Applicability-Domain-Aware Evaluation of Machine-Learning NMR Chemical-Shift Prediction in Small drug Molecules: GT-NMR for 13C and CASCADE for 1H
Year: 2026
Type: preprint
Abstract: A small external evaluation was performed to examine the practical accuracy and applicability domain of two pretrained machine-learning models for solution-state NMR chemical-shift prediction in five pharmaceutically relevant small molecules: amylmetacresol, benzocaine, dopamine, betazole, and betahistine.The pretrained GT-NMR graphtransformer model was used for atom-specific 13C chemical-shift prediction, while the CASCADE 3D graph neural-network workflow was used for 1H chemical-shift prediction.Predictions were compared with primary-literature NMR data where numerical assignments could be verified.Exact dataset overlap, nearest-training-set Morgan-fingerprint Tanimoto similarity, Murcko-scaffold coverage, and molecular complexity were evaluated as applicability-domain diagnostics.None of the five exact structures was detected in the checked GT-NMR or CASCADE proton train/validation/test partitions, although all five Murcko scaffolds were represented in the corresponding training sets.For GT-NMR, audited 13C molecule-level MAEs were 0.50 ppm for benzocaine, 2.59 ppm for
Source: ChemRxiv
Author: Diba Allameh Zadeh
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7211994047