pith. sign in

arxiv: 2604.23134 · v1 · submitted 2026-04-25 · 💻 cs.LG

h-MINT: Modeling Pocket-Ligand Binding with Hierarchical Molecular Interaction Network

Yanru Qu , Yijie Zhang , Wenjuan Tan , Xiangzhe Kong , Xiangxin Zhou , Chaoran Cheng , Mathieu Blanchette , Jiaxuan You
show 1 more author
Ge Liu
This is my paper

Pith reviewed 2026-05-08 08:34 UTC · model grok-4.3

classification 💻 cs.LG
keywords molecular tokenizationbinding affinity predictionvirtual screeninghierarchical neural networkprotein-ligand interactiondrug discoverymachine learning
0
0 comments X

The pith

Overlapping molecular fragments and a hierarchical network improve ligand-protein binding predictions.

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

The paper addresses the difficulty of representing molecules so that local chemical environments supporting interactions such as hydrogen bonds and pi-stacking are captured for pocket-ligand binding tasks. Atom-level graphs lose higher-order details like stereochemistry and conjugation, while standard fragment methods use rigid non-overlapping pieces that discard essential context. OverlapBPE creates data-driven tokens that can overlap to reflect fuzzy substructure boundaries and carries richer chemical information at each token. The h-MINT architecture then jointly processes atom and fragment levels to handle the resulting many-to-many mappings. If the gains hold, drug discovery pipelines could identify better binders with modestly higher accuracy on affinity and screening benchmarks.

Core claim

The central claim is that OverlapBPE tokenization, which permits overlapping fragments to preserve fuzzy boundaries and enriched chemical context including chirality and ionic states, combined with the h-MINT hierarchical molecular interaction network that models interactions at both atom and fragment levels, produces measurable improvements: 2-4% higher Pearson and Spearman correlations for binding affinity on PDBBind and LBA, 1-3% gains in key virtual screening metrics on DUD-E and LIT-PCBA, and the best overall high-throughput screening results on PubChem assays.

What carries the argument

OverlapBPE, a data-driven tokenization scheme that generates overlapping molecular fragments, together with the h-MINT hierarchical network that jointly models atom-level and fragment-level interactions to accommodate the induced many-to-many mappings.

If this is right

  • Better retention of stereochemistry, aromaticity, and ionic state information in representations used for binding tasks.
  • Higher Pearson and Spearman correlations for binding affinity prediction on PDBBind and LBA benchmarks.
  • Improved enrichment and other screening metrics on DUD-E and LIT-PCBA datasets.
  • Stronger overall results on high-throughput screening assays from PubChem.

Where Pith is reading between the lines

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

  • The overlapping-fragment approach could be tested on other molecular property tasks such as predicting solubility or metabolic stability.
  • Inspecting which overlapping tokens receive high attention in the hierarchical layers might highlight recurring interaction patterns across different protein families.
  • Pairing the token-level hierarchy with explicit 3D pocket geometry encoders offers a direct route to models that reason about both chemical context and spatial fit.

Load-bearing premise

The observed performance gains arise from the OverlapBPE overlapping tokenization and hierarchical architecture rather than from differences in training data splits, hyperparameter tuning, or preprocessing choices.

What would settle it

Retraining the strongest baseline models on the exact same data splits and with matching hyperparameters and preprocessing but replacing OverlapBPE with a standard non-overlapping tokenizer and removing the hierarchical component, then checking whether the 2-4% and 1-3% advantages disappear.

Figures

Figures reproduced from arXiv: 2604.23134 by Chaoran Cheng, Ge Liu, Jiaxuan You, Mathieu Blanchette, Wenjuan Tan, Xiangxin Zhou, Xiangzhe Kong, Yanru Qu, Yijie Zhang.

