pith. sign in

arxiv: 2604.23115 · v1 · submitted 2026-04-25 · 💻 cs.LG

HBGSA: Hydrogen Bond Graph with Self-Attention for Drug-Target Binding Affinity Prediction

Junxiao Kong , Chupei Tang , Di Wang , Jixiu Zhai , Yi He , Moyu Tang , Tianchi Lu This is my paper

Pith reviewed 2026-05-08 08:16 UTC · model grok-4.3

classification 💻 cs.LG
keywords drug-target binding affinityhydrogen bond graphgraph neural networkself-attentionPearson correlation lossvirtual screeningPDBbindCSAR-HiQ
0
0 comments X

The pith

HBGSA improves drug-target binding affinity prediction by modeling hydrogen bond spatial features with graph neural networks, self-attention, and Pearson correlation loss.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents HBGSA as a method to predict how tightly a drug molecule binds its protein target. Existing approaches either lose three-dimensional spatial details by working only with sequences or overlook hydrogen bond patterns even when using structures, while typical training objectives do not emphasize how well predictions match the true affinity values. HBGSA builds a graph representation focused on hydrogen bonds, processes it with neural networks plus attention to capture spatial relationships, and adds a Pearson correlation term to the loss function. If the approach holds, virtual screening can rank candidate compounds more reliably, reducing the number that must be tested in the lab. The model remains compact at 3.06 million parameters and reports stronger results than prior methods on the PDBbind Core Set and CSAR-HiQ data.

Core claim

HBGSA encodes hydrogen bond spatial features by applying graph neural networks to model the spatial topology of hydrogen bonds, with self-attention enhancement, and trains using Pearson correlation loss together with conventional objectives. This design directly targets three limitations: loss of geometric constraints in sequence models, underuse of hydrogen bond information in structure models, and neglect of prediction-target correlation in standard losses. On the PDBbind Core Set and CSAR-HiQ dataset the model outperforms baselines and exhibits strong generalization, with ablation experiments isolating the contributions of the hydrogen bond graph and the correlation loss.

What carries the argument

The Hydrogen Bond Graph with Self-Attention mechanism, which represents hydrogen bonds as a graph whose spatial topology is processed by graph neural networks augmented with self-attention layers, combined with Pearson correlation loss to align predicted and measured affinities.

If this is right

  • More accurate ranking of high-affinity compounds during virtual screening, reducing experimental workload.
  • Better exploitation of three-dimensional hydrogen bond geometry that sequence-based methods discard.
  • Training objectives that explicitly reward correlation between predictions and targets improve identification of strong binders.
  • Ablation results confirm that both the hydrogen bond graph and Pearson loss contribute measurably to the reported gains.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same graph-construction strategy focused on specific interaction types could be reused to model other non-covalent contacts such as pi-stacking or salt bridges.
  • Because the model contains only 3.06 million parameters it may remain practical for screening compound libraries that contain millions of molecules on ordinary hardware.
  • If hydrogen bond topology proves broadly informative, similar lightweight graph layers could be added to existing structure-based predictors without requiring full retraining.

Load-bearing premise

Hydrogen bond spatial topology modeled as a graph and processed by neural networks with attention, together with a Pearson correlation term in the loss, captures the dominant factors that set binding affinity and generalizes to new drug-target pairs.

What would settle it

An independent test set of drug-target complexes, drawn from a source outside PDBbind and CSAR-HiQ, on which HBGSA shows no improvement in accuracy or correlation over standard baselines, or on which removing the hydrogen-bond graph component leaves performance unchanged.

Figures

Figures reproduced from arXiv: 2604.23115 by Chupei Tang, Di Wang, Jixiu Zhai, Junxiao Kong, Moyu Tang, Tianchi Lu, Yi He.

