BindCLIP: A Unified Contrastive-Generative Representation Learning Framework for Virtual Screening

Anjie Qiao; Zhen Wang; Yaliang Li; Jiahua Rao; Yuedong Yang

cs.LGFeb 16, 2026

BindCLIP: A Unified Contrastive-Generative Representation Learning Framework for Virtual Screening

Authors: Anjie Qiao, Zhen Wang, Yaliang Li, Jiahua Rao, Yuedong Yang

Organizations: Sun Yat-sen University, Guangzhou, China · Alibaba Group, Hangzhou, China

Abstract

Virtual screening aims to efficiently identify active ligands from massive chemical libraries for a given target pocket. Recent CLIP-style models such as DrugCLIP enable scalable virtual screening by embedding pockets and ligands into a shared space. However, our analyses indicate that such representations can be insensitive to fine-grained binding interactions and may rely on shortcut correlations in training data, limiting their ability to rank ligands by true binding compatibility. To address these issues, we propose BindCLIP, a unified contrastive-generative representation learning framework for virtual screening. BindCLIP jointly trains pocket and ligand encoders using CLIP-style contrastive learning together with a pocket-conditioned diffusion objective for binding pose generation, so that pose-level supervision directly shapes the retrieval embedding space toward interaction-relevant features. To further mitigate shortcut reliance, we introduce hard-negative augmentation and a ligand-ligand anchoring regularizer that prevents representation collapse. Experiments on two public benchmarks demonstrate consistent improvements over strong baselines. BindCLIP achieves substantial gains on challenging out-of-distribution virtual screening and improves ligand-analogue ranking on the FEP+ benchmark. Together, these results indicate that integrating generative, pose-level supervision with contrastive learning yields more interaction-aware embeddings and improves generalization in realistic screening settings, bringing virtual screening closer to real-world applicability.

Figures & tables

Panel of Figure 1 — Figure 1. Illustrative probe cases from the test set.

Figure 2. Overview of the BindCLIP framework. (a) Pocket-conditioned ligand binding pose generation objective; (b) Contrastive learning objective; (c) Hard-negative augmentation workflow.

Figure 3. Visualization of pocket, ligand, and hard-negative embeddings with/without the anchoring regularizer (t-SNE).

	AUROC	BEDROC	$\text{EF}^{0.5\%}$	$\text{EF}^{1\%}$	$\text{EF}^{5\%}$
Surflex	51.47	-	-	2.50	-
Glide-SP	53.15	4.00	3.17	3.41	2.01
Planet	57.31	-	4.64	3.87	2.43
Gnina	60.93	5.40	-	4.63	-
DeepDTA	56.27	2.53	-	1.47	-
BigBind	60.80	-	-	3.82	-

Table 1. Results on LIT-PCBA. Best results are shown in bold.

	AUROC	BEDROC	$\text{EF}^{0.5\%}$	$\text{EF}^{1\%}$	$\text{EF}^{5\%}$
Vina	71.60	-	9.13	7.32	4.44
Glide-SP	76.70	40.70	19.39	16.18	7.23
RFscore	65.21	12.41	4.90	4.52	2.98
Pafnucy	63.11	16.50	4.24	3.86	3.76
OnionNet	59.71	8.62	2.84	2.84	2.20
Planet	71.60	-	10.23	8.83	5.40

Table 2. Results on DUD-E. Best results are shown in bold.

Figure 4. Out-of-distribution evaluation on the MF-PCBA subset, excluding assays whose proteins have more than 30% sequence identity with any protein in the training set.

Figure 5. Evaluation on the 4-target FEP subset (CDK2, TYK2, JNK1, and P38) for ligand activity ranking within a chemical series.

Ablation Setting			AUROC	BEDROC	$\text{EF}^{0.5\%}$	$\text{EF}^{1\%}$	$\text{EF}^{5\%}$
Diffusion	Rand. Neg.	Hard Neg.	AUROC	BEDROC	$\text{EF}^{0.5\%}$	$\text{EF}^{1\%}$	$\text{EF}^{5\%}$
✗	✗	✗	55.45 $\pm$ 1.31	6.41 $\pm$ 0.87	8.24 $\pm$ 1.52	5.21 $\pm$ 1.11	2.14 $\pm$ 0.35
✓	✗	✗	56.71 $\pm$ 1.10	6.85 $\pm$ 0.41	9.11 $\pm$ 0.67	5.72 $\pm$ 0.32	2.27 $\pm$ 0.22
✓	✓	✗	58.91 $\pm$ 0.89	7.21 $\pm$ 0.53	9.17 $\pm$ 0.75	5.89 $\pm$ 0.60	2.66 $\pm$ 0.24
✓	✗	✓	59.15 $\pm$ 1.26	7.88 $\pm$ 0.25	9.84 $\pm$ 0.65	6.26 $\pm$ 0.35	2.90 $\pm$ 0.18

Table 3. Ablation studies on LIT-PCBA.

