Fig 1.
Quantifying time concordance based on first activation.
(A) The event activation of the events A-D and the later event is shown over time, as well as their timepoint of first activation, at which the event first passes the defined activation criteria. If an event takes place before a defined later event, which in our study is adverse histopathology, it is time-concordant. Time concordance indicates that there is potentially a causal relation between both events, and is distinct from time-dependence which is defined based on the correlation to the later event or time. (B) Based on the frequency of an event before or at the same time as the later event and its frequency in background time-series without the later event, a confusion matrix and different time concordance metrics can be derived. (C) Time concordance can both prioritize novel links and provide further evidence on potential mechanistic links between events. Events are indicated as nodes and mechanistic links between them as edges.
Fig 2.
Workflow to quantify time concordance between preceding gene expression-derived events and later adverse histopathology.
First, events are derived from the gene expression and histopathology data. Pathway and TF activity is inferred based on the expression of the respective gene sets using GSVA [26] and binary histopathology labels are derived from the continuous toxscores. Secondly, we derive the first activation of expression-based events as well as of adverse histopathology. Lastly, we quantify the time concordance between potential preceding events (PE) which are derived from gene expression and adverse histopathology as potential later event (LE).
Table 1.
Metrics quantifying the time concordance between a potential preceding event PE and potential later event LE, and their relation to the original Bradford Hill (BH) considerations.
Fig 3.
Distribution and relation of histopathological findings across time series.
A) Histopathology labels are defined for each histopathological finding at 3 different toxscore cut-offs, namely “null” (toxscore>0), “low” (toxscore>0.67) and “high” (toxscore>1.34). For each label, the number of occurrences in the 40 adverse time series and the fraction of adverse time series among all occurrences of the given histopathology label are shown. Histopathological findings, out of which at least 50% and at least 5 of the occurrences were found in adverse conditions timeseries were considered adverse B) Number of conditions with histopathological findings at different timepoints, as well as the frequency of the respective first activations C) Time of first activation across timeseries labelled as adverse or non-adverse. Each time series is annotated with the dose level in repeat-dose studies, as well as with whether or not the time series was considered adverse by Sutherland et al. [15].
Fig 4.
Enrichment of known events in DILI before adverse histopathology based on gene sets as well as individual gene members.
The enrichment of first activation before or at adverse histopathology is shown for gene sets mapping to known key events in DILI, for which first activation was defined as first timepoint of differential GSVA-derived gene set activity. Furthermore, also enrichment of individual genes within these genesets is shown and was derived based on the first timepoints of differential expression. Aligning with the expected direction, a significant down-regulation of Liver X Receptor (LXR) signalling and bile acid-related pathways is observed, while all other gene sets were found to be more significantly up-regulated. Only for peroxisomal pathways, both directions were significantly enriched indicating that dysregulation in direction might be linked to adverse histopathology.
Fig 5.
|logFC| before adverse histopathology. For known processes in DILI which correspond to significantly enriched events before adverse histopathology, the max |logFC| before adverse histopathology is shown. In comparison to other known pathways and the overall background distribution, a high logFC is found for mitochondrial beta oxidation followed by peroxisomal beta oxidation and mitophagy.
Fig 6.
True positive rate (TPR) and positive predictive value (PPV) before or at histopathology of genes and gene sets in known key events in DILI.
Events related to the given known key event are shown in red or blue indicating an up- or downregulation, respectively. Genes with a p-value < 0.0001 involved in known key events in DILI are additionally labelled. The background distribution of all significantly enriched genes or gene sets is shown in grey (p-value < 0.05).
Fig 7.
Temporal concordance of nuclear receptors and adaptive response transcription factors (TFs) in DILI.
For known TFs in DILI the following time concordance metrics are shown: A) The enrichment significance before or at first adverse histopathology, B) Positive Predictive Value (PPV) and True Positive Rate (TPR), C) Max. mean |logFC| before or at first adverse histopathology. As background distribution in grey, the statistics for all inferred TFs is shown.
Fig 8.
Highest ranking events by time concordance metrics.
The ten transcription factor (TF)- and pathway-level events ranking highest by enrichment p-value, median max. |logFC| and true positive rate (TPR) before or at histopathology are shown.
Fig 9.
Transcription Factor (TF) activity and expression before adverse histopathology.
A) Significance of enrichment in adverse conditions for matched TF activity and expression-based events. Events only found on the expression or TF level are not included in the figure due to the inability to perform a statistical test for those. B) For significantly enriched TF activity-based events, the True Positive Rate (TPR) of observing TF activity before or at the time of adverse histopathology is shown, as well as the TPR for observing TF expression changes before TF activity in the time series where it precedes adverse histopathology.
Fig 10.
Causal relationships between TFs supported by time concordance.
For TFs which are significantly enriched before or at adverse histopathology, known causal relations are shown in which the upstream event is found before or at the downstream event in at least 20% of adverse cases. For induced TFs for which expression is found before regulon activity and significantly enriched, not only protein-protein interactions are considered but also upstream TF-target gene interactions annotated with DoRothEA [69] confidence scores (A: High confidence, C: Medium confidence).
Fig 11.
Combining time dependence and concordance to identify mechanistically supported biomarkers.
A) The relation between time concordance, quantified by the enrichment p-value for event activation before adverse histopathology, and time dependence, quantified by the meta p-value for Spearman correlation between time and event activation across adverse conditions, is shown. B) For events with the most significant time-dependence, the distribution of correlation coefficients is shown providing further insight into the strength of correlation and consistency across adverse conditions.