Figure 1
Figure 1. Figure 1: Overall architecture of the HBGSA model. fined Set; (2) remove 82 validation and 5 train￾ing complexes to maintain consistency with Deep￾DTAF [2] and Pafnucy [12]. After cleaning, General Set contains 9,221 complexes and Refined Set con￾tains 3,685 complexes. Data partition: Test Set uses Core Set 2016 (290 complexes); Validation Set randomly samples 1,000 from cleaned Refined Set; Training Set com￾bines G… view at source ↗
Figure 3
Figure 3. Figure 3: Hydrogen bond graph neural network encoder. second layer adds residual connections: H(2) = GELU(H(1) +LayerNorm(Linear(A·H(1)))) (6) Finally, global max pooling yields shb = max(H(2) , dim = 0) ∈ R 128 . 3.6 Prediction and Optimization We concatenate features from all branches into a unified representation: scat = [sseq; spkt; ssmi; shb] ∈ R 512 (7) A three-layer fully connected network progressively downs… view at source ↗
Figure 4
Figure 4. Figure 4: visualizes these trade-offs using a radar chart, clearly showing that λ = 50 achieves the most balanced performance across all metrics view at source ↗
Figure 5
Figure 5. Figure 5: Prediction performance on 4 sets. 8 view at source ↗
Figure 6
Figure 6. Figure 6: Sorted bar chart analysis for 4 sets view at source ↗
Figure 8
Figure 8. Figure 8: Relationship between hydrogen bond density and binding affinity. Yellow dashed lines indicate hydrogen bonds. higher affinity due to its superior hydrogen bond density (26.3% vs. 15.8%) view at source ↗
read the original abstract

Accurate prediction of drug-target binding affinity accelerates drug discovery by prioritizing compounds for experimental validation. Current methods face three limitations: sequence-based approaches discard spatial geometric constraints, structure-based methods fail to exploit hydrogen bond features, and conventional loss functions neglect prediction-target correlation, a key factor for identifying high-affinity compounds in virtual screening. We developed HBGSA (Hydrogen Bond Graph with Self-Attention), a 3.06M-parameter model that encodes hydrogen bond spatial features. HBGSA uses graph neural networks to model hydrogen bond spatial topology with self-attention enhancement and Pearson correlation loss. Experimental results on PDBbind Core Set and CSAR-HiQ dataset demonstrate that HBGSA outperforms baseline methods with strong generalization capability. Ablation studies confirm the effectiveness of hydrogen bond modeling and Pearson correlation loss.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces HBGSA, a 3.06M-parameter model that encodes hydrogen bond spatial topology using graph neural networks augmented with self-attention and trained under a Pearson correlation loss. It claims superior performance over baseline methods on the PDBbind Core Set and CSAR-HiQ datasets, asserts strong generalization capability, and presents ablation studies supporting the contributions of hydrogen-bond modeling and the Pearson loss.

Significance. If the performance claims are substantiated with complete quantitative results and appropriate validation, the emphasis on explicit hydrogen-bond graph features plus correlation-aware training could offer a practical advance for structure-based affinity prediction in virtual screening, particularly given the modest parameter count.

major comments (3)
  1. [Abstract] Abstract: the central claim that HBGSA 'outperforms baseline methods with strong generalization capability' is stated without any numerical results (e.g., RMSE, Pearson r, or baseline values), error bars, or statistical tests, preventing verification of the asserted improvement.
  2. [Experimental results] Experimental results section: both PDBbind Core Set and CSAR-HiQ are drawn from the same overall PDBbind collection; no sequence-identity filtering, temporal split, or external OOD benchmark (e.g., BindingDB or kinase-specific sets) is described, so the 'strong generalization' assertion rests on in-distribution performance only.
  3. [Methods] Methods: the construction of the hydrogen-bond graph, the precise integration of self-attention with the GNN layers, and the exact form of the Pearson loss are not supplied with equations or pseudocode, which are load-bearing for reproducing or assessing the claimed gains.
minor comments (2)
  1. [Abstract] The 3.06 M parameter count is given but without an architecture table or comparison to the baselines' sizes.
  2. Baseline methods should be explicitly named with citations and implementation details (e.g., whether re-implemented or taken from original papers).

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful and constructive comments, which have helped us improve the clarity and rigor of the manuscript. We address each major comment point by point below and have made revisions where appropriate to strengthen the presentation.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that HBGSA 'outperforms baseline methods with strong generalization capability' is stated without any numerical results (e.g., RMSE, Pearson r, or baseline values), error bars, or statistical tests, preventing verification of the asserted improvement.

