308
D. Wang
median or mean BMD of genes in the pathway to represent the pathway-wise BMD,
other quantiles or summary statistics can also be used.
Determining the transcriptional PODs. After pathway-wise BMDs have been
calculated, one needs to summarize them into a single value for the POD. The first
question is whether all pathways in the consideration (use the most sensitive pathway
or some other quantiles) or only pathways relevant to an AOP should be used. If there
is solid knowledge linking the chemical under consideration with one or more AOPs,
considering pathways or gene sets associated with the relevant AOPs will reduce noise
and anchor the transcriptional PODs with specific biological mechanisms, which
may lead to more accurate estimate. However, for a large number of chemicals, the
existing knowledge is not sufficient to clearly identify relevant AOPs; though it might
be possible to use transcriptomic data to construct biologically relevant pathways and
potential AOPs (see [30–32] and others for more discussion) while estimating PODs.
The alternative is to use an AOP-independent approach (e.g., relying on the most
sensitive pathway), which can give valuable information on PODs without detailed
knowledge of AOPs. This is especially true if the toxicity effects of the chemical are
broad and disturbing many different pathways.
Early efforts using the AOP-independent approach have focused on the most sensitive pathways, i.e., using the lowest BMD or BMDL from all pathways (after filtering)
as the transcriptomic POD. Though this approach has been shown to be successful
with a number of studies, other alternatives can also be considered. Farmahin et al.
[20] compared 11 different approaches for deriving transcriptomic PODs with different criteria for gene and pathway selection. They concluded that three approaches
have the best performance: (1) using the 20 significantly enriched pathways with the
lowest BMDs, (2) using the 20 genes with the largest fold changes relative to controls, and (3) using the 20 genes that contribute to the greatest number of enriched
pathways. Reassuringly, other approaches, including the most sensitive pathway
approach, usually also give reasonable results. From the author’s experience, the
most sensitive pathway approach is sometimes unstable when the number of time
points and sample size is small, and the tenth percentile of all pathways can give
good results. Thus, trying multiple approaches in determining the POD should be
beneficial.
15.3 PODs Based on In Vitro Assays
Discussions so far have focused on transcriptional profiling with the same tissues used
to derive apical endpoints. This provides the most direct comparison of transcriptional
PODs with the apical endpoint based counterpart. On the other hand, a significant
amount of data has become available using in vitro assays. The ToxCast project
(including Tox21 assays) generated data for a diverse array of chemicals using a dose
response format that usually involves a range of doses in triplets. ToxCast provides
a specialized software package (the R package tcpl, [33]) to fit dose response curves
with constant, Hill, or gain-loss models. It is therefore feasible to perform BMD-type
D. Wang
median or mean BMD of genes in the pathway to represent the pathway-wise BMD,
other quantiles or summary statistics can also be used.
Determining the transcriptional PODs. After pathway-wise BMDs have been
calculated, one needs to summarize them into a single value for the POD. The first
question is whether all pathways in the consideration (use the most sensitive pathway
or some other quantiles) or only pathways relevant to an AOP should be used. If there
is solid knowledge linking the chemical under consideration with one or more AOPs,
considering pathways or gene sets associated with the relevant AOPs will reduce noise
and anchor the transcriptional PODs with specific biological mechanisms, which
may lead to more accurate estimate. However, for a large number of chemicals, the