On MoleculeNet BBBP (n=2039), Dynamic Random Forest with combined features reached the highest mean AUC of 0.970, while after orthogonalization no conditional LogP–BBB permeability effects remained significant

Predictive performance depends jointly on feature representation and algorithm rather than on either alone; Dynamic Random Forest with combined features achieved the highest mean AUC of 0.970 (95% CI: 0.963-0.977). Prior work often treats feature engineering and model selection separately; this study examines them as coupled design choices by systematically ablating three feature families and comparing four algorithms on the same dataset. A systematic ablation on the MoleculeNet BBBP dataset (n = 2039) reporting mean AUC with a 95% confidence interval, constituting a controlled comparison within a single dataset.

After constructing a pseudo-treatment from a LogP median split and applying double/debiased machine learning to account for confounding, orthogonalization substantially attenuated the heterogeneity detected by naive causal forests, and no conditional effects remained significant after false discovery rate correction (smallest adjusted p = 0.082). The result re-examines causal-forest heterogeneity estimates under an orthogonalization framework, indicating that unorthogonalized causal forests risk overstating genuine treatment effect heterogeneity. An exploratory analysis using double/debiased machine learning for orthogonalization with false discovery rate correction; the authors frame this as insufficient evidence rather than confirmed absence of effects.

After orthogonalization, feature importance shifted toward residual structural information in SMILES bigrams. This suggests that once observed confounding is controlled, information associated with LogP–BBB associations comes more from structural residuals than from LogP itself, offering a new empirical clue for feature representation choices. A descriptive comparison of orthogonalized feature importance, an exploratory finding without an independent validation cohort.

The work targets blood-brain barrier permeability prediction in central nervous system drug discovery and applies to settings where feature representation and algorithm are jointly selected on labeled molecular datasets such as MoleculeNet BBBP, as well as to researchers wishing to use generalized random forests to explore heterogeneous associations between molecular structure and properties. Its pipeline design—first ablating feature spaces, then performing orthogonalized heterogeneity estimation with double/debiased machine learning—can serve as a reference template for other molecular property prediction tasks.

The heterogeneity analysis is exploratory, and no conditional effects remained significant after false discovery rate correction (smallest adjusted p = 0.082), so the conclusion should be read as insufficient evidence after controlling observed confounding rather than proof that heterogeneity is absent. The pseudo-treatment is constructed from a LogP median split, and its interpretation depends on the reasonableness of that split. The finding that orthogonalized feature importance shifts toward residual structural information in SMILES bigrams has not been validated on independent data. In addition, the loaded text is abstract-level and lacks figures and full methodological detail, so implementation specifics and robustness checks still require consulting the original article.