    Authors: We agree that the abstract should include quantitative support for the performance claims. In the revised version, we have updated the abstract to report key metrics including RMSE and Pearson r values for HBGSA alongside the main baselines, with reference to error bars obtained from repeated runs. revision: yes

  2. Referee: [Experimental results] Experimental results section: both PDBbind Core Set and CSAR-HiQ are drawn from the same overall PDBbind collection; no sequence-identity filtering, temporal split, or external OOD benchmark (e.g., BindingDB or kinase-specific sets) is described, so the 'strong generalization' assertion rests on in-distribution performance only.

    Authors: The referee is correct that both sets are subsets of PDBbind and that no explicit sequence-identity filtering or external OOD benchmark was performed. CSAR-HiQ is a distinct and commonly used held-out collection with different characteristics, but this does not fully address out-of-distribution concerns. We have revised the text to moderate the generalization claim, clarified the dataset relationship, and added a limitations paragraph discussing this point with plans for future external validation. revision: partial

  3. Referee: [Methods] Methods: the construction of the hydrogen-bond graph, the precise integration of self-attention with the GNN layers, and the exact form of the Pearson loss are not supplied with equations or pseudocode, which are load-bearing for reproducing or assessing the claimed gains.

    Authors: We appreciate this observation and apologize for the insufficient detail in the original submission. We have substantially expanded the Methods section to provide the explicit equations for hydrogen-bond graph construction, the self-attention integration within the GNN layers, and the precise Pearson correlation loss formulation, together with pseudocode to ensure reproducibility. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical model with independent experimental validation

full rationale

The paper describes a standard GNN + self-attention architecture for hydrogen-bond graphs, trained end-to-end with Pearson correlation loss on public PDBbind Core and CSAR-HiQ sets. No equations, uniqueness theorems, or ansatzes are presented that reduce the reported performance or generalization claim to a fitted parameter or self-citation by construction. Ablation studies and baseline comparisons constitute independent empirical content. The derivation chain consists of conventional architectural choices whose outputs are not definitionally equivalent to the inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The abstract supplies no equations, hyperparameters, or new postulated entities, so the ledger is empty; full text would be required to audit free parameters or domain assumptions.

pith-pipeline@v0.9.0 · 5451 in / 1198 out tokens · 44194 ms · 2026-05-08T08:16:57.088497+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 3 canonical work pages · 2 internal anchors

  1. [1]

    Öztürk, H., Özgür, A., & Ozkirimli, E. (2018). DeepDTA: deep drug–target binding affinity prediction. Bioinformatics, 34(17), i821-i829

    2018

  2. [2]

    Jiang, M., Li, Z., Zhang, S., Wang, S., Wang, X., Yuan, Q., & Wei, Z. (2021). Drug–target affinity prediction using graph neural network and contact maps. RSC Advances , 10(35), 20701-20712

    2021

  3. [3]

    P., Nguyen, T., Le, T

    Nguyen, T., Le, H., Quinn, T. P., Nguyen, T., Le, T. D., & Venkatesh, S. (2021). GraphDTA: predicting drug–target binding affinity with graph neural networks. Bioinformatics, 37(8), 1140-1147

    2021

  4. [4]

    Wang, K., Zhou, R., Tang, J., & Li, M. (2022). InteractionGraphNet: A Novel and Efficient Deep Graph Representation Learning Frame- work for Accurate Protein–Ligand Interaction Predictions. Journal of Medicinal Chemistry , 65(10), 7155-7171

    2022

  5. [5]

    Lu, W., Wu, Q., Zhang, J., Rao, J., Li, C., & Zheng, S. (2022). TANKBind: Trigonometry- Aware Neural NetworKs for Drug-Protein Binding Structure Prediction. Advances in Neural Information Processing Systems , 35, 7236-7249

    2022

  6. [6]

    Zhao, Q., Duan, G., Yang, M., Cheng, Z., Li, Y., & Wang, J. (2023). MMPD-DTA: Integrat- ing Multi-Modal Deep Learning with Pocket- Drug Graphs for Drug-Target Binding Affinity Prediction. Bioinformatics, 39(5), btad234

    2023

  7. [7]

    Yang, Z., Zhong, W., Lv, Q., Dong, T., & Chen, C. Y. C. (2023). ML-PLA: Enhanc- ing Protein-Ligand Binding Affinity Prediction with Microenvironment and Long-Range In- teraction Aware. Briefings in Bioinformatics , 24(4), bbad451

