Graph Neural Network and Transformer Fusion for Molecular Property Prediction: What a Strict Evaluation Protocol Finds

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 and an across-dataset 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. The results are largely negative and are reported as such: the proposed fusion architecture does not beat a fingerprint-plus-descriptor MLP (4 of 8 datasets, p = 0.55); a 16,513-parameter gate matches a 1,169,793-parameter fusion block; roughly eleven GPU-hours of LoRA adaptation buy nothing over a cached frozen embedding; two published external baselines (AttentiveFP, Chemprop) are statistically indistinguishable from a two-layer graph network; a nominal 90% conformal guarantee covers 9.7-14.1% of active compounds for three of fifteen models, with mean prediction-set size predicting which; a fully seeded pipeline moved to a different GPU is a different model, moving 18% of single-split numbers by more than the field's reported effect sizes; and the significance of the paper's own mechanism claim is not stable to an architectural constant that does not otherwise matter. The baseline pipeline is the author's own earlier work, preserved runnable so that claims about fixing its evaluation are checkable against archived predictions.