RGCN

RGCN on GlycanML class Accuracy: GlycanGT study: class Accuracy

Taxonomy: 13,209 glycans total across eight levels, 4–1,737 classes per level. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported

RGCN on GlycanML class Macro-F1: GlycanGT study: class Macro-F1

Taxonomy: 13,209 glycans total across eight levels, 4–1,737 classes per level. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported

RGCN on GlycanML domain Accuracy: GlycanGT study: domain Accuracy

Taxonomy: 13,209 glycans total across eight levels, 4–1,737 classes per level. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported

RGCN on GlycanML domain Macro-F1: GlycanGT study: domain Macro-F1

Taxonomy: 13,209 glycans total across eight levels, 4–1,737 classes per level. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported

RGCN on GlycanML family Accuracy: GlycanGT study: family Accuracy

Taxonomy: 13,209 glycans total across eight levels, 4–1,737 classes per level. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported

RGCN on GlycanML family Macro-F1: GlycanGT study: family Macro-F1

Taxonomy: 13,209 glycans total across eight levels, 4–1,737 classes per level. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported

RGCN on GlycanML genus Accuracy: GlycanGT study: genus Accuracy

Taxonomy: 13,209 glycans total across eight levels, 4–1,737 classes per level. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported

RGCN on GlycanML genus Macro-F1: GlycanGT study: genus Macro-F1

Taxonomy: 13,209 glycans total across eight levels, 4–1,737 classes per level. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported

RGCN on GlycanML glycosylation Accuracy: GlycanGT study: glycosylation Accuracy

Glycosylation: 1,683 glycans total; N-linked/O-linked/free. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported

RGCN on GlycanML glycosylation Macro-F1: GlycanGT study: glycosylation Macro-F1

Glycosylation: 1,683 glycans total; N-linked/O-linked/free. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported

RGCN on GlycanML immunogenicity Accuracy: GlycanGT study: immunogenicity Accuracy

Immunogenicity: 1,320 glycans total; binary immune activity. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported

RGCN on GlycanML immunogenicity AUPRC: GlycanGT study: immunogenicity AUPRC

Immunogenicity: 1,320 glycans total; binary immune activity. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported

RGCN on GlycanML immunogenicity Macro-F1: GlycanGT study: immunogenicity Macro-F1

Immunogenicity: 1,320 glycans total; binary immune activity. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported

RGCN on GlycanML kingdom Accuracy: GlycanGT study: kingdom Accuracy

Taxonomy: 13,209 glycans total across eight levels, 4–1,737 classes per level. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported

RGCN on GlycanML kingdom Macro-F1: GlycanGT study: kingdom Macro-F1

Taxonomy: 13,209 glycans total across eight levels, 4–1,737 classes per level. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported

RGCN on GlycanML order Accuracy: GlycanGT study: order Accuracy

Taxonomy: 13,209 glycans total across eight levels, 4–1,737 classes per level. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported

RGCN on GlycanML order Macro-F1: GlycanGT study: order Macro-F1

Taxonomy: 13,209 glycans total across eight levels, 4–1,737 classes per level. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported

RGCN on GlycanML phylum Accuracy: GlycanGT study: phylum Accuracy

Taxonomy: 13,209 glycans total across eight levels, 4–1,737 classes per level. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported

RGCN on GlycanML phylum Macro-F1: GlycanGT study: phylum Macro-F1

Taxonomy: 13,209 glycans total across eight levels, 4–1,737 classes per level. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported

RGCN on GlycanML species Accuracy: GlycanGT study: species Accuracy

Taxonomy: 13,209 glycans total across eight levels, 4–1,737 classes per level. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported

RGCN on GlycanML species Macro-F1: GlycanGT study: species Macro-F1

Taxonomy: 13,209 glycans total across eight levels, 4–1,737 classes per level. Official fixed GlycanML motif-based train/validation/test splits (8:1:1). Section 2.5 reports class-balanced classifiers, train ∪ validation hyperparameter selection by randomized search with 3-fold cross-validation, followed by one evaluation on the held-out test set; complete procedure repeated with three random seeds, reporting mean and standard deviation. GlycanGT large model pretrained with 35% masking provides [Graph] embeddings to SVM/LightGBM; the selected classifier for each S4 row is not identified. Section 2.6 states that all four graph baselines were trained and evaluated on the same datasets/splits; it does not establish that each baseline used the GlycanGT downstream classifier search. SVM search: 10 iterations, RBF/linear, C logU(1e-3,1e2), gamma logU(1e-4,1e-1); LightGBM: 15 randomized iterations. Do not combine with separate original GlycanML-paper protocols.

Aggregation: Not reported