Reliable Yield Modeling for 2-Arylbenzimidazole Synthesis: Source Bias, Endpoint Definition, and Validation Boundaries

Introduction: Literature-derived reaction-yield models often report strong predictive performance, yet random data splitting can obscure publication-specific experimental designs and inconsistent yield definitions. We investigated this problem using the synthesis of 2-arylbenzimidazoles as a chemically focused case study.

Methods: A targeted archive of 185 reaction records was assembled from five publications covering enzyme, supported-gold, supported-base, inorganic-salt, rare-earth, and catalyst-free conditions. Chemical descriptors were rebuilt from a manually audited substrate registry, while isolated, crude NMR, unspecified, and nonnumerical outcomes were separated. Models were evaluated using repeated random five-fold, substrate-grouped, and leave-one-publication-out validation. Random-forest and ridge models were compared with matched mean baselines, and catalyst/solvent support overlap was examined for each held-out publication.

Results: Among 120 exact isolated-yield records, the coarse random-forest model achieved a mean absolute error of 11.1 percentage points under repeated random validation but 15.0 points under leave-one-publication-out validation. At the publication boundary, its performance was essentially equivalent to a global mean baseline of 14.7 points. A structural ridge model performed worse, with an error of 20.1 points. The loss of transferability was associated with minimal or absent overlap in catalyst and solvent regimes across publications. Pooling heterogeneous endpoints further increased the prediction error.

Conclusions: For literature-based yield modeling, apparent accuracy can reflect interpolation within publication-specific condition spaces rather than chemically transferable learning. Reliable reaction informatics requires audited descriptors, explicit endpoint definitions, validation across publication boundaries, and datasets designed to share catalyst and solvent support across studies.