Method Profile
Structure, target-interaction and pathway fingerprints were compared on the same drug set; a majority-vote model was tested internally, on dissimilar compounds and on an independent data set.
Exploratory classification of drugs for steatosis-related safety signals, with attention to whether the model generalizes beyond its development set.
Separate random-forest and gradient-boosted classifiers used PubChem structural fingerprints, liver-relevant human target interactions or mapped pathways. A consensus classifier combined selected single-fingerprint models by majority vote.
A common 372-drug set (183 steatosis-positive, 189 negative) was split into 335 for training and 37 low-similarity holdouts. A separate nonoverlapping published set contained 19 pharmaceuticals (10 positive, nine negative).
Outcome labels
Steatosis-associated reports curated from FAERS for the positive class; DILIst compounds classified as drug-induced-liver-injury negative used as a proxy steatosis-negative class.
Feature sources
PubChem structure bits plus drug–human-target interaction data filtered for liver expression and mapped to pathways. The same common compound set was used to compare input types.
Compare the three fingerprint representations and their majority-vote combination under repeated cross-validation; then examine the 37 least-similar compounds and an independent data set after removing overlapping drugs.
Five repetitions of two-fold cross-validation on the development set.
335 training compounds from the common 372-drug set.
Structure-only versus target- and pathway-based models; consensus majority vote.
Biological fingerprints outperformed structure alone in the internal comparison. The consensus balanced accuracy was 0.72 and sensitivity was 0.69.
Low-similarity holdout followed by a separate published steatosis data set after overlap removal.
37 low-similarity holdouts; 19 independent pharmaceuticals, including 10 positive and nine negative.
Curated steatosis labels in each test set.
Sensitivity, precision, Matthews correlation coefficient and balanced accuracy declined on the low-similarity holdout while specificity remained comparatively stable. Both individual models and the consensus performed poorly on the independent 19-drug set.
A University of Vienna team compared chemical-structure, target and pathway fingerprints for drug-induced liver steatosis. Biological inputs performed better during internal validation, while accuracy fell on less-similar compounds and a separate 19-drug set.
