Abstract

Introduction: Current in-vitro screens for liver safety rely on cytotoxicity
endpoints. In contrast, organ failure in patients occurs due to functional
damage, much earlier than cell death. Liver injury can be caused by different
molecular initiation events in multiple connected pathways. We investigated how
measuring metabolites in cell lysates of in-vitro cultured liver cell lines
could improve our predictions for hepatotoxicity compared to current assays and
provide mechanistic insight.

Objectives: We demonstrate a proof-of-concept high-throughput metabolomic
cell-based assay to reveal mechanistic and functional readouts for key events
in adverse outcome pathways for drug-induced liver injury (DILI).

Methods: Galactose-cultured HepG2 C3a cells were treated with 150 DILI-relevant
compounds in 9-point, half-log spaced dose-response format with 316µM top
concentration. Metabolomic profiling was done in 384-well plates using a
high-throughput acoustic mist ionisation mass spectrometry (AMI-MS) platform
consisting of an Echo acoustic dispenser and a time-of-flight mass detector. Plates
treated in the same way were also analysed in a 9-parameter imaging-based assay
using Hoechst, TMRM, TOTO-3 and a classic cytotoxicity endpoint (ATP).


Results: Using the GeneData Expressionist software, we detected over 3000 peaks
and isotope clusters. Close to a thousand of the peaks had annotation in the
Human Metabolite Database. Out of those 1000 metabolites, 300 matched an
in-house expert-curated database of metabolites.


To analyse the full
multidimensional dataset in an untargeted way, we have developed a new data
analysis pipeline to process high-throughput MS data. The dataset contained
features spanning multiple orders of magnitude and a high proportion of
variable features with missing values. Moreover, experiments performed on
different days clustered separately in principal component analyses, indicating
batch effects. Our pipeline in R addresses these challenges by filtering out peaks
with many missing values, imputing missing values for the rest and correcting
for batch effects. The pipeline also includes different normalisation
techniques with and without internal standard correction.   


Many of the
metabolites exhibited non-monotonic and hormetic dose-responses with initial
stimulation at low and intermediate concentrations, followed by a decrease at
higher toxic concentrations. We developed an automatic curve fitting workflow
in R which selects the best fit between constant, linear, hormetic, logistic,
and multiphasic models using Bayesian information criterion (BIC) to penalize
more complex models. We implemented this on multiple experiments with more than
60 000 concentration-responses per experiment.

This facilitates feature
selection and concentration-response quantification in an unbiased way for
multiparametric data. It also reduces analysis time from months to hours and
enables the development of a predictive model of compound toxicity.