    2023

  8. [8]

    Li, S., Wan, F., Shu, H., Jiang, T., Zhao, D., & Zeng, J. (2021). GIGN: Learning Geometry- Aware Interaction Graph Neural Network for Protein-Ligand Binding Affinity Prediction. Bioinformatics, 37(18), 2988-2995

    2021

  9. [9]

    & Dou, D

    Li, S., Zhou, J., Xu, T., Huang, L., Wang, F., Xiong, H., ... & Dou, D. (2021). Structure- Aware Interactive Graph Neural Networks for the Prediction of Protein-Ligand Binding Affin- ity. KDD, 975-985

    2021

  10. [10]

    Li, M., Cao, Y., Liu, X., & Ji, H. (2024). Structure-Aware Graph Attention Diffusion Network for Protein–Ligand Binding Affinity Prediction. IEEE Transactions on Neural Net- works and Learning Systems , 35(12), 18370- 18380

    2024

  11. [11]

    Li, M., Zhang, Y., Li, Y., & Wang, J. (2025). Knowledge-enhanced and structure- enhanced representation learning for protein– ligand binding affinity prediction. Pattern Recognition, 166, 111701

    2025

  12. [12]

    M., Zielenkiewicz, P., & Siedlecki, P

    Stepniewska-Dziubinska, M. M., Zielenkiewicz, P., & Siedlecki, P. (2018). Development and evaluation of a deep learning model for protein– ligand binding affinity prediction. Bioinformat- ics, 34(21), 3666-3674

    2018

  13. [13]

    H., Ko, J., & Lee, J

    Kwon, Y., Shin, W. H., Ko, J., & Lee, J. (2020). AK-Score: Accurate Protein-Ligand Binding Affinity Prediction Using an Ensem- ble of Convolutional Neural Networks. Inter- national Journal of Molecular Sciences , 21(22), 8424

    2020

  14. [14]

    Zheng, L., Fan, J., & Mu, Y. (2019). OnionNet: A Multiple-Layer Intermolecular- Contact-Based Convolutional Neural Network for Protein–Ligand Binding Affinity Predic- tion. ACS Omega, 4(14), 15956-15965

    2019

  15. [15]

    Öztürk, H., Ozkirimli, E., & Özgür, A. (2019). WideDTA: Prediction of drug-target binding affinity. arXiv preprint arXiv:1902.04166

    work page arXiv 2019

  16. [16]

    Xu, W., Wang, X., Luo, H., Shan, W., Liu, B., & Huang, X. (2025). UAMRL: multi- granularity uncertainty-aware multimodal rep- resentation learning for drug-target affinity prediction. Bioinformatics, 41(10), btaf512

    2025

  17. [17]

    Lai, H., Gao, Y., Tan, C., Huang, P., Ron- grong, J., & Cheng, J. (2024). Interformer: an interaction-aware model for protein-ligand docking and affinity prediction. Nature Com- munications, 15(1), 10223

    2024

  18. [18]

    V., Kc, D

    Samudrala, M. V., Kc, D. B., & Bhattacharya, D. (2025). PLAIG: Protein–Ligand Binding Affinity Prediction Using a Novel Interaction- Based Graph Neural Network Framework. ACS Bio & Med Chem Au , 5, 447-463. 12 REFERENCES REFERENCES

    2025

  19. [19]

    A., Hoffmann, M., Steinmann, C., & Hensen, U

    Moesser, M. A., Hoffmann, M., Steinmann, C., & Hensen, U. (2022). PLIG: A structure- informed approach for protein-ligand interac- tion prediction. Journal of Chemical Informa- tion and Modeling , 62(13), 3170-3183

    2022

  20. [20]

    Yi, Y., Zhao, Z., Sun, J., & Huang, B. (2024). Equivariant Line Graph Neural Net- work for Protein-Ligand Binding Affinity Pre- diction. IEEE Journal of Biomedical and Health Informatics, 28(7), 4336-4347

    2024

  21. [21]

    Truong Jr, T. F. (2020). Interpretable Deep Learning Framework for Binding Affinity Pre- diction. Master’s thesis, Massachusetts Insti- tute of Technology

    2020

  22. [22]

    W., Brent, R

    Kyro, G. W., Brent, R. I., & Batista, V. S. (2023). HAC-Net: A Hybrid Attention-Based Convolutional Neural Network for Highly Ac- curate Protein–Ligand Binding Affinity Pre- diction. Journal of Chemical Information and Modeling, 63(6), 1947-1960

    2023

  23. [23]

    & Wang, R

    Liu, Z., Li, Y., Han, L., Li, J., Liu, J., Zhao, Z., ... & Wang, R. (2015). PDB-wide collection of binding data: current status of the PDBbind database. Bioinformatics, 31(3), 405-412

    2015

  24. [24]

    B., Smith, R

    Dunbar Jr, J. B., Smith, R. D., Yang, C. Y., Ung, P. M. U., Lexa, K. W., Khazanov, N. A., ... & Carlson, H. A. (2011). CSAR bench- mark exercise of 2010: selection of the protein– ligand complexes. Journal of Chemical Infor- mation and Modeling , 51(9), 2036-2046

    2011

  25. [25]

    A., Grabowski, H

    DiMasi, J. A., Grabowski, H. G., & Hansen, R. W. (2016). Innovation in the pharmaceutical industry: New estimates of R&D costs. Jour- nal of Health Economics , 47, 20-33

    2016

  26. [26]

    & Zheng, S

    Lu, W., Zhang, J., Huang, W., Zhang, Z., Jia, X., Wang, Z., ... & Zheng, S. (2024). Dynamic- Bind: predicting ligand-specific protein-ligand complex structure with a deep equivariant gen- erative model. Nature Communications, 15(1), 1071

    2024

  27. [27]

    & Dou, D

    Li, S., Zhou, J., Xu, T., Huang, L., Wang, F., Xiong, H., ... & Dou, D. (2021). Structure- aware interactive graph neural networks for the prediction of protein-ligand binding affin- ity. Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, 975-985

    2021

  28. [28]

    F., Maziarz, K., Misztela, H., Lanini, J., Segler, M.,

    Stanley, M., Bronskill, J. F., Maziarz, K., Misztela, H., Lanini, J., Segler, M., ... & Brockschmidt, M. (2021). FS-Mol: A few- shot learning dataset of molecules. Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks , 1

    2021

  29. [29]

    & Rives, A

    Lin, Z., Akin, H., Rao, R., Hie, B., Zhu, Z., Lu, W., ... & Rives, A. (2023). Evolutionary-scale prediction of atomic-level protein structure with a language model. Science, 379(6637), 1123-1130

    2023

  30. [30]

    Bissantz, C., Kuhn, B., & Stahl, M. (2010). A medicinal chemist’s guide to molecular interac- tions. Journal of Medicinal Chemistry , 53(14), 5061-5084

    2010

  31. [31]

    R., Klein, R

    Arunan, E., Desiraju, G. R., Klein, R. A., Sadlej, J., Scheiner, S., Alkorta, I., ... & Nes- bitt, D. J. (2011). Definition of the hydrogen bond (IUPAC Recommendations 2011). Pure and Applied Chemistry , 83(8), 1637-1641

    2011

  32. [32]

    Vaswani, A., Shazeer, N., Parmar, N., Uszko- reit, J., Jones, L., Gomez, A. N., ... & Polo- sukhin, I. (2017). Attention is all you need. Ad- vances in Neural Information Processing Sys- tems, 30, 5998-6008

    2017

  33. [33]

    Loshchilov, I., & Hutter, F. (2017). Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101

    work page internal anchor Pith review arXiv 2017

  34. [34]

    Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., & Salakhutdinov, R. (2014). Dropout: a simple way to prevent neural net- works from overfitting. The Journal of Machine Learning Research, 15(1), 1929-1958

    2014

  35. [35]

    He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition , 770-778

    2016

  36. [36]

    Semi-Supervised Classification with Graph Convolutional Networks

    Kipf, T. N., & Welling, M. (2016). Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907

    work page internal anchor Pith review arXiv 2016

  37. [37]

    Benesty, J., Chen, J., Huang, Y., & Cohen, I. (2009). Pearson correlation coefficient. In Noise reduction in speech processing (pp. 1-4). Springer, Berlin, Heidelberg. 13

    2009