Graph Neural Network and Transformer Fusion for Molecular Property Prediction under a Strict Evaluation Protocol: Notation Leakage, Descriptor Baselines and Conformal Coverage

Abstract

An evaluation protocol for multi-view molecular property prediction, and a fusion architecture measured under it. Eight MoleculeNet datasets, six scaffold splits (the DeepChem canonical split plus five seeded), paired comparisons with Holm correction, an across-dataset test and an equivalence test, split-conformal uncertainty (LAC, APS, RAPS, class-conditional, CQR), and machine-checked reproducibility gates that fail the build when a number, a config or a featuriser drifts. In ClinTox and BBBP the notation of a SMILES string (aromatic or Kekulé) predicts the label (notation-bit AUC 1.000 for ClinTox toxicity); canonicalising the strings lowers a frozen ChemBERTa model from 0.988 to 0.795 AUC on ClinTox. With the leak removed, no fusion variant outperforms a fingerprint-plus-descriptor MLP, the proposed fusion model's advantage over a two-layer graph network is not significant across datasets (favoured on 7 of 8, p ≥ 0.070), and a 16,513-parameter gate is not shown to be equivalent to the 1,169,793-parameter fusion block. Under marginal split conformal prediction on Tox21, three of fifteen models cover only 10-14% of active compounds; retraining the descriptor MLP without class weighting reproduces the failure (77.9% to 16.3%) at unchanged AUC. A change of accelerator moves single-split results about as much as a change of seed. The baseline pipeline is the author's own earlier work, preserved runnable so that claims about fixing its evaluation are checkable against the archived per-split metrics.