acceptodds
Under review as a conference paper at ICLR 2027

Reliability-Calibrated Pose-by-Metric Fusion for Target-Local Virtual Screening

Abstract

Structure-based virtual screening ranks candidate ligands against a target protein by aggregating heterogeneous evidence from predicted protein-ligand complexes, structural confidence, docking affinity, interface geometry, interaction fingerprints, and ligand descriptors, yet within-target ranking quality remains decisive for downstream hit rates. Two difficulties complicate this aggregation: binding is pose-dependent with uncertain pose quality at inference time, and supervised ranking signals are sharply non-uniform across targets, with metric reliability also varying per target. We propose **PoT-Rank** (**Po**se-**T**rust Ranking), a target-local learning-to-rank framework that represents each candidate as a masked pose-by-metric tensor and learns hierarchical softmax weights over feature suites, metrics, and poses. Metric trust is calibrated without label leakage through a training-only reliability gate based on same-target pairwise concordance, paired with a target-normalized hard-decoy curriculum. To address differing statistical risks, PoT-Rank splits into a dense pose-aware ranker with a target-balanced pairwise hinge loss and a sparse antisymmetric comparator with signed-vote synthesis, with all stateful components frozen before evaluation. Under a unified frozen protocol with protein-side MMseqs2 splits at 40% identity, PoT-Rank attains the strongest within-target ranking in both regimes, reaching within-target pairwise accuracy 0.6602 and Spearman 0.5343 on the dense split (3 targets, 565 within-target pairs) and 0.6897 and 0.4313 on the sparse split (79 targets). It surpasses classical scoring functions, deep affinity predictors, graph-based models, pose-aware rescorers, and structure-prediction-based affinity models, and matches or exceeds strong tabular baselines in both regimes, while remaining stable under leakage audits, multi-seed perturbation, and corrupted-channel stress tests. Our code is available at https://anonymous.4open.science/r/PoT-Rank-4CD1/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Virtual Screening, Protein-Ligand Binding, Learning To Rank, Pose-Aware Ranking, Reliability Calibration, Target-Local Ranking

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_iXKxOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the problem of within-target ligand ranking in structure-based virtual screening, addressing pose uncertainty, target-dependent metric reliability, and heterogeneous supervision density across protein targets. To tackle these challenges, the paper propose PoT-Rank, a target-local learning-to-rank framework that integrates multiple protein–ligand poses and heterogeneous scoring metrics through hierarchical feature fusion, reliability calibration, and hard-decoy curriculum learning, while employing a separate pairwise comparator for sparsely supervised targets. To evaluate its effectiveness, the authors conduct experiments using protein-level MMseqs2 splits at 40% sequence identity, comparing against classical docking scores, deep affinity predictors, graph-based models, pose-aware methods, and tabular baselines, alongside ablation studies, leakage audits, and robustness analyses. The proposed framework achieves within-target pairwise accuracies of 0.6602 on three dense test targets and 0.6897 on 79 sparse test targets, outperforming the evaluated domain-specific baselines in the dense setting and matching XGBoost's pairwise accuracy in the sparse setting.

Critical strengths

Instead of simply performing structure-based virtual screening using various information, such as structural confidence (the reliability of predicted complex structures), docking affinity (energy-based scores of binding poses), interface geometry (protein–ligand contacts and surface area), interaction fingerprints (hydrogen bonds, hydrophobic interactions, and π-stacking), and ligand descriptors (physicochemical properties of compounds), this study attempts to learn which features, poses, and metrics are more useful for ranking. I believe this approach addresses an important direction for future research in virtual screening.

Critical weaknesses

This approach seems to be a form of feature selection and metric selection, and there are insufficient ablation studies or further analyses to demonstrate which types of information have higher reliability. Also, the study seems to address a relatively small search-space problem by integrating 13 existing structure-based features and heterogeneous metrics to identify which information is more useful for target-specific screening. To address this problem more concretely, it would be helpful to conduct broader experiments covering not only SBVS but also LBVS, 3D-based, and diffusion-based methods, and to demonstrate which metrics and feature information are actually important. The current experimental results seem insufficient to fully support these claims.

