer’s concern, we evaluated several widely-used LBVS methods, covering both 2D and 3D paradigms, including ECFP (2D topology-based), USR (3D shape-based), and Shape from Schrödinger suite (3D shape overlap). The results are summarized in Table 2.

Table 2. Performance of common LBVS methods on DUD-E benchmark

[1] J Comput Chem. 2007, 28, 10, 1711-23. [2] J. Chem. Inf. Model. 2011, 51, 10, 2455–2466. [3] J. Chem. Inf. Model. 2010, 50, 5, 742–754.

The notably strong performance of ECFP4 is likely due to the fact that many actives in virtual screening benchmarks such as DUD-E share similar topological scaffolds, whereas decoys are specifically designed to have similar physicochemical properties but divergent topologies. As a result, ECFP4 can effectively distinguish actives from decoys based on topological similarity alone, leading to higher enrichment factors. However, such advantages may not always generalize to real-world screening scenarios, where novel actives are often needed.

Q2: Is removing overlapping UniProt IDs sufficient to prevent information leakage between training and test sets? How do the authors ensure that models are not indirectly exposed to structurally or functionally similar targets? L3: The data split removes overlapping UniProt IDs but does not filter by sequence similarity.

We thank the reviewer for raising this important concern regarding potential overlap between training and test data.

To further rule out information leakage, we performed an additional experiment using MMseqs2 to remove any proteins in our training sets (PDBbind and ChEMBL) that contain chains sharing ≥90% sequence identity with any DUD-E target. We then compared performance under both the original de-redundant setting and this stricter de-homologous setting. As shown in Table 3, our method exhibits only a minor drop in performance, maintaining strong results across all settings. This demonstrates the robust generalization of our approach, even in the absence of closely related training proteins.

Table 3. Performance on DUD-E under de-redundant and de-homologous (90%) settings.

L1: The model relies on external cavity detection (e.g., Fpocket). If the cavity predictions are poor—especially in complex or flexible proteins—the model’s performance may degrade.

We thank the reviewer for raising this important point.

1. Robustness across cavity detection tools: Our method does not rely on any specific cavity detection tool. As shown in Table 4, AANet demonstrates consistent performance across different pocket detection methods.

2. Limitation in flexible or disordered proteins: We acknowledge that for highly flexible or intrinsically disordered proteins, where well-formed pockets are absent and ligand-induced fit is essential, existing pocket-based SBVS methods, including ours, face inherent challenges. Our work primarily addresses uncertainty in binding site localization, not full flexible binding modeling.

We agree this is a critical direction, and we consider end-to-end pocket prediction and flexible modeling promising avenues for future research.

L2: No confidence or uncertainty estimation in predictions.

We appreciate the reviewer’s insightful comment and agree that quantifying prediction confidence would enhance the method’s practical value. While AANet does not currently produce explicit uncertainty estimates, its design of modeling interactions with both binding and non-binding cavities naturally provides multiple levels of structural signal that could support future confidence modeling. Leveraging this structure-aware framework for uncertainty estimation is one of our priority directions for future work.

W1: Even though the results are strong on DUD-E and LIT-PCBA, the method is not tested on more diverse or unbiased datasets, such as DEKOIS, or in real-world drug discovery tasks.

We thank the reviewer for pointing this out. Due to time constraints, we were unable to identify and verify the experimental apo structures corresponding to DEKOIS targets. However, we mapped the DEKOIS targets to UniProt ID and successfully downloaded 77 target structures from the AlphaFold Database, excluding 4 virus-related targets that are not available (HIV1PR & HIV1RT: P03366, SARS-HCoV: P0C6X7, NA: P27907).

We evaluated our method and baselines under three settings:

• Holo: using the known holo pocket.

• Apo-pred (annotated): pocket detected near the ligand (used only for annotation).

• Apo-pred (blind): fully blind screening setting.

As shown in Table 1, our method consistently outperforms DrugCLIP across all settings, and surpasses traditional docking baselines under holo pocket settings (italic) when tested under the most challenging apo-pred blind setup (bolded).

Table 1. Performance on DEKOIS (77 targets) under holo, apo-pred-annotated (the cavity detected by FPocket that is closest to the holo pocket), and apo-pred-blind settings.

W2: The model relies on external tools like Fpocket to find candidate pockets, which prevents end-to-end learning and may cause biases depending on the tool used.

