The primary comparison is fixed before evaluation: GNN + descriptors versus classical baselines, on the same partitions. All model variants are reported. These are descriptive results from three seeds, not claims of statistical significance or a standard MoleculeNet leaderboard score.
GNN + descriptors averages 0.593 ± 0.039 logS RMSE on random splits, compared with 0.705 ± 0.064 for the descriptor random forest. On scaffold splits, the GNN averages 0.764 ± 0.063, with per-seed scores of approximately 0.732, 0.837 and 0.723. The descriptor forest scores approximately 0.817, 0.806 and 0.853 on those same partitions: the GNN wins two of three, not every run.
That supports a modest within-study conclusion: the combined representation is competitive and improves the mean error over this baseline under the chosen training budget. It does not establish that a GNN will always beat a tuned classical model, nor that the difference will survive an external dataset.
On the random splits, removing global descriptors increases mean RMSE by 0.051. Masking bond features changes it by only +0.005, while adding single-conformer shape features changes it by +0.012. Positive differences mean worse error than GNN + descriptors.
The experiment therefore supports using the descriptors but provides no clear random-split improvement from the bond-feature channels or the single-conformer additions. Adjacency and atom chemistry remain available in the bond-masked model. The geometry variant also has more input parameters, approximation error and missing values. It would be incorrect to label all three feature choices as demonstrated improvements.
The complete paired results, including scaffold ablations and sample SD, are in ablations.csv. All three ablations use the same partitions, optimization settings and corresponding member seeds.
For GNN + descriptors, nominal 90% absolute conformal intervals achieve 87.5% average random-split coverage, with mean width 1.967 logS. Per-seed coverage ranges from 78.6% to 92.9%. Those differences are visible rather than averaged away in the per-run reports.
On scaffold splits, average coverage is 91.4% and mean width is 2.619 logS. The wider intervals accompany a different calibration distribution and larger errors. The favorable aggregate scaffold coverage does not establish a general guarantee under distribution shift.
Absolute and spread-scaled intervals are both retained in the results. Read coverage, width and interval score together. Ensemble standard deviation alone is not calibrated uncertainty, and reporting a nominal level alone is insufficient.
The diagnostic example is fixed to GNN + descriptors, scaffold split, seed 17. Its largest three errors are:
| Molecule ID | Measured logS | Predicted logS | Absolute error | Nearest training Tanimoto |
|---|---|---|---|---|
| esol-0998 | -4.445 | -7.786 | 3.341 | 0.220 |
| esol-0956 | -3.931 | -6.413 | 2.482 | 0.175 |
| esol-0999 | -2.840 | -5.042 | 2.202 | 0.262 |
These three cases have low similarity to their nearest training fingerprints. That is a useful diagnostic observation, not proof that similarity causes the error or that high similarity guarantees accuracy. A 3.341-log-unit error is very large on a concentration scale, even when overall RMSE looks respectable.
The five largest error cases include full structures, interval bounds and similarity values. Unknown assay conditions, solid forms, ionization and measurement provenance limit causal explanations. The appropriate next step is to inspect external measurements and relevant chemical context, rather than invent a mechanistic explanation from SMILES alone.
The dataset is small enough to reproduce locally without GPU hardware. Its 1,117 unique structures have 269 scaffold groups, including 316 acyclic molecules in one group. The splitting rule keeps this group together and primarily assigns large groups to training; this changes the chemical distribution of the held-out sets.
Twenty-seven conformers failed or did not converge; those molecules were retained with flagged, train-imputed geometric features. A larger conformer ensemble or explicit solvation treatment could be scientifically interesting, but this study does not demonstrate that those additions would improve performance.
The most valuable next experiment is an externally held-out solubility dataset with a molecule/scaffold overlap audit and documented assay conditions. Continue model selection on validation data; avoid treating repeated iterations on these published test partitions as new unbiased evidence.