Scaffold-aware benchmarking of GCN and GIN for DFT-derived molecular Gibbs energy in traditional Chinese medicine molecules.
Absolute gas-phase molecular Gibbs energy G is an extensive quantum-chemical quantity that can be dominated by elemental composition, yet graph-model benchmarks often omit a composition control. This study established a composition-controlled, scaffold-aware benchmark for 2089 traditional Chinese medicine molecular records to compare four model tracks under identical held-out frameworks. The records, representing 2071 unique PubChem CIDs and 1986 distinct canonical SMILES, were audited for repeated structures, stereochemical information and ORCA target provenance. One Bemis-Murcko partition was shared by ordinary least-squares regression on explicit-hydrogen element counts (EC-LR), a five-layer graph convolutional network (GCN), a five-layer graph isomorphism network (GIN), and ordinary least-squares regression on binary Morgan fingerprints (FP-LR). Fold-level statistics were summarized as unweighted means and sample standard deviations, whereas pooled diagnostics used one out-of-fold prediction per record and model. Direct output comparison matched 2084 of 2089 values exactly to the ORCA Final Gibbs free energy field in Eh; all 2089 values were retained unchanged and analyzed identically. EC-LR achieved R² = 0.9998 ± 0.0003 and MAE = 0.189 ± 0.229 Eh, showing that absolute G is composition dominated. GCN was more accurate than GIN and FP-LR, but EC-LR remained superior. Out-of-fold atom feature occlusion and node deletion characterized GCN sensitivity without assigning causal thermodynamic contributions. The results support an audited control framework rather than a new architecture and do not establish that graph learning is necessary or universally superior. More discriminating studies should examine atomization, formation or reaction free energies, or predict residuals after removal of an explicit composition baseline.