Answers are grounded in this paper and cite their sources.
As descriptor-based machine learning becomes widely used for reaction yield prediction and catalyst prescreening, controlled validation testbeds with known ground-truth data-generating processes (DGPs) are useful for auditing model behavior, cross-validation protocols, and post-hoc explainability methods. Here we construct and evaluate a synthetic chemoinformatics benchmark of N = 5,000 catalytic alcohol oxidation reactions across 20 independently seeded dataset realizations (100,000 total simulated reactions). By pairing a 22-substrate alcohol library characterized by RDKit exact monoisotopic masses (ExactMolWt) with categorical catalyst and terminal oxidant tiers, continuous thermal profiles, and non-causal synthetic decoy variables (Decoy_Cat_Load, Decoy_Rot_Bonds, Decoy_TPSA), we compare four regression models: Ordinary Least Squares (OLS) Linear Regression, eXtreme Gradient Boosting (XGBoost), Random Forest (RF), and Radial Basis Function Support Vector Regression (RBF-SVR). First, we compute the analytical unbounded Gaussian noise floor (3.989% MAE) alongside the 20-seed empirical clamped-oracle benchmark (3.994 ± 0.043% MAE, R² = 0.870 ± 0.003), showing that OLS closely approaches the oracle error floor (4.001 ± 0.044% MAE) under fold-specific standardization. Second, to illustrate how validation splits respond to substrate-level effects, we evaluate a positive-control regime that varies substrate-specific reactivity offsets (σ_sub ∈ {0.0, 4.0, 8.0, 12.0} percentage points), quantifying the divergence between random five-fold cross-validation and substrate-grouped hold-out (GroupKFold). Third, a paired four-model decoy ablation shows that while OLS, XGBoost, and RF exhibit minimal sensitivity to uninformative continuous descriptors (|ΔMAE| ≤ 0.008%, with RF changing by +0.008 ± 0.002%), untuned isotropic RBF-SVR experiences a larger error increase (+0.360 ± 0.026% MAE) consistent with Euclidean distance dilution. Fourth, multi-seed TreeSHAP evaluations show consistent global signal-to-decoy separation across all 20 realizations. Finally, we apply an illustrative multi-criteria Green Performance Index (GPI) to examine the trade-off between thermal activation and environmental penalty in post-hoc reaction ranking.
No takes yet. Share an insight, caveat, or question.
Gagan Sai Amith Pennada (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: