REVIEW 4 major objections 5 minor 20 references
Modeling All-Atom Glycan Structures via Hierarchical Message Passing and Multi-Scale Pre-training
T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read The paper claims that glycans should be modeled by reading every heavy atom, not just sugar units, and that a hierarchical all-atom graph encoder plus multi-scale pre-training makes that practical and effective.
desk verdict Solid architecture and consistently strong per-task results, but the headline ranking is internally inconsistent (1.5/2.56 vs the 1.0/2.0 the table implies), so the SOTA claim needs a repair pass before it can be trusted. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the all-atom heterogeneous glycan graph $g=(V_a,V_m,E)$, whose atom nodes and monosaccharide nodes are connected by three edge families: covalent bonds ($E_{aa}$), atom–monosaccharide memberships ($E_{am}$), and glycosidic bonds ($E_{mm}$). The argument is carried by hierarchical message passing built from relational graph convolution (RGConv), run in three stages per block—atom–atom, atom–monosaccharide, monosaccharide–monosaccharide—so that local atomic information is distilled into backbone nodes before global backbone messages are exchanged. Pre-training rests on multi-scale mask prediction: with atom mask ratio $\rho_a=0.45$ and monosaccharide mask ratio $\rho_m=0.15$, masked atom and monosaccharide types are recovered by MLP predictors under cross-entropy loss on a cleaned 40,781-sample corpus. Monosaccharide-wise readout, meaning mean and max pooling over monosaccharide nodes only, converts the node representations into a glycan-level vector and is supported by ablation as superior to all-node readout.
What would settle it
Re-run every baseline in Table 1 under the authors' exact protocol, using the same three seeds, epoch counts, validation-based model selection, and data splits, then recompute weighted mean ranks; if PreRGCN or another baseline then matches or beats GlycanAA's 2.56, the reported ranking would not survive.
Extended reading notes
Core claim
The central claim is that previous all-atom glycan modeling failed not because atomic detail is useless, but because the atomic graphs were fed to encoders designed for small molecules. GlycanAA instead keeps two granularities in one heterogeneous graph: heavy atoms that form each monosaccharide and monosaccharide nodes that form the glycan backbone, with covalent bonds, atom–monosaccharide memberships, and glycosidic bonds as separate relation types. Hierarchical message passing first updates atoms, then pushes atomic information into monosaccharide nodes, then exchanges messages along glycosidic bonds, and the final readout pools only the monosaccharide nodes. Pre-training then masks a fraction of atoms and entire monosaccharides and trains the model to recover their types, teaching dependencies at both scales. On the GlycanML benchmark, PreGlycanAA and GlycanAA achieve weighted mean ranks of 1.5 and 2.56, versus 5.34 for the strongest baseline PreRGCN, and PreGlycanAA improves on GlycanAA across all 11 tasks.
Load-bearing premise
The state-of-the-art claim assumes that the baseline numbers in Table 1 were produced under the same protocol as the authors' runs, since the paper does not state whether those baselines were re-run with identical seeds, epochs, splits, and checkpoint selection.
Editorial extensions
If this is right
- All-atom modeling is affordable: GlycanAA is only about 22% slower and uses about 19% more memory than the monosaccharide-level RGCN, yet beats it on all 11 benchmark tasks.
- Pre-training from unlabeled glycan repositories transfers to downstream tasks, since PreGlycanAA outperforms GlycanAA on every one of the 11 tasks.
- Multi-scale mask prediction is a better pre-training objective than attribute masking or context prediction, beating both GlycanAA-Attribute and GlycanAA-Context as well as the pre-trained monosaccharide-level PreRGCN.
- Small-molecule 3D encoders such as Graphormer, GraphGPS, Uni-Mol+, and VabsNet do not transfer directly to glycan-sized all-atom graphs, so architecture matters for atomic-level glycan modeling.
- Mask ratios are consequential: downstream taxonomy performance peaks near an atom mask ratio of 0.45 and a monosaccharide mask ratio of 0.15, indicating a balance between masked and observed information.
Reading between the lines
- The same two-granularity hierarchical design could transfer to other branched biopolymers, such as glycogen, peptidoglycan, or glycosaminoglycans, where local residue chemistry and global linkage pattern jointly determine function.