Also, as has been widely demonstrated in research on deep learning-based VS, various approaches have been proposed to enhance virtual screening performance through systematically constructed representations. As such, without presenting comparison results against SOTA models such as DrugCLIP[1] or PUResNET[2] etc.., it is difficult to ascertain whether the proposed method has achieved a significant improvement in ranking performance compared to these approaches, and if so, how. Including such comparisons would make the experimental results more convincing.

[References]

[1] Gao, Bowen, et al. "Drugclip: Contrastive protein-molecule representation learning for virtual screening." Advances in Neural Information Processing Systems 36 (2023): 44595-44614.

[2] Jeevan, Kandel, et al. "PUResNetV2. 0: a deep learning model leveraging sparse representation for improved ligand binding site prediction." Journal of Cheminformatics 16.1 (2024): 66.

Questions for authors
  • Is the manifest-level regime assignment defined by the authors, or are there previous studies that have divided datasets into these regimes in a similar way? How were the regime hints specified in the manifests determined?

  • If these concepts were introduced specifically for the authors' experimental setup, I think additional experiments under various settings would be necessary. Since the reported results are based on datasets constructed by the authors rather than established benchmark datasets, it is difficult to fully assess the reliability of the experimental results. Based on the dataset statistics, it is also unclear how challenging these tasks actually are(There are insufficient data points).

  • The number of data points is quite small. Have the authors considered data augmentation using AlphaFold-based simulations? There would likely be a gap between real experimental data and simulated information. How do the authors intend to address this gap?

  • For the sparse regime, I think comparisons with methods that perform well under low-label settings, such as SubDyve[1] or network propagation-based approaches, would be useful.

  • Recent studies have shown that subgraph information can help predict protein–ligand binding affinity, and that certain subgraphs can capture pharmacologically meaningful characteristics or contribute to docking affinity prediction. What advantages does the proposed method offer compared to these approaches?

  • SBVS can be computationally expensive, but one of its advantages is that it can provide accurate predictions. However, the number of compounds that need to be screened continues to increase. Therefore, I am also curious whether the proposed method offers any advantages in computational speed or memory efficiency compared to computationally efficient approaches such as DrugHash[2] or other models.

  1. It would be beneficial to conduct cold-split experiments on a large-scale set of protein–ligand pairs and analyze the results in comparison with a wider range of existing VS methods. Such experiments would help to demonstrate the generalizability and practical advantages of the proposed framework.

  2. There are quite a few model variants in this study, making it difficult to identify which components actually contribute to the performance improvements. Although the paper provides several ablation studies, it is difficult to regard these results as providing sufficient support for the authors’ main claims, particularly regarding the effectiveness of the hierarchical scorer and reliability calibration. It is unclear whether the reported improvements are directly related to these core components or mainly come from the additional policy priors, memory-based calibration, and other variants. A clearer analysis of how each component contributes to the final performance would help to support the paper’s key methodological claims.

[1] Yi, Jungseob, et al. "SubDyve: Subgraph-driven dynamic propagation for virtual screening enhancement controlling false positive." International Conference on Learning Representations. Vol. 2026. 2026.

[2] Han, Jin, Yun Hong, and Wu-Jun Li. "Drughash: Hashing based contrastive learning for virtual screening." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 39. No. 16. 2025.

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

PoT-Rank ensembles structural confidence, docking affinity, interface geometry, interaction fingerprints and ligand descriptors across multiple poses to rank ligands for a target protein. It uses separate dense and sparse ranking branches. It also applies reliability gating and curriculum-based reweighting, a cached training policy, memory-based score calibration. On the dense split containing 3 targets, 565 within-target pairs and the sparse split containing 79 targets, it surpasses classical scoring functions, deep affinity predictors, graph-based models, pose-aware rescorers, and structure-prediction-based affinity models, and matches or exceeds strong tabular baselines.

Critical strengths

Combining complementary structural, energetic, and interaction signals across shared poses is practically motivated and potentially useful for virtual screening.