Figure 1
Figure 1. Figure 1: Illustration of the OverlapBPE tokenization process. (i) Starting from the molecule in A, we first extract all basic tokens from the atom graph. (ii) After identifying all basic tokens, we obtain the initial fragments (left) and token graph (right), as shown in B, which contains 4 tokens in 3 types: c1ccccc1 (freq=3778), Cc (freq=3496), and Sc (freq=637). (iii) We then enumerate all adjacent token pairs an… view at source ↗
Figure 2
Figure 2. Figure 2: Overall model architecture. (A) Global node, fragments, and atoms in the ligand molecule of an input pair. (B) The aggregation of fragment and global embeddings. (C) An encoder layer of h-MINT. Note: Solid lines indicate connection within the same level. Dashed lines indicate connections across different levels. charged chlorine atom, while [n+] indicates a positively charged aromatic nitrogen. In contrast… view at source ↗
Figure 3
Figure 3. Figure 3: OverlapBPE (ours) better preserves aromatic bond integrity, and ionic states. (A) An interaction formed between the ligand and the protein pocket. The ligand contains a positively charged [N+], which forms two π-cation interactions with two aromatic rings in the protein pocket. (B) Representation using our tokenization method. Green colors indicate fragments without charge. Red colors indicate charged frag… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of overlap (top) and non-overlap tokenization (bottom). Overlap vs. Non-overlap Tokenization In view at source ↗
Figure 5
Figure 5. Figure 5: Noise robustness comparison on LBA. We report results from 3 runs. In this section, we analyze the robustness and generalization of GET, GET-PS, and our model under different noise scales on LBA. We consider two noise settings: adding noise only to the training set for simulating scenarios when training data is of low quality or is predicted, and adding noise to both training and test sets for simulating s… view at source ↗
Figure 6
Figure 6. Figure 6: t-SNE Visualization of Fragment Embeddings view at source ↗
Figure 7
Figure 7. Figure 7: H-bond acceptors distribution across clusters. F.5 ADDITIONAL EXPERIMENTS ON MOLECULAR PROPERTY PREDICTION We include 3 property prediction tasks from MoleculeNet. The baseline follows MoleculeNet directly, which extracts ECFP features and trains XGBoost with grid search. The ECFP features are chosen from 128-bit, 512-bit, 1024-bit and 2048-bit according to dataset. We augment the ECFP features with bag-of… view at source ↗
Figure 8
Figure 8. Figure 8: Top-100 composite tokens from LBA vocabulary. 26 view at source ↗
read the original abstract

Accurate molecular representations are critical for drug discovery, and a central challenge lies in capturing the chemical environment of molecular fragments, as key interactions, such as H-bond and {\pi} stacking, occur only under specific local conditions. Most existing approaches represent molecules as atom-level graphs; however, atom-level representations can hardly express higher-order chemical context (e.g., stereochemistry, lone pairs, conjugation). Fragment-based methods (e.g., principal subgraph, predefined functional groups) fail to preserve essential information such as chirality, aromaticity, and ionic states. This work addresses these limitations from two aspects. (i) OverlapBPE tokenization. We propose a novel data-driven molecule tokenization method. Unlike existing approaches, our method allows overlapping fragments, reflecting the inherently fuzzy boundaries of small-molecule substructures and, together with enriched chemical information at the token level, thereby preserving a more complete chemical context. (ii) h-MINT model. OverlapBPE induces many-to-many atom-fragment mappings, which necessitate a new hierarchical architecture. We therefore develop a hierarchical molecular interaction network capable of jointly modeling interactions at both atom and fragment levels. By supporting fragment overlaps, the model naturally accommodates the many-to-many atom-fragment mappings introduced by the OverlapBPE scheme. Extensive evaluation against state-of-the-art methods shows our method improves binding affinity prediction by 2-4% Pearson/Spearman correlation on PDBBind and LBA, enhances virtual screening by 1-3% in key metrics on DUD-E and LIT-PCBA, and achieves the best overall HTS performance on PubChem assays. Further analysis demonstrates that our method effectively captures interactive information while maintaining good generalization.

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

2 major / 1 minor

Summary. The paper proposes OverlapBPE, a data-driven tokenization allowing overlapping molecular fragments to capture richer chemical context (including chirality, aromaticity, and ionic states), and the h-MINT hierarchical architecture that jointly models atom- and fragment-level interactions via many-to-many mappings. It claims 2-4% gains in Pearson/Spearman correlation for binding affinity prediction on PDBBind and LBA, 1-3% improvements in virtual screening metrics on DUD-E and LIT-PCBA, and best overall HTS performance on PubChem assays, attributing these to better preservation of local interaction contexts.

