Peptides-struct

Predict 11 graph-level geometric descriptors computed from each peptide’s 3D structure. The targets cover three mass-weighted inertias, three valence-weighted inertias, three principal-axis lengths, sphericity, and distance from a best-fit plane; each target is standardized to zero mean and unit variance.

The 15,535 peptides are the same SATPdb-derived molecules used by Peptides-func. Nodes are non-hydrogen atoms and edges are covalent bonds, giving much larger molecular graphs than typical small-molecule benchmarks.

Inputs use OGB’s nine categorical atom fields and three categorical bond fields derived from SMILES. Although the targets were computed from 3D peptide structures, neither coordinates nor other 2D/3D structural information are included in the input graphs.

Peptides-struct uses the same fixed 70%/15%/15% stratified split as Peptides-func, where stratification is based on pooled combinations of the functional labels. The primary score is the unweighted mean absolute error across the 11 standardized targets; per-target R-squared is supplementary.

The task asks a model to infer global 3D properties from covalent connectivity without a supplied conformation. Target standardization is part of the benchmark, so MAE on unstandardized physical quantities is not comparable. The split is stratified by function rather than structural scaffold, and correlated geometric targets must not be treated as independent datasets.