Critical weaknesses
  1. The evaluation is limited. The test set contains only three dense targets and 79 sparse targets. It also lacks direct Boltz-2 benchmarking and end-to-end cost comparisons.
  2. The core model performs worse than XGBoost and CNN in Spearman correlation.
  3. The conclusion reports accuracy 0.6602 / 0.6897 for the dense and sparse Core models. But table1 reports 0.6460 / 0.5862. Also, the released policy cache in Supplementary Material is a placeholder and provides no sample-specific policy reweighting.
Questions for authors
  1. Could you provide comparison with Boltz2? Boltz2 emphasizes intra-assay relative-affinity supervision as well.
  2. Could the authors provide computational cost, including generating the input features, model training and inference?
  3. Could you provide performance on FEP+4, OpenFE and CASP16 datasets?
Reviewer_8HiaOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents PoT-Rank, a multi-conformation, multi-metric fusion framework for molecular affinity ranking within protein targets. Integrating structural prediction confidence, docking scores, interaction features and molecular descriptors, the method optimizes candidate molecule ranking via hierarchical weighting, reliability calibration and pairwise ranking. For different data scales, the authors design a Hierarchical Ranking Model for the dense regime and a Pairwise Ranking Model for the sparse regime. On the dense test set, PoT-Rank achieves a within-target pairwise accuracy of 0.6602, outperforming XGBoost’s 0.5770; it reaches 0.6897 on the sparse test set. The manuscript also includes comparisons across various scoring methods, ablation experiments and model stability analysis.

Critical strengths
  1. The paper addresses a practically meaningful problem in virtual screening: how to effectively integrate diverse binding poses and multiple scoring metrics to improve the ranking of candidate molecules within the same target. Using Pose-by-Metric Representation, Hierarchical Fusion and Reliability Calibration, the authors exploit complementary information including structural confidence, docking scores and molecular interactions. The overall design is reasonable and achieves promising results in the dense regime. Pairwise accuracy increases from XGBoost’s 0.5770 to 0.6602, and Spearman correlation rises from 0.2737 to 0.5343.

  2. The manuscript provides extensive experimental analyses. The authors compare traditional scoring functions, deep-learning affinity prediction models and learned baselines, and conduct comprehensive ablation studies, feature importance analysis, multi-random-seed experiments and data leakage checks. For instance, the dense model attains a pairwise accuracy of 0.6336 over five random seeds, still outperforming XGBoost at 0.5759. These experiments offer support for the effectiveness of the proposed method.

Critical weaknesses
  1. The experimental scale remains limited, and the generalization ability of the model is not fully demonstrated. Although notable performance gains are observed in the dense regime, its test set contains only 3 protein targets and 56 molecules. In the sparse regime with more test targets, PoT-Rank’s main metric is on par with XGBoost (both yield a pairwise accuracy of 0.6897). Moreover, the sparse setting contains roughly 1.66 candidate molecules per target on average, which deviates substantially from real-world virtual screening scenarios. Therefore, existing results cannot fully prove the stable advantages of this method across diverse targets and practical screening scales.

  2. The comparisons in Table 2 provide relatively limited evidence for validating the effectiveness of the proposed method. PoT-Rank essentially learns to fuse scoring signals from existing methods including Boltz-1, Smina and RTMScore using experimental affinity labels in the training set. Since Table 2 mainly compares this supervised trained fusion model against standalone scoring functions or affinity prediction models, superior ranking performance is not surprising. These results mostly demonstrate the value of supervised score fusion, yet they are insufficient to show the unique advantages of the proposed Hierarchical Fusion and Reliability Calibration over alternative fusion approaches.

Questions for authors
  1. In the dense regime, PoT-Rank-Core achieves a Spearman correlation of only 0.1728, while the full model reaches 0.5343. Interestingly, removing either Policy Prior or Memory Calibration alone leads to only minor performance degradation, yet removing both together drastically reduces the Spearman correlation. Could the authors further explain this phenomenon and discuss whether strong interactions exist between these two modules?

  2. Have the authors considered validating the dense model on more independent targets and adding learning-to-rank baselines with identical features and training objectives, to clearly demonstrate the independent contribution of the core fusion approach?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.