- Because atomic substructures are shared across many monosaccharides, the all-atom representation may generalize better to rare taxonomic classes than monosaccharide-level encoders; per-class accuracy across the eight taxonomy tasks would test this directly.
- A testable extension is to pre-train on the full 219,857 GlyTouCan entries with the same cleaning criteria rather than the filtered 40,781, to see whether downstream gains continue to scale with unlabeled data.
- The multi-scale mask objective is a form of graph denoising, so it could be combined with conformer- or geometry-based pre-training if reliable 3D glycan conformations become available for the same molecules.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript introduces GlycanAA, a heterogeneous all-atom glycan graph encoder that performs hierarchical message passing over atom, atom–monosaccharide, and monosaccharide–monosaccharide relations, followed by a monosaccharide-wise readout. It further introduces PreGlycanAA, a self-supervised variant pre-trained on 40,781 curated GlyTouCan glycans using atom-scale and monosaccharide-scale mask prediction. The central empirical claim is that PreGlycanAA and GlycanAA rank first and second on the GlycanML benchmark, outperforming both monosaccharide-level glycan encoders and all-atom small-molecule encoders. The paper also presents ablations, an efficiency comparison with RGCN, and t-SNE visualizations.
Significance. If the comparative results are correct, the contribution is significant: all-atom glycan modeling is an underexplored representation level, hierarchical message passing is a natural inductive bias for glycans, and the multi-scale masked pre-training scheme with a leakage-screened corpus is a useful methodological addition. The paper includes extensive experiments, ablations, an efficiency study, visualizations, and a public repository link. However, the headline state-of-the-art claim is not verifiable from the printed evidence because Table 1 is internally inconsistent, and because the provenance of the baseline numbers is not documented. These are fixable issues, but they are load-bearing and need to be resolved before the empirical claims can be accepted.
major comments (4)
- [Table 1 and Section 5.2] The reported Weighted Mean Rank column is internally inconsistent with the per-task values and with the definition in Section 5.2. In every one of the 11 task columns, PreGlycanAA is strictly the largest value; with the stated weighting (each taxonomy task weight 1/8 and each of the other three tasks weight 1), its weighted mean rank must be exactly 1.0, not 1.5. For GlycanAA, the printed table gives ranks of [2, 4, 2, 2, 2, 2, 2, 2, 2, 2, 2] across the 11 tasks, because on Kingdom the GlycanAA-Attribute (0.687) and GlycanAA-AN (0.685) rows are above GlycanAA (0.683); this yields a weighted mean rank of about 2.06, not 2.56. The same section also states that GlycanAA outperforms the best baseline on 10 of 11 tasks, whereas a direct comparison with PreRGCN shows GlycanAA is ahead on all 11 columns. Please report the per-task rank vectors and recompute the summary column from a single, consistent evaluation; the current table cannot support the 'ranks first and second' claim.
- [Section 5.2, Baselines] The manuscript does not state whether the baseline numbers in Table 1 were re-run under the identical protocol or copied from the original GlycanML benchmark paper. Section 5.2 says the authors compare 'with the baselines studied in the GlycanML benchmark' but does not specify the seeds, epochs, splits, model selection procedure, or preprocessing used for each baseline row. Given that the printed weighted mean rank column is already inconsistent with the per-task values, the possibility that different rows come from different evaluation runs cannot be dismissed. Please state for every row whether it is a re-run or a reported value, and provide the exact configuration needed to reproduce all rows.
- [Section 5.3, Sensitivity of PreGlycanAA to mask ratio] The hyperparameters ρa and ρm are selected by pre-training models under 36 mask-ratio pairs and evaluating them on eight glycan taxonomy prediction tasks. The manuscript does not state whether this evaluation uses the same test sets as Table 1 or a separate validation split. If test labels were used for model selection, the reported gains of PreGlycanAA over GlycanAA on the taxonomy tasks would be optimistically biased, which would undermine the claim that the multi-scale mask prediction objective is responsible for the improvement. Please clarify the selection protocol and, if validation was not used, re-report the affected results.
- [Section 5.2, Results] No statistical significance testing is reported, and several of the claimed improvements are within one standard deviation of the comparison model on individual tasks. For example, PreGlycanAA versus GlycanAA on Kingdom is 0.688±0.001 versus 0.683±0.002, and on Glycosylation it is 0.961±0.011 versus 0.950±0.020. With only three seeds, the repeatedly stated conclusion that PreGlycanAA 'outperforms GlycanAA on all 11 tasks' is not statistically supported. Please report paired significance tests or additional seeds, or soften the claims to observed-score comparisons.
minor comments (5)
- [Section 5.3, Figure 4] The caption reports 'Average Macro-F1 score of PreGlycanAA on eight taxonomy prediction tasks' but the text does not state whether the plotted values are averaged over the three seeds used elsewhere or computed from a single seed; please clarify.
- [Table 1] The weighted mean rank column cannot be reconstructed from the current table formatting. Printing the per-task rank for each model, or a separate rank matrix, would make the summary column checkable.
- [Section 5.5 and Appendix A.3] The t-SNE visualizations do not include a color legend or the number of classes per panel, which limits the interpretation of the claimed 'better clustering' behavior.
- [Appendix A.4] The efficiency study reports throughput and memory on a single dataset at batch size 256; please state whether the throughput includes data loading and whether the memory cost is peak CUDA memory, so the comparison with RGCN is reproducible.
- [Reproducibility] The GitHub repository link is given, but the paper does not indicate whether pre-trained checkpoints, the curated pre-training dataset, and the exact train/validation/test split files are included; please add this information.
Circularity Check
No circular derivation: GlycanAA and PreGlycanAA rest on an independent architectural construction and self-supervised objective, though the GlycanML benchmark is co-authored by the present first author and Table 1's printed weighted mean ranks are internally inconsistent with its own per-task columns.
full rationale
Walking the derivation chain, there is no step in which a claimed prediction is equivalent by construction to an input. GlycanAA's representations are computed by the hierarchical RGConv message passing in Eqs. (1)-(4) on a heterogeneous graph defined from atom and monosaccharide structure, and the glycan-level readout is a mean/max pooling over monosaccharide nodes. PreGlycanAA's pre-training loss in Eq. (5) is a cross-entropy recovery objective on masked atom and monosaccharide types. None of these equations defines the downstream benchmark performance in terms of the model's own fitted parameters, and no reported result is forced by the architecture or by the pre-training objective alone. Two caveats prevent a clean verification of the headline claim, but neither is a circular reduction. First, the evaluation is conducted on GlycanML (Xu et al., 2024), a benchmark co-authored by the present first author, and the baselines are described as 'the baselines studied in the GlycanML benchmark' without stating whether all baselines were re-run under identical protocol; this is a self-citation and provenance concern, not a derivation-level circularity. Second, Table 1's printed weighted mean ranks are not consistent with the printed per-task columns under the metric defined in Section 5.2: PreGlycanAA is the maximum in all 11 columns and GlycanAA the second maximum in all 11 columns, which would give weighted mean ranks of 1.0 and 2.0 rather than the reported 1.5 and 2.56. This is an internal arithmetic or bookkeeping inconsistency in the summary metric, not an equivalence of a prediction to its inputs. The central architecture and multi-scale pre-training method are self-contained new designs, and the empirical comparison is in principle falsifiable on independent evaluation, so no significant circularity is found.
Assumptions & free parameters
free parameters (3)
- atom mask ratio rho_a =
0.45
- monosaccharide mask ratio rho_m =
0.15
- number of hierarchical blocks L =
3
assumptions (4)
- domain assumption The GlycanML benchmark tasks and metrics are a valid and comprehensive evaluation of glycan property prediction.
- domain assumption The all-atom glycan graph (heavy atoms plus covalent and glycosidic bond types, no hydrogen atoms, no 3D coordinates) captures sufficient structural signal for the benchmark tasks.
- domain assumption Self-supervised pre-training via multi-scale mask prediction transfers to downstream tasks.
- standard math Standard stochastic optimization and i.i.d. evaluation assumptions hold for all reported results.
