Research map / Probabilistic and scientific modeling / Molecules and proteins
GlycanML: A Multi-Task and Multi-Structure Benchmark for Glycan Machine Learning
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.
Builds GlycanML, a benchmark of 11 glycan tasks (8 taxonomy levels, immunogenicity, glycosylation type, protein-glycan interaction) with sequence and graph representations, plus a multi-task testbed comparing MTL loss-weighting methods.
Problem
There is no standard machine learning benchmark for glycan property and function prediction, and it is unclear how sequence-based models, GNNs, small-molecule encoders, and multi-task learning perform on glycan tasks.
Approach
Provides motif-based structural splits for generalization to unseen glycan structures, evaluates sequence encoders, homogeneous/heterogeneous GNNs and all-atom molecular encoders on both tokenized-sequence and planar-graph glycan representations, and introduces GlycanML-MTL with eight taxonomy tasks to test MTL methods.
Key results
- RGCN ranks first by weighted mean rank (2.19) and is best on 6 of 11 tasks in single-task learning.
- Temperature scaling gives the best mean Macro-F1 on shallow CNN, a 2.75% improvement over single-task learning.
- Multi-task learning with taxonomy tasks improves immunogenicity and glycosylation prediction under both Shallow CNN and RGCN, but not interaction prediction.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use temperature-scaling MTL instead of single-task training | Shallow CNN · Glycan taxonomy prediction, mean across 8 tasks | Mean Macro-F1: 0.327 (0.003) -> 0.336 (0.014); 2.75% improvement | Helped |
| Use temperature-scaling MTL instead of single-task training | RGCN · Glycan taxonomy prediction, mean across 8 tasks | Mean Macro-F1: 0.360 (0.009) -> 0.368 (0.016) | Helped |
| Use naive equal-weight MTL instead of single-task training | Shallow CNN · Glycan taxonomy prediction, mean across 8 tasks | Mean Macro-F1: 0.327 (0.003) -> 0.312 (0.010) | Hurt |
| Use gradient normalization MTL instead of single-task training | Shallow CNN · Glycan taxonomy prediction, mean across 8 tasks | Mean Macro-F1: 0.327 (0.003) -> 0.319 (0.012) | Hurt |
| Use uncertainty-weighting MTL instead of single-task training | RGCN · Glycan taxonomy prediction, mean across 8 tasks | Mean Macro-F1: 0.360 (0.009) -> 0.341 (0.019) | Hurt |
| Use dynamic task prioritization MTL instead of single-task training | RGCN · Glycan taxonomy prediction, mean across 8 tasks | Mean Macro-F1: 0.360 (0.009) -> 0.362 (0.007) | No clear effect |
| Jointly train immunogenicity prediction with taxonomy prediction | Shallow CNN · Immunogenicity prediction | AUPRC: 0.776 (0.027) -> 0.792 (0.028) | Helped |
| Jointly train glycosylation prediction with taxonomy prediction | RGCN · Glycosylation type prediction | Macro-F1: 0.948 (0.004) -> 0.963 (0.009) | Helped |
| Jointly train interaction prediction with taxonomy prediction | Shallow CNN · Protein-glycan interaction prediction | Spearman's ρ: 0.261 (0.008) -> 0.257 (0.002) | Hurt |
| Jointly train interaction prediction with taxonomy prediction | RGCN · Protein-glycan interaction prediction | Spearman's ρ: 0.262 (0.005) -> 0.259 (0.030) | Hurt |
Surprising findings
- Graphormer and GraphGPS, strong small-molecule encoders, rank last on glycan tasks, performing poorly on glycans.
- Most MTL methods underperform single-task learning on the taxonomy testbed; only TS and UW (shallow CNN) and TS and DTP (RGCN) beat single-task on mean Macro-F1.
Limitations
- Glycan taxonomy and function tasks are evaluated only with a fixed set of encoders and a simple MLP head; no pretrained glycan language models are tested.
- MTL methods generally fail to beat single-task learning, so the MTL testbed remains open.
- Interaction prediction is not improved by MTL with taxonomy.
Experimental setup
Datasets: SugarBase GlyConnect LectinOracle GlyTouCan
Benchmarks: GlycanML GlycanML-MTL
Metrics: Macro-F1 AUPRC Spearman's rho weighted mean rank
Baselines: Shallow CNN ResNet LSTM Transformer GCN GAT GIN MPNN CompGCN Graphormer GraphGPS
Builds on: RGCN CompGCN Graphormer GraphGPS Gradient Normalization Uncertainty Weighting Dynamic Weight Averaging Dynamic Task Prioritization
Closest papers by meaning
- Modeling All-Atom Glycan Structures via Hierarchical Message Passing and Multi-Scale Pre-training (ICML 2025 · similarity 0.92)
- Prot2Text: Multimodal Protein’s Function Generation with GNNs and Transformers (AAAI 2024 · similarity 0.84)
- BioXArena: Benchmarking LLM Agents on Multi-Modal Biomedical Machine Learning Tasks (NeurIPS 2026 · similarity 0.82)
- Expressivity and Generalization: Fragment-Biases for Molecular GNNs (ICML 2024 · similarity 0.82)
- Kernel-Based Evaluation of Conditional Biological Sequence Models (ICML 2024 · similarity 0.82)
- DPLM-2: A Multimodal Diffusion Protein Language Model (ICLR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.