Method Profile
A molecular-descriptor classifier evaluated on an independently sourced set, with performance reported before and after confidence filtering.
Early-stage ranking of candidate compounds for possible drug-associated kidney toxicity while identifying predictions that the model treats as uncertain.
Random-forest classifier using RDKit two-dimensional descriptors; tree-to-tree variation defines four confidence tiers.
Ten internal data partitions for model development and a separate DIRIL external set; compare full-set performance with a retained high/medium-confidence subset.
Source labels
Three curated datasets with structure deduplication and label reconciliation; 1,829 development compounds and 247 retained external compounds.
Inputs and outputs
Two-dimensional molecular descriptors, binary nephrotoxicity classification, probability and confidence tier; no injury-subtype prediction.
Report external AUC and coverage together. SHAP feature attribution is a statistical explanation of predictions, not direct evidence of a renal-injury mechanism.
Source-separated external comparison followed by confidence-filtered analysis.
1,829 development compounds; 247 external compounds; about 54 external compounds retained at high or medium confidence.
DIRIL external nephrotoxicity labels; full external set versus the selected confidence subset.
Full external AUC 0.727; retained-subset AUC 0.902 at 21.7% coverage. Low/very-low tiers contained 88.3% of false negatives and 94.5% of false positives.
An academic kidney-toxicity model reached an AUC of 0.727 across 247 external compounds. Restricting the analysis to high- or medium-confidence predictions raised AUC to 0.902, but retained only 21.7% of that set.