Cite this review
Pith. "Pith review of Modeling All-Atom Glycan Structures via Hierarchical Message Passing and Multi-Scale Pre-training." pith.science (2026). https://pith.science/paper/FJGRF3TW
@misc{pith2026250601376,
author = {Pith},
title = {Pith review of: Modeling All-Atom Glycan Structures via Hierarchical Message Passing and Multi-Scale Pre-training},
year = {2026},
howpublished = {\url{https://pith.science/paper/FJGRF3TW}},
note = {Machine review of arXiv:2506.01376}
}
read the original abstract
Understanding the various properties of glycans with machine learning has shown some preliminary promise. However, previous methods mainly focused on modeling the backbone structure of glycans as graphs of monosaccharides (i.e., sugar units), while they neglected the atomic structures underlying each monosaccharide, which are actually important indicators of glycan properties. We fill this blank by introducing the GlycanAA model for All-Atom-wise Glycan modeling. GlycanAA models a glycan as a heterogeneous graph with monosaccharide nodes representing its global backbone structure and atom nodes representing its local atomic-level structures. Based on such a graph, GlycanAA performs hierarchical message passing to capture from local atomic-level interactions to global monosaccharide-level interactions. To further enhance model capability, we pre-train GlycanAA on a high-quality unlabeled glycan dataset, deriving the PreGlycanAA model. We design a multi-scale mask prediction algorithm to endow the model about different levels of dependencies in a glycan. Extensive benchmark results show the superiority of GlycanAA over existing glycan encoders and verify the further improvements achieved by PreGlycanAA. We maintain all resources at https://github.com/kasawa1234/GlycanAA
Figures
Figures from the paper (3 more)
Reference graph
Works this paper leans on
-
[2]
Dai, B., Mattox, D. E., and Bailey-Kellogg, C. Atten- tion please: modeling global and local context in glycan structure-function relationships. bioRxiv, pp. 2021–10,
work page 2021
-
[7]
Liang, S.-Y ., Wu, S.-W., Pu, T.-H., Chang, F.-Y ., and Khoo, K.-H. An adaptive workflow coupled with random forest algorithm to identify intact n-glycopeptides detected from mass spectrometry. Bioinformatics, 30(13):1908–1916,
work page 1908
-
[9]
N., Bloem, P., Van Den Berg, R., Titov, I., and Welling, M
Schlichtkrull, M., Kipf, T. N., Bloem, P., Van Den Berg, R., Titov, I., and Welling, M. Modeling relational data with graph convolutional networks. In The semantic web: 15th international conference, ESWC 2018, Heraklion, Crete, Greece, June 3–7, 2018, proceedings 15 , pp. 593–607. Springer,
work page 2018
-
[11]
Benchmarking graphormer on large-scale molecular modeling datasets
Shi, Y ., Zheng, S., Ke, G., Shen, Y ., You, J., He, J., Luo, S., Liu, C., He, D., and Liu, T.-Y . Benchmarking graphormer on large-scale molecular modeling datasets. arXiv preprint arXiv:2203.04810,
-
[14]
Learn- ing hierarchical protein representations via complete 3d graph networks
Wang, L., Liu, H., Liu, Y ., Kurtin, J., and Ji, S. Learn- ing hierarchical protein representations via complete 3d graph networks. In International Conference on Learning Representations, 2023a. Wang, X., Gu, R., Chen, Z., Li, Y ., Ji, X., Ke, G., and Wen, H. Uni-rna: universal pre-trained models revolutionize rna research. bioRxiv, pp. 2023–07, 2023b. X...
arXiv 2023
-
[15]
How powerful are graph neural networks? arXiv preprint arXiv:1810.00826,
Xu, K., Hu, W., Leskovec, J., and Jegelka, S. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826,
-
[16]
GlycanML: A Multi-Task and Multi-Structure Benchmark for Glycan Machine Learning
Xu, M., Geng, Y ., Zhang, Y ., Yang, L., Tang, J., and Zhang, W. Glycanml: A multi-task and multi-structure benchmark for glycan machine learning. arXiv preprint arXiv:2405.16206,
-
[17]
Zhang, D. et al. Dnagpt: A generalized pre-trained tool for versatile dna sequence analysis tasks. Preprint at https://doi. org/10.48550/arXiv, 2307, 2023a. Zhang, X.-L. Roles of glycans and glycopeptides in immune system and immune-related diseases. Current medicinal chemistry, 13(10):1141–1147,
Show all 20 references
-
[19]
Torchdrug: A powerful and flexible machine learning platform for drug discovery
Zhu, Z., Shi, C., Zhang, Z., Liu, S., Xu, M., Yuan, X., Zhang, Y ., Chen, J., Cai, H., Lu, J., et al. Torchdrug: A powerful and flexible machine learning platform for drug discovery. arXiv preprint arXiv:2202.08320,
-
[20]
Appendix A.1
12 Modeling All-Atom Glycan Structures via Hierarchical Message Passing and Multi-Scale Pre-training A. Appendix A.1. Accuracy and Perplexity Curves during Pre-training : Monosaccharide : Atom (a) Accuracy over Epochs during Pre-training (b) Perplexity over Epochs during Pre-t...
