RESULTS.md

Measured benchmark results

Mean ± sample standard deviation across split seeds. These are descriptive repeat variations, not confidence intervals or a standard MoleculeNet leaderboard. RMSE, MAE and interval width are in log10(mol/L). Absolute conformal intervals target 90% coverage. Scaffold shift can invalidate exchangeability.

Random split

Model	RMSE ↓	MAE ↓	R² ↑	Coverage	Width ↓	Train seconds
GNN · graph only	0.644 ± 0.064	0.484	0.896	88.9%	2.081	33.8
GNN + desc. + 3D	0.605 ± 0.084	0.443	0.908	91.1%	2.012	40.8
GNN + descriptors	0.593 ± 0.039	0.436	0.912	87.5%	1.967	41.1
GNN + desc. / no bond features	0.598 ± 0.033	0.442	0.910	88.7%	2.036	42.0
Training mean	2.003 ± 0.113	1.615	-0.003	92.3%	7.319	0.0
RF · descriptors	0.705 ± 0.064	0.516	0.876	88.9%	2.130	1.5
RF · Morgan	1.199 ± 0.050	0.895	0.636	91.3%	4.215	12.4
Ridge · descriptors	0.979 ± 0.028	0.775	0.760	92.7%	3.474	0.0

Scaffold split

Model	RMSE ↓	MAE ↓	R² ↑	Coverage	Width ↓	Train seconds
GNN · graph only	0.785 ± 0.059	0.596	0.881	91.4%	2.705	38.8
GNN + desc. + 3D	0.807 ± 0.091	0.591	0.873	90.0%	2.669	42.5
GNN + descriptors	0.764 ± 0.063	0.572	0.887	91.4%	2.619	42.5
GNN + desc. / no bond features	0.762 ± 0.100	0.565	0.887	91.4%	2.595	40.6
Training mean	2.915 ± 0.033	2.356	-0.657	82.4%	8.318	0.0
RF · descriptors	0.825 ± 0.024	0.602	0.867	93.4%	3.111	1.4
RF · Morgan	2.147 ± 0.042	1.735	0.105	80.2%	5.850	9.6
Ridge · descriptors	1.076 ± 0.062	0.835	0.775	91.0%	3.287	0.0

Ablations

ΔRMSE = variant minus GNN + descriptors; lower is better.

Split	Variant	Paired ΔRMSE
random	GNN · graph only	+0.051 ± 0.026
random	GNN + desc. / no bond features	+0.005 ± 0.010
random	GNN + desc. + 3D	+0.012 ± 0.045
scaffold	GNN · graph only	+0.021 ± 0.036
scaffold	GNN + desc. / no bond features	-0.001 ± 0.037
scaffold	GNN + desc. + 3D	+0.043 ± 0.029

See summary.csv for both absolute and spread-scaled intervals, including the interval score (width plus a penalty for missed targets).