Research map / Graph learning / Graph self-supervised learning

Modeling All-Atom Glycan Structures via Hierarchical Message Passing and Multi-Scale Pre-training

ICML 2025 · Graph learning › Graph self-supervised learning · approach: Brain and biological graph learning · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes GlycanAA, an all-atom heterogeneous glycan graph encoder with hierarchical message passing, and PreGlycanAA, which is pre-trained with multi-scale atom/monosaccharide mask prediction on 40,781 GlyTouCan glycans, evaluated on GlycanML.

Problem

Existing glycan machine learning models model glycans only at the monosaccharide level, ignoring atomic structure, while small-molecule all-atom encoders perform poorly on glycans with hundreds of atoms.

Approach

Represent a glycan as a heterogeneous graph with atom nodes and monosaccharide nodes, apply hierarchical message passing (atom-atom, then atom-monosaccharide, then monosaccharide-monosaccharide) with RGConv, and pre-train with multi-scale masking of atoms and monosaccharides.

Key results

• PreGlycanAA achieves the best weighted mean rank of 1.5 on GlycanML, outperforming all baselines and improving over GlycanAA on all 11 tasks.
• GlycanAA outperforms the best baseline on 10 of 11 tasks with weighted mean rank 2.56, while all-atom small-molecule encoders (Graphormer, GraphGPS, Uni-Mol+) perform poorly on glycans.
• GlycanAA is about 22% slower than RGCN in training and inference and uses about 19% more memory.

Measured effects reported in the paper

Surprising findings

• Off-the-shelf small-molecule all-atom encoders (Graphormer, GraphGPS, Uni-Mol+) perform worse than monosaccharide-level models on glycans with hundreds of atoms.
• All-node readout including atoms underperforms monosaccharide-only readout on most tasks.

Limitations

• Atomic-level modeling incurs extra computational cost relative to monosaccharide-level RGCN.
• Pre-training dataset is limited to 40,781 fully solved, single-component glycans after filtering.

Experimental setup

Datasets: GlyTouCan GlycanML

Benchmarks: GlycanML

Metrics: Macro-F1 AUPRC Spearman's rho weighted mean rank

Baselines: Transformer Shallow CNN LSTM ResNet MPNN GCN GAT GIN CompGCN RGCN GearNet GearNet-Edge

Builds on: relational graph convolution (RGCN) attribute masking context prediction GlycanML

Closest papers by meaning

• GlycanML: A Multi-Task and Multi-Structure Benchmark for Glycan Machine Learning (ICLR 2025 · similarity 0.92)
• Bridging Molecular Graphs and Large Language Models (AAAI 2025 · similarity 0.84)
• Proteina: Scaling Flow-based Protein Structure Generative Models (ICLR 2025 · similarity 0.83)
• Atomic Diffusion Models for Small Molecule Structure Elucidation from NMR Spectra (NeurIPS 2025 · similarity 0.83)
• Hierarchical Graph Tokenization for Molecule-Language Alignment (ICML 2025 · similarity 0.83)
• On the Scalability of GNNs for Molecular Graphs (NeurIPS 2024 · similarity 0.83)