BDE-db Bond Dissociation Enthalpy Prediction

What this research found

Bond-dissociation enthalpy — the energy needed to split a covalent bond into two radicals — controls radical chemistry, combustion and drug metabolism, but computing it by density functional theory is too slow for screening. A gradient-boosted tree ensemble was trained to predict the enthalpy of an individual bond from two-dimensional structure alone, using 290,664 reference values from the BDE-db dataset under a strict molecule-disjoint split. It reached a held-out mean absolute error of 1.865 kcal/mol, and stratifying the error showed accuracy tracks how often each bond type appears in training.

• The tuned model reached a test mean absolute error of 1.865 kcal/mol and a root-mean-square error of 3.650 kcal/mol on 57,767 bonds from 8515 unseen molecules — a 52.0% error reduction against a Ridge-regression baseline (3.882) and 28.8% against untuned gradient boosting (2.620).
• Errors are sharply peaked with a heavy tail: the median absolute error is 0.83 kcal/mol, and 55.2%, 73.5% and 82.3% of test bonds fall within 1, 2 and 3 kcal/mol of the reference calculation respectively.
• Accuracy tracks training abundance almost exactly. C–H bonds, 51.5% of the test set, were predicted to 1.373 kcal/mol; C–C bonds to 1.813; the aggregated heteroatom bonds (C–O, C–N, O–H, N–H) to 2.781; and the rare N–N, N–O and O–O bonds, just 1.2% of the data, to 3.933. At element-pair resolution N–N is worst at 4.501 kcal/mol.
• Predictions are essentially unbiased where data is plentiful, with mean signed residuals below 0.1 kcal/mol for C–H and C–C. The sparse N–O category is the exception, showing a +1.635 kcal/mol bias consistent with reverting toward the population mean when local training support is thin.
• Generalization holds up under scrutiny: the test error of 1.865 kcal/mol is slightly better than the molecule-grouped five-fold cross-validation error of 1.941 (fold standard deviation 0.030), and the training error of 1.63 sits only 0.24 kcal/mol below test error despite a 2063-tree ensemble on 1041 features.
• The pipeline does not match the roughly 0.6 kcal/mol reported for the specialist ALFABET graph neural network on the same data; the threefold larger error is the cost of using fixed interpretable descriptors and an off-the-shelf tabular learner.

How it was done

Reference enthalpies came from BDE-db, computed at the M06-2X/def2-TZVP level for 42,577 neutral molecules of carbon, hydrogen, oxygen and nitrogen with at most ten heavy atoms. Independent re-validation confirmed the element and size constraints and recovered exactly the published 290,664 unique dissociations, spanning roughly −1 to 165 kcal/mol. Each bond was encoded as a 1041-dimensional vector combining descriptors of its two terminal atoms, bond-level descriptors, and radius-2 Morgan fingerprints centred on each terminal atom, with the atoms canonically ordered so the encoding does not depend on which is listed first. Featurization had to materialize explicit hydrogens, since C–H, N–H and O–H bonds exist as explicit bonds only then. Molecules — not bonds — were split 80/20, keeping every bond of a molecule on one side and giving 232,897 training bonds against 57,767 test bonds with zero overlap. Ridge regression, default gradient boosting, and a model tuned by sequential search against molecule-grouped cross-validation error were then compared.

Data sources

• BDE-db (St. John et al., Nature Communications 11:2328, 2020) — 290,664 homolytic bond-dissociation enthalpies at M06-2X/def2-TZVP for 42,577 molecules
• RDKit for featurization, LightGBM for the tree ensembles, Optuna for hyperparameter search, scikit-learn for cross-validation