We thank the reviewer for this insightful comment. While our current framework employs Fpocket for cavity detection during training, we would like to clarify and justify this design with the following points:

1. Robustness across different cavity detectors: Although our model uses Fpocket to extract training cavities, the downstream performance remains stable across different cavity detection tools (cf. Table 4). These conventional detectors are grounded in physicochemical definitions of binding pockets, such as surface concavity, void volume, and geometric constraints. The consistency across tools suggests that AANet is not overfitting to a specific detection algorithm, but instead learns underlying structural priors intrinsic to proteins, such as spatial enclosures or clefts.

2. Alignment with practical structure-based pipelines: Our method follows the typical structure-based virtual screening pipeline, where pocket detection and binding assessment are distinct stages. This modularity aligns with standard workflows in drug discovery and allows our method to plug into existing infrastructure while remaining adaptable to alternative cavity detectors.

3. Physically meaningful supervision: By grounding learning on cavities defined via physical heuristics (e.g., void detection), we introduce inductive bias from medicinal chemistry intuition into the model. This helps prevent overfitting to spurious binding patterns and encourages the model to focus on biophysically plausible pockets. Concerns about physical plausibility have recently gained attention in the structure prediction and docking communities, as highlighted by works such as PoseBusters in deep learning-based docking. This emphasis on biophysical realism is equally relevant to virtual screening and motivates the design of our method.

Although we currently do not explore end-to-end cavity prediction due to time and resource constraints, we agree that learning pocket detection in an end-to-end fashion is a promising direction, especially on novel targets including flexible proteins. As datasets grow and modeling techniques evolve, we believe this integration could offer even more robust and generalizable solutions.

W3: Lack of comparison with recent methods like FlowDock and ProFSA.

While we appreciate the reviewer’s suggestion regarding FlowDock and ProFSA, which enrich the diversity of baseline methods, we would like to clarify that flexible docking frameworks like FlowDock are not primarily designed for large-scale virtual screening tasks. Instead, they are better suited for post-screening re-ranking or refinement stages, where a small number of candidate actives are already identified.

Flexible docking excels at capturing subtle conformational changes and fine-grained interactions between a specific ligand and a pocket, making it valuable for discriminating between a few similar binders. However, in large-scale virtual screening where the key objective is to enrich true actives from a vast set of mostly inactive molecules, such methods are less effective when screening against millions of decoys, as required in virtual screening settings. We will include a discussion of flexible docking methods like FlowDock in the Related Work section of the paper to provide a more complete picture in the context of apo setting.

To address the reviewer’s concern, we included results from DrugCLIP (ProFSA) using the checkpoint publicly released in the supplementary data of bioRxiv 2024.09.02.610777, which was fine-tuned from ProFSA with the DrugCLIP task. Interestingly, DrugCLIP (ProFSA) underperforms the original DrugCLIP baseline in our virtual screening benchmark. We hypothesize that this may be due to its pseudo-ligand-based pretraining, which may have enhanced the model’s ability to recognize general protein–ligand interaction patterns but also made it more prone to misinterpreting peptide-binding interfaces as small-molecule binding pockets.

Although FlowDock’s computational cost prevented us from performing large-scale comparisons on the full dataset, we conducted a direct empirical evaluation on a representative subset. Specifically, we randomly sampled 10% of actives and 5% of decoys across 38 apo targets from DUD-E, which may approximately halve the number of Enrichment Factors (EF) reported. For efficiency, we generated only one ligand conformer per molecule (instead of the default five), which may degrade FlowDock’s performance. This reduced the total number of complexes to < 30k and allowed us to complete all FlowDock runs within 3 GPU days. These results are shown in Table 2.

In contrast, our method directly targets binder enrichment under structural uncertainty and is optimized for efficiency, robustness, and early enrichment metrics. Nevertheless, AANet significantly outperforms FlowDock across all key metrics on the subset, as shown below. We also acknowledge FlowDock’s strengths in generating high-quality complex structures, which may provide additional interpretability or utility in downstream stages of the drug discovery pipeline.

Table 2. Performance comparison on apo-blind virtual screening subset.

Q1: Can the authors clarify whether any target-level or structure-level overlap exists between training and test sets?

We thank the reviewer for raising this important point. On PDBbind, we removed duplicate complexes based on structure IDs, following prior work to ensure fair comparison. However, this may not fully eliminate target-level redundancy. For some of baselines, duplicates were not removed (cf. Supplementary Section B.7). On ChEMBL, we filtered data using UniProt IDs, which more reliably removes overlapping targets.

To further address potential leakage, we conducted an additional experiment by using MMseqs2 to identify and remove any training proteins (from PDBbind and ChEMBL) containing chains with ≥90% sequence identity to any DUD-E structure / target sequences. We then compared our method and DrugCLIP under both the original and this more stringent de-redundant / de-homologous setting. As shown in Table 3, our method maintains robust performance even after removing homologous sequences, demonstrating better generalization.

Table 3. Performance on DUD-E under de-redundant and de-homologous (90%) settings.

W1: Lack of discussion on computational efficiency

To address the reviewer's concern, we provide a direct comparison of overall computational cost among representative methods on the LIT-PCBA dataset, using a single NVIDIA A100 80GB GPU and a 128 vCPU server.

*: Cost that may be shared across subsequent virtual screening (VS) projects; **: One highest-resolution PDB structure was selected for each subset (cf. Supplementary Section B.6).

As our dual-tower backbone does not involve fusion prior to the adapter and supports embedding reuse, our method (AANet) introduces minimal overhead beyond DrugCLIP while providing significantly improved performance, particularly in unbound settings. The majority of runtime during the scoring phase arises from data loading and CPU–GPU communication, rather than model inference itself. On typical commercial-scale libraries, AANet maintains ultra-fast screening throughput comparable to DrugCLIP.

W2: Clarification on the 10 Å pocket extraction radius

We appreciate the reviewer’s concern regarding the selection of the 10 Å threshold for pocket extraction. This choice was based on quantitative analysis on the PDBbind dataset, supported by the following three key findings:

1. Coverage of original holo pockets: The oracle 10 Å cavities, which are defined as those with the highest overlap (IoU) with holo pockets, fully cover 75.38% of the original 6 Å holo pockets, and 90.94% achieve ≥ 85% coverage. This indicates that 10 Å cavities provide a reliable approximation of the ground-truth pocket regions.

2. Tradeoff between feature integrity and computational cost: The average atom count in 10 Å oracle cavities is 612.68 ± 240.75, over 3× larger than the 6 Å holo pockets (199.32 ± 55.04). Since computational cost scales non-linearly with the number of input atoms, we uniformly downsample pockets to ≤ 256 atoms during training to ensure fair comparisons across different pocket definitions. Larger regions would require more aggressive downsampling, which may compromise spatial features or increase computational overhead. Thus, 10 Å represents a practical upper bound that balances representational fidelity with efficiency.

3. Avoiding bias from over-extension: Extending the radius beyond 10 Å risks incorporating off-target cavities, especially on cluttered protein surfaces. As our 10 Å is defined by minimal atom-to-atom distance (not center-based), it already captures ample structural context. Further expansion would dilute the binding signal and potentially introduce irrelevant patterns.

Thus, the 10 Å threshold strikes a practical balance: it ensures high coverage of the ground-truth binding pocket, preserves structural integrity, and avoids unnecessary computational overhead or learning from irrelevant patterns.

Q1: What is the role of aligning with the 'holo pocket' in tri-modal contrastive learning?

We thank the reviewer for this insightful question. The alignment with holo pockets plays a critical and bidirectional role in our tri-modal contrastive framework, motivated by the following:

1. Holo → Cavity - Guidance for learning from noisy input: Holo pockets, shaped by actual ligand binding, serve as reliable supervisory anchors (as stated by the reviewer). Aligning detected cavities with holo pockets helps the model extract meaningful features even from noisy or imprecise cavity detection.

2. Cavity → Holo - Addressing structural uncertainty and generalization: Holo pockets are biased toward a specific ligand series(cf. Suppl. A.1-3), while other actives may bind elsewhere. By aligning cavities (which are ligand-agnostic) to holo pockets, the model learns features that are spatially intrinsic to the protein. This improves the model’s robustness to structural uncertainty and enhances generalization. This alignment also boosts performance on holo structures, since the learned features are less tied to ligand-specific conformations.

3. Cavity ↔ Holo - Dual modeling of general and specific signals: The model jointly learns ligand-independent (intrinsic to proteins) and ligand-specific pocket representations, which improves robustness. This enables stronger performance across both holo and apo structures, not just mitigating the performance gap but outperforming DrugCLIP in both cases.

In summary, holo-cavity alignment is essential for modeling both precise and generalizable binding patterns, supporting our goal of robust virtual screening under structural uncertainty.