2018
-
[1997]
Strategies for pre-training graph neural networks
Hu, W., Liu, B., Gomes, J., Zitnik, M., Liang, P., Pande, V ., and Leskovec, J. Strategies for pre-training graph neural networks. arXiv preprint arXiv:1905.12265,
1905 arXiv
-
[2008]
Composition-based multi-relational graph convolutional networks
Vashishth, S., Sanyal, S., Nitin, V ., and Talukdar, P. Composition-based multi-relational graph convolutional networks. arXiv preprint arXiv:1911.03082,
1911 arXiv
-
[2014]
Language models of protein sequences at the scale of evolution enable accurate structure prediction
Lin, Z., Akin, H., Rao, R., Hie, B., Zhu, Z., Lu, W., dos Santos Costa, A., Fazel-Zarandi, M., Sercu, T., Candido, S., et al. Language models of protein sequences at the scale of evolution enable accurate structure prediction. BioRxiv, 2022:500902,
2022
-
[2017]
Graph attention networks.arXiv preprint arXiv:1710.10903,
Veliˇckovi´c, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., and Bengio, Y . Graph attention networks.arXiv preprint arXiv:1710.10903,
-
[2018]
Is trans- fer learning necessary for protein landscape prediction? arXiv preprint arXiv:2011.03443,
10 Modeling All-Atom Glycan Structures via Hierarchical Message Passing and Multi-Scale Pre-training Shanehsazzadeh, A., Belanger, D., and Dohan, D. Is trans- fer learning necessary for protein landscape prediction? arXiv preprint arXiv:2011.03443,
2011 arXiv
-
[2020]
P., Kozl´ıkov´a, B., Krone, M., Ritschel, T., and Ropinski, T
Hermosilla, P., Sch ¨afer, M., Lang, M., Fackelmann, G., V´azquez, P. P., Kozl´ıkov´a, B., Krone, M., Ritschel, T., and Ropinski, T. Intrinsic-extrinsic convolution and pooling for learning on 3d protein structures. arXiv preprint arXiv:2007.06252,
2007 arXiv
-
[2021]
Bert: Pre-training of deep bidirectional trans- formers for language understanding
Devlin, J. Bert: Pre-training of deep bidirectional trans- formers for language understanding. arXiv preprint arXiv:1810.04805,
-
[2022]
M., and Collins, J
Bojar, D., Camacho, D. M., and Collins, J. J. Using natural language processing to learn the grammar of glycans. bioRxiv, pp. 2020–01, 2020a. Bojar, D., Powers, R. K., Camacho, D. M., and Collins, J. J. Sweetorigins: Extracting evolutionary information from glycans. bioRxiv, p...
2020
-
[2023]
J., Oktay, D., Lin, Z., Verkuil, R., Tran, V
Hayes, T., Rao, R., Akin, H., Sofroniew, N. J., Oktay, D., Lin, Z., Verkuil, R., Tran, V . Q., Deaton, J., Wiggert, M., et al. Simulating 500 million years of evolution with a language model. bioRxiv, pp. 2024–07,
2024
-
[2024]
Generrna: A generative pre-trained language model for de novo rna design
11 Modeling All-Atom Glycan Structures via Hierarchical Message Passing and Multi-Scale Pre-training Zhao, Y ., Oono, K., Takizawa, H., and Kotera, M. Generrna: A generative pre-trained language model for de novo rna design. bioRxiv, pp. 2024–02,
2024
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.