Peptides-func

Predict all applicable therapeutic-function labels for each peptide. A peptide may have several of the 10 labels, so this is graph-level multi-label classification evaluated with the unweighted mean average precision across labels.

The benchmark contains 15,535 peptides selected from SATPdb, which itself aggregates and manually curates peptide records from 22 source databases. Each peptide is represented as a molecular graph whose nodes are non-hydrogen atoms and whose edges are covalent bonds.

The graphs reuse OGB molecular featurization computed from SMILES: each atom has nine categorical feature fields and each bond has three. The benchmark deliberately omits peptide 2D folding and 3D conformation; its graph describes covalent connectivity rather than the peptide’s folded spatial structure.

The fixed split is 70% training, 15% validation, and 15% test. It is stratified by the complete 10-label combination, with combinations occurring fewer than 10 times pooled into one meta-class. The labels are strongly imbalanced, which is why average precision rather than accuracy is the primary metric.

The 10 labels are correlated and non-exclusive; treating the task as single-label classification is incorrect. SATPdb combines records from many upstream databases, and many peptides have multiple functions. Randomly resplitting the data or adding predicted structural information changes the published LRGB protocol. The benchmark’s long atom-level chains should not be mistaken for experimentally determined peptide conformations.