Task	Model	RMSE	MAE
Pocket–Ligand Affinity Prediction	DrugCLIP	1.63	1.30
Pocket–Ligand Affinity Prediction	BindCLIP	1.35	1.06
Atom–Target Distance Prediction	DrugCLIP	1.05	0.79
Atom–Target Distance Prediction	BindCLIP	0.87	0.61

Table 4. The performance of two probing experiments using linear probing on global (pocket–ligand affinity) and atom-level (atom–target distance) embeddings, averaged over five random seeds.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6. Atom-level embedding visualization (t-SNE). Ligand atom embeddings from DrugCLIP (left) and UniMol (right), colored by atom type (C/N/O/S/Halogens, where halogens include F, Cl, Br, and I).

Figure 7. DrugCLIP ranking vs. Vina re-ranking. For pocket 7KSI, we take the top-100 candidates retrieved by DrugCLIP and re-rank them with AutoDock Vina (exhaustiveness=16).

Figure 8. A case of an activity cliff in JNK1. Two structurally similar ligands (18634-1 and 18636-1) exhibit a large difference in binding free energy ( $\Delta\Delta G$ =2.48 kcal/mol). BindCLIP better captures the affinity ordering across this activity cliff.

Virtual Screening

Virtual screening uses computational methods to prioritize molecules for testing against biological targets or desired properties.

See all 20 papers

PETA:Parameter-Efficient Test-Time Adaptation for Virtual Screening
Aug 20, 2026Jia-Qi Lin, Yinghua Yao, Chang-Dong Wang +2Virtual Screening Test-Time Adaptation
Centre for Frontier AI Research (CFAR), Agency for Science, Technology and Research (A*STAR), Singapore 138632 · School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China
Structure-guided molecular design with contrastive 3D protein-ligand learning
Apr 21, 2026Carles Navarro, Philipp Tholke, Gianni de FabritiisContrastive Learning Virtual Screening
Acellera Labs, C/ Doctor Trueta 183, 08005 Barcelona, Spain · Computational Science Laboratory, Universitat Pompeu Fabra, PRBB, C/ Doctor Aiguader 88, 08003 Barcelona, Spain · Instituci´o Catalana de Recerca i Estudis Avan¸cats (ICREA), Passeig Llu´ıs Companys 23, 08010 Barcelona, Spain
CausalBind: Causal Modeling and Learning for Protein-Molecule Virtual Screening
Oct 5, 2026Loka Li, Jin Tian, Kun ZhangCausal Representation Learning Structured Sparsity
Mohamed bin Zayed University of Artificial Intelligence · Carnegie Mellon University

Aug 20, 2026J/K move · Enter open · S save

PETA:Parameter-Efficient Test-Time Adaptation for Virtual Screening

Jia-Qi Lin, Yinghua Yao, Chang-Dong Wang, Yew-Soon Ong+1

Centre for Frontier AI Research (CFAR), Agency for Science, Technology and Research (A*STAR), Singapore 138632 · School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China

Virtual Screening Test-Time Adaptation Drug Discovery Parameter-Efficient Fine-Tuning

Open paper AskPDF arXiv

Accurately ranking active ligands for a target protein pocket from massive chemical libraries remains a central challenge in virtual screening. DrugCLIP and its recent extensions substantially accelerate this process by encoding protein pockets and molecules into a shared embedding space. Despite this progress, further performance improvements typically require retraining the entire model, incurring substantial computational overhead and making target-specific customization inefficient. In this work, we formulate the specialization of pretrained virtual screening models to individual pockets as a test-time adaptation problem and propose PETA, a parameter-efficient framework that directly adapts pretrained model at test time. Given a target pocket, PETA constructs pocket-specific negatives through molecular diffusion and chemical validity filtering, and further moves them toward the reference ligand retrieved from structural databases via embedding-space mixup to create more challenging ranking tasks. A ranking objective then places greater emphasis on suppressing high-scoring invalid candidates that could contaminate the top-ranked screening results, providing structured supervision for lightweight adaptation. Experiments across diverse benchmarks demonstrate that this lightweight, pocket-specific adaptation outperforms both pretrained and fully retrained baselines while updating only the LayerNorm parameters, which account for approximately

0.03\%

of the full model.

Abstract

Virtual Screening

Explore similar work

PETA:Parameter-Efficient Test-Time Adaptation for Virtual Screening

Structure-guided molecular design with contrastive 3D protein-ligand learning

CausalBind: Causal Modeling and Learning for Protein-Molecule Virtual Screening