Physics-informed conformal prediction of aromatic C-H site selectivity: guaranteed coverage prediction sets for...

Predicting the site of electrophilic aromatic substitution (EAS) is a longstanding challenge in synthetic and medicinal chemistry, yet existing quantum-chemical and machine-learning predictors return a single most likely site or an uncalibrated ranking, with no statistical guarantee that the experimentally observed position is captured. Here, we reformulate aromatic C-H site selectivity as a set-valued prediction problem and solve it with conformal prediction, a distribution-free framework that converts any predictive score into a prediction set carrying a user-specified, finite-sample coverage guarantee. Because this guarantee holds irrespective of the underlying model, gains in descriptor quality appear directly as smaller guaranteed sets at fixed coverage, yielding a quantitative measure of descriptor informativeness. On 535 experimentally characterized EAS substrates, we evaluate progressively richer atom-level representations combining topological and Gasteiger descriptors, GFN1-xTB atomic charges, and condensed Fukui f^- indices from a single additional semiempirical calculation. At 90% target coverage, the electronic descriptors increase the top-1 accuracy from 0.749 to 0.862 while shrinking the mean guaranteed set from 2.80 to 2.62 candidate sites without loss of coverage, and class-conditional calibration preserves reliability across substitution patterns. Descriptor attribution identifies the Fukui indices as the most informative descriptors, which also reproduce established regioselectivity trends across diverse aromatic systems, and a conformal credibility score flags out-of-domain substrates. Conformal prediction thus delivers EAS site-selectivity models that are accurate, efficient, interpretable, and statistically reliable.