Significance. If the performance lifts are causally attributable to OverlapBPE and the hierarchical layers rather than uncontrolled experimental factors, the work would advance fragment-aware molecular representations beyond standard atom graphs or fixed functional-group approaches. The emphasis on overlapping substructures addresses a genuine limitation in current methods for modeling context-dependent interactions such as H-bonds and pi-stacking.

major comments (2)
  1. [Abstract and §4] Abstract and §4 (Experiments): The reported 2-4% correlation improvements and 1-3% virtual-screening gains are presented without ablation studies that isolate OverlapBPE (with its overlapping fragments and enriched token features) and the h-MINT many-to-many interaction layers while holding data splits, random seeds, optimizer schedules, and preprocessing pipelines fixed against the cited baselines. This leaves the central attribution claim vulnerable to alternative explanations such as more favorable partitioning or hyperparameter differences.
  2. [Abstract] Abstract: No information is supplied on baseline re-implementations, error bars, statistical significance tests (e.g., paired t-tests or bootstrap confidence intervals), or variance across multiple runs, making it impossible to judge whether the stated improvements exceed experimental noise.
minor comments (1)
  1. [Abstract] The abstract's contrast between atom-level graphs and fragment-based methods could be sharpened by citing concrete failure modes (e.g., loss of stereochemistry in principal-subgraph approaches) with a brief illustrative example.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and the recommendation for major revision. The comments highlight important areas for strengthening the attribution of our results and improving experimental rigor. We address each major comment below and commit to revisions that will enhance the manuscript.

read point-by-point responses

  1. Referee: [Abstract and §4] Abstract and §4 (Experiments): The reported 2-4% correlation improvements and 1-3% virtual-screening gains are presented without ablation studies that isolate OverlapBPE (with its overlapping fragments and enriched token features) and the h-MINT many-to-many interaction layers while holding data splits, random seeds, optimizer schedules, and preprocessing pipelines fixed against the cited baselines. This leaves the central attribution claim vulnerable to alternative explanations such as more favorable partitioning or hyperparameter differences.

    Authors: We agree that the absence of controlled ablation studies weakens the causal attribution of the reported gains to OverlapBPE and the hierarchical layers. In the revised manuscript, we will add a new subsection in §4 dedicated to ablations. These will include: (i) OverlapBPE versus standard non-overlapping BPE and atom-level baselines using identical data splits, random seeds, optimizer schedules, and preprocessing; (ii) h-MINT versus a flat (non-hierarchical) model with the same tokenization; and (iii) variants ablating the many-to-many interaction mappings. All experiments will be run under fixed conditions to directly address alternative explanations such as partitioning or hyperparameter differences. revision: yes

  2. Referee: [Abstract] Abstract: No information is supplied on baseline re-implementations, error bars, statistical significance tests (e.g., paired t-tests or bootstrap confidence intervals), or variance across multiple runs, making it impossible to judge whether the stated improvements exceed experimental noise.

    Authors: We concur that details on re-implementations, variance, and statistical testing are required to evaluate whether the 2-4% and 1-3% gains exceed noise. In the revised manuscript we will: expand the abstract and §4 with explicit descriptions of baseline re-implementations (including code availability or matching hyperparameter settings); report mean ± standard deviation over at least five independent runs with different random seeds; and include statistical significance results (paired t-tests and bootstrap confidence intervals) comparing our method against each baseline. These additions will allow readers to assess the improvements relative to experimental variability. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical evaluation on external benchmarks is self-contained.

full rationale

The paper introduces OverlapBPE tokenization and the h-MINT hierarchical architecture as modeling choices, then reports performance on standard external benchmarks (PDBBind, LBA, DUD-E, LIT-PCBA, PubChem assays) against prior SOTA methods. No derivation chain is presented that reduces a claimed result to its own fitted inputs or self-citations by construction. The reported 2-4% correlation lifts and 1-3% screening gains are outputs of trained models evaluated on held-out data, which is standard empirical validation rather than a self-definitional or fitted-input prediction. No load-bearing self-citation, uniqueness theorem, or ansatz smuggling is invoked in the provided text.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The performance claims rest on the assumption that overlapping fragments capture chemically meaningful contexts and that the hierarchical network can learn useful joint representations from standard binding datasets.

free parameters (1)
  • neural network weights
    All parameters of the h-MINT model are fitted during training on PDBBind and related datasets.
axioms (1)
  • domain assumption Overlapping molecular fragments preserve essential chemical properties such as chirality and aromaticity better than non-overlapping or atom-only representations
    This premise underpins the design of OverlapBPE and is invoked to justify the hierarchical architecture.

pith-pipeline@v0.9.0 · 5635 in / 1302 out tokens · 66521 ms · 2026-05-08T08:34:42.840999+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

42 extracted references · 23 canonical work pages

  1. [1]

    Learning 3d representations of molecular chirality with invariance to bond rotations.arXiv preprint arXiv:2110.04383,

    Keir Adams, Lagnajit Pattanaik, and Connor W Coley. Learning 3d representations of molecular chirality with invariance to bond rotations.arXiv preprint arXiv:2110.04383,

    work page arXiv

  2. [2]

    Protein structure and sequence generation with equivariant denoising diffusion probabilistic models.arXiv preprint arXiv:2205.15019, 2022

    Namrata Anand and Tudor Achim. Protein structure and sequence generation with equivariant denoising diffusion probabilistic models.arXiv preprint arXiv:2205.15019,

    work page arXiv

  3. [3]

    Learning protein sequence embeddings using information from structure.arXiv preprint arXiv:1902.08661, 2019

    Tristan Bepler and Bonnie Berger. Learning protein sequence embeddings using information from structure.arXiv preprint arXiv:1902.08661,

    work page arXiv 1902

  4. [4]

    A new perspective on building efficient and expressive 3d equivariant graph neural networks.arXiv preprint arXiv:2304.04757,

    Weitao Du, Yuanqi Du, Limei Wang, Dieqiao Feng, Guifeng Wang, Shuiwang Ji, Carla Gomes, and Zhi-Ming Ma. A new perspective on building efficient and expressive 3d equivariant graph neural networks.arXiv preprint arXiv:2304.04757,

    work page arXiv

  5. [5]

    Aosong Feng, Chenyu You, Shiqiang Wang, and Leandros Tassiulas

    doi: 10.1109/TPAMI.2021.3095381. Aosong Feng, Chenyu You, Shiqiang Wang, and Leandros Tassiulas. Kergnns: Interpretable graph neural networks with graph kernels. InProceedings of the AAAI conference on artificial intelligence, volume 36, pp. 6614–6622,

    work page doi:10.1109/tpami.2021.3095381 2021

  6. [6]

    A foundation model for protein-ligand affinity prediction through jointly optimizing virtual screening and hit-to-lead optimization.bioRxiv, pp

    Bin Feng, Zijing Liu, Mingjun Yang, Junjie Zou, He Cao, Yu Li, Lei Zhang, and Sheng Wang. A foundation model for protein-ligand affinity prediction through jointly optimizing virtual screening and hit-to-lead optimization.bioRxiv, pp. 2025–02,

    2025

  7. [7]

    Protein-ligand binding representation learning from fine-grained interactions

    Shikun Feng, Minghao Li, Yinjun Jia, Wei-Ying Ma, and Yanyan Lan. Protein-ligand binding representation learning from fine-grained interactions. InThe Twelfth International Conference on Learning Representations. 11 Published as a conference paper at ICLR 2026 Pablo Gainza, Freyr Sverrisson, Frederico Monti, Emanuele Rodola, D Boscaini, MM Bronstein, and ...

    2026

  8. [8]

    Pifold: Toward effective and efficient protein inverse folding.arXiv preprint arXiv:2209.12643,

    Zhangyang Gao, Cheng Tan, Pablo Chacón, and Stan Z Li. Pifold: Toward effective and efficient protein inverse folding.arXiv preprint arXiv:2209.12643,

    work page arXiv

  9. [9]

    arXiv preprint arXiv:2011.14115 (2020)

    Johannes Gasteiger, Shankari Giri, Johannes T Margraf, and Stephan Günnemann. Fast and uncertainty-aware directional message passing for non-equilibrium molecules.arXiv preprint arXiv:2011.14115,

    work page arXiv 2011

  10. [10]

    De novo molecular generation via connection-aware motif mining.arXiv preprint arXiv:2302.01129,

    Zijie Geng, Shufang Xie, Yingce Xia, Lijun Wu, Tao Qin, Jie Wang, Yongdong Zhang, Feng Wu, and Tie-Yan Liu. De novo molecular generation via connection-aware motif mining.arXiv preprint arXiv:2302.01129,

    work page arXiv

  11. [11]

    Intrinsic-extrinsic convolution and pooling for learning on 3d protein structures.arXiv preprint arXiv:2007.06252,

    Pedro Hermosilla, Marco Schäfer, Mat ˇej Lang, Gloria Fackelmann, Pere Pau Vázquez, Barbora Kozlíková, Michael Krone, Tobias Ritschel, and Timo Ropinski. Intrinsic-extrinsic convolution and pooling for learning on 3d protein structures.arXiv preprint arXiv:2007.06252,

    work page arXiv 2007

  12. [12]

    Deep contrastive learning enables genome-wide virtual screening

    Yinjun Jia, Bowen Gao, Jiaxin Tan, Jiqing Zheng, Xin Hong, Wenyu Zhu, Haichuan Tan, Yuan Xiao, Liping Tan, Hongyi Cai, et al. Deep contrastive learning enables genome-wide virtual screening. bioRxiv, pp. 2024–09,

    2024

  13. [13]

    Antibody-antigen docking and design via hierarchical equivariant refinement.arXiv preprint arXiv:2207.06616,

    Wengong Jin, Regina Barzilay, and Tommi Jaakkola. Antibody-antigen docking and design via hierarchical equivariant refinement.arXiv preprint arXiv:2207.06616,

    work page arXiv

  14. [14]

    arXiv preprint arXiv:2009.01411 , year=

    Bowen Jing, Stephan Eismann, Patricia Suriana, Raphael JL Townshend, and Ron Dror. Learning from protein structure with geometric vector perceptrons.arXiv preprint arXiv:2009.01411,

    work page arXiv 2009

  15. [15]

    Equivariant graph neural networks for 3d macromolecular structure.arXiv preprint arXiv:2106.03843,

    12 Published as a conference paper at ICLR 2026 Bowen Jing, Stephan Eismann, Pratham N Soni, and Ron O Dror. Equivariant graph neural networks for 3d macromolecular structure.arXiv preprint arXiv:2106.03843,

    work page arXiv 2026

  16. [16]

    Conditional antibody design as 3d equivariant graph translation.arXiv preprint arXiv:2208.06073, 2022a

    Xiangzhe Kong, Wenbing Huang, and Yang Liu. Conditional antibody design as 3d equivariant graph translation.arXiv preprint arXiv:2208.06073, 2022a. Xiangzhe Kong, Wenbing Huang, Zhixing Tan, and Yang Liu. Molecule generation by principal subgraph mining and assembling.Advances in Neural Information Processing Systems, 35: 2550–2563, 2022b. Xiangzhe Kong, ...

    work page arXiv

  17. [17]

    Frequent subgraph discovery

    Michihiro Kuramochi and George Karypis. Frequent subgraph discovery. InProceedings 2001 IEEE international conference on data mining, pp. 313–320. IEEE,

    2001

  18. [18]

    arXiv preprint arXiv:2206.11990 , year=

    Yi-Lun Liao and Tess Smidt. Equiformer: Equivariant graph attention transformer for 3d atomistic graphs.arXiv preprint arXiv:2206.11990,

    work page arXiv

  19. [19]

    Antigen-specific antibody design and optimization with diffusion-based generative models.bioRxiv, pp

    Shitong Luo, Yufeng Su, Xingang Peng, Sheng Wang, Jian Peng, and Jianzhu Ma. Antigen-specific antibody design and optimization with diffusion-based generative models.bioRxiv, pp. 2022–07,

    2022

  20. [20]

    Large-scale assessment of binding free energy calculations in active drug discovery projects.Journal of Chemical Information and Modeling, 60(11):5457–5474,

    13 Published as a conference paper at ICLR 2026 Christina EM Schindler, Hannah Baumann, Andreas Blum, Dietrich Bose, Hans-Peter Buchstaller, Lars Burgdorf, Daniel Cappel, Eugene Chekler, Paul Czodrowski, Dieter Dorsch, et al. Large-scale assessment of binding free energy calculations in active drug discovery projects.Journal of Chemical Information and Mo...

    2026

  21. [21]

    Graphbpe: Molecular graphs meet byte-pair encoding.arXiv preprint arXiv:2407.19039,

    Yuchen Shen and Barnabás Póczos. Graphbpe: Molecular graphs meet byte-pair encoding.arXiv preprint arXiv:2407.19039,

    work page arXiv

  22. [22]

    Stepniewska-Dziubinska, Piotr Zielenkiewicz, and Paweł Siedlecki

    Marta M. Stepniewska-Dziubinska, Piotr Zielenkiewicz, and Paweł Siedlecki. Pafnucy - a deep neural network for structure-based drug discovery.ArXiv, abs/1712.07042,

    work page arXiv

  23. [23]

    TorchMD-NET: Equivariant Transform ers for Neu- ral Network based Molecular Potentials

    Philipp Thölke and Gianni De Fabritiis. Torchmd-net: equivariant transformers for neural network based molecular potentials.arXiv preprint arXiv:2202.02541,

    work page arXiv

  24. [24]

    arXiv preprint arXiv:2012.04035 , year=

    Raphael JL Townshend, Martin Vögele, Patricia Suriana, Alexander Derry, Alexander Powers, Yianni Laloudakis, Sidhika Balachandar, Bowen Jing, Brandon Anderson, Stephan Eismann, et al. Atom3d: Tasks on molecules in three dimensions.arXiv preprint arXiv:2012.04035,

    work page arXiv 2012

  25. [25]

    On the art of compiling and using’drug-like’chemical fragment spaces

    14 Published as a conference paper at ICLR 2026 Christof Wegscheid-Gerlach, Andrea Zaliani, and Matthias Rarey. On the art of compiling and using’drug-like’chemical fragment spaces. Daniel S Wigh, Jonathan M Goodman, and Alexei A Lapkin. A review of molecular representation in the age of machine learning.Wiley Interdisciplinary Reviews: Computational Mole...

    2026

  26. [26]

    Geodiff: A geometric diffusion model for molecular conformation generation

    Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. Geodiff: A geometric diffusion model for molecular conformation generation.arXiv preprint arXiv:2203.02923,

    work page arXiv

  27. [27]

    International Conference on Learning Representations , year =

    Sheheryar Zaidi, Michael Schaarschmidt, James Martens, Hyunjik Kim, Yee Whye Teh, Alvaro Sanchez-Gonzalez, Peter Battaglia, Razvan Pascanu, and Jonathan Godwin. Pre-training via denoising for molecular property prediction.arXiv preprint arXiv:2206.00133,

    work page arXiv

  28. [28]

    medRxiv (2023).https://doi.org/10.1101/2023

    URL https://doi.org/10.1101/2023. 02.01.526585. Liangzhen Zheng, Jingrong Fan, and Yuguang Mu. Onionnet: a multiple-layer intermolecular-contact- based convolutional neural network for protein–ligand binding affinity prediction.ACS omega, 4 (14):15956–15965,

    work page doi:10.1101/2023 2023

  29. [29]

    URL https: //openreview.net/forum?id=6K2RM6wVqKu. 15 Published as a conference paper at ICLR 2026 A ETHICSSTATEMENT Small-molecule modeling plays a critical role in drug discovery, with broad potential applications in therapeutic development, virtual screening, and rational design of ligands targeting protein pockets. Advances in representation learning a...

    2026

  30. [30]

    The code and models used for evaluation are also publicly accessible and cited in the appendix

    codebases. The code and models used for evaluation are also publicly accessible and cited in the appendix. Furthermore, we describe the training hyperparameters in detail in the appendix, thereby ensuring that the entire experimental process is fully reproducible. C METHOD C.1 OVERLAPBPE c1cscn1c1c[nH]cn1c1ccccc1cS(N)(=O)=O c1cscn1nc[nH]NS(=O)(=O)c1ccccc1...

    2026

  31. [31]

    The other is the invariant channel, which mainly encodes and predicts embeddings (i.e., H)

    architecture with 2-channel updates: One is the equivariant channel, which mainly encodes and predict the coordinates of molecules following SE-(3) symmetry. The other is the invariant channel, which mainly encodes and predicts embeddings (i.e., H). Thus, in general, our model is an SE-(3) equivariant model. Since we only use the embedding channel in this...

    2026

  32. [32]

    Inspired by the good performance and trends in joint encoder models, we also adopt a joint encoder architecture

    instead use a joint encoder for pockets and ligands. Inspired by the good performance and trends in joint encoder models, we also adopt a joint encoder architecture. For theLBAdataset, SchNet (Schütt et al., 2018), DimeNet++ (Gasteiger et al., 2020), GemNet (Gasteiger et al.,

    2018

  33. [33]

    EGNN (Satorras et al., 2021), TorchMD-Net (ET) (Thölke & De Fabritiis, 2022), and LEFTNet (Du et al.,

    are invariant models based on invariant geometric features (i.e., distance and angle). EGNN (Satorras et al., 2021), TorchMD-Net (ET) (Thölke & De Fabritiis, 2022), and LEFTNet (Du et al.,

    2021

  34. [34]

    We also include atom-level pretrained models, UniMol (Zhou et al., 2023), ProFSA (Gao et al.) and BigBind (Feng et al.)

    utilize harmonic and irreducible representations to preserve high-order equivariant features. We also include atom-level pretrained models, UniMol (Zhou et al., 2023), ProFSA (Gao et al.) and BigBind (Feng et al.). In general, all these models mainly use their invariant channel for affinity prediction, similar to GET (Kong et al.,

    2023

  35. [35]

    and our model; thus, we can ignore how these models deal with equivariant features in these experiments. We take the baseline results mainly from GET (Kong et al., 2024), which provides a complete comparison of all the above models in 3 representation settings: atom-level, fragment-level, and bi-level. To save space, we include each model’s best represent...

    2024

  36. [36]

    And the results show p-values <0.005 for these models in both PDBBind and LBA tasks

    Besides, we also conduct a significance test on the prediction results of GET, GET-PS and our model. And the results show p-values <0.005 for these models in both PDBBind and LBA tasks. This evidence suggests that our model performs significantly better than the strong baselines, GET, and GET-PS. PDBBind Results Analysis.From Table 1, we can mainly draw t...

    2024

  37. [37]

    Table 7: Ablation Study of OverlapBPE and h-MINT on LBA. RMSE↓Pearson↑Spearman↑ GET 1.331 ± 0.008 0.618 ± 0.005 0.607 ± 0.005 GET+PS 1.312 ± 0.016 0.631 ± 0.011 0.642 ± 0.011 h-MINT+PS 1.321 ± 0.010 0.633 ± 0.007 0.641 ± 0.008 GET+OverlapBPE N/A N/A N/A Ours (h-MINT+OverlapBPE)1.276 ± 0.011 0.660 ± 0.001 0.661 ± 0.001 As can be seen in this table, GET is ...

    2026

  38. [38]

    Therefore, we trained it for 25 epochs and averaged the last 3 checkpoints for evaluation purposes

    We trained the model for 100 epochs initially, and observed that it converged at around the 25th epoch. Therefore, we trained it for 25 epochs and averaged the last 3 checkpoints for evaluation purposes. E.3 EVALUATIONMETRICS We assess model performance using the following metrics: 20 Published as a conference paper at ICLR 2026 Table 8: Zero-shot Virtual...

    work page arXiv 2026

  39. [39]

    Effect of the proposed auxiliary loss

    The results on DUDE and LIT-PCBA are shown in the table above, which confirms the following findings: (i). Effect of the proposed auxiliary loss. Using our proposed auxiliary loss consistently improves LigUnity across almost all metrics on both datasets. For example, on DUDE, AUC improves from 81.69 to 82.57, and BEDROC from 46.01 to 47.58. On LIT-PCBA, e...

    work page arXiv 2026

  40. [40]

    Run-to-run variation is small (std <0.006 across metrics), indicating stable behavior. These trends suggest a bias-variance trade-off: overly strict thresholds (very small vocab) underfit by missing informative fragments, whereas overly lax thresholds (very large vocab) admit rare or redundant fragments that increase sparsity and noise. A moderate thresho...

    2026

  41. [41]

    Statistical analysis of fragments in each cluster, as shown in Figure 7, revealed that Cluster 5 exhibited significant chemical specificity

    The results showed that all fragments clustered into 6 distinct categories in the latent space. Statistical analysis of fragments in each cluster, as shown in Figure 7, revealed that Cluster 5 exhibited significant chemical specificity. This cluster was predominantly enriched with functional groups containing lone electron pairs on N and O atoms, such as ...

    2026

  42. [42]

    Table 13: Molecular Property Prediction Benchmarks from MoleculeNet

    The significant improvements in prediction error confirm that OverlapBPE provides discriminative representations for molecules. Table 13: Molecular Property Prediction Benchmarks from MoleculeNet. RMSE ESOL↓FreeSolv↓Lipo↓ ECFP 1.5668 3.9498 0.8875 ECFP + OverlapBPE1.2972 3.3409 0.8270 24 Published as a conference paper at ICLR 2026 F.6 EXAMPLETOKENS ANDCH...

    2026