Method Profile
Human liver-cell toxicity, mitochondrial respiration, bile-salt transport and drug properties are combined in interpretable flags and separately trained classifiers.
Prioritize small molecules for mechanistic follow-up during early drug discovery and lead optimization.
170 reference drugs: 80 clinically DILI-positive and 90 DILI-negative; 2D primary human hepatocytes, 3D human liver microtissues, HepG2 respiration and BSEP vesicles.
48-hour 2D viability, seven-day 3D repeated exposure, short mitochondrial-respiration testing and BSEP inhibition; exposure-inclusive and exposure-free random forests.
Human liver systems
BioIVT primary human hepatocytes and InSphero microtissues containing pooled hepatocytes from 10 donors plus non-parenchymal cells; ATP viability measured against vehicle.
Mechanistic assays
HepG2 oxygen-consumption testing and BSEP taurocholate transport in vesicles; additional metabolism, physicochemical and glutathione-trapping features.
Reference data
Literature/curated clinical DILI classification and total clinical Cmax, or Css where Cmax was unavailable.
Treat 2D/3D cytotoxicity as one hazard category alongside mitochondrial and BSEP categories. Evaluate random forests with 100 repetitions of stratified tenfold cross-validation and explicitly retain each probability threshold.
Compare hazard combinations and two classifier feature sets against clinical labels within the 170-drug reference set.
80 DILI-positive and 90 DILI-negative reference drugs; each molecule appears in 100 held-out test folds across repeated cross-validation.
Clinical reference labels; exposure-inclusive versus exposure-free models; single versus multiple hazard categories.
At least two hazard categories gave 98% specificity and 40% sensitivity. At probability 0.50, exposure-inclusive RF balanced accuracy/sensitivity/specificity were 0.72/0.71/0.72, versus 0.60/0.55/0.66 without exposure.
Apply the assay and model framework to internal drug-development examples with differing evidence availability.
Eight Genentech compounds, four with known clinical DILI; other examples have limited clinical or preclinical information.
Mechanistic flags, RF probabilities and the available development outcomes for each compound.
The exposure-free model classified all eight positive; the exposure-inclusive model classified five positive and three negative at 0.50, including a known clinical DILI compound with probability 0.47.
A Genentech-led study combines mechanistic assays and drug properties across 170 reference drugs. Two hazard categories gave 98% specificity but 40% sensitivity; random-forest results depended on exposure information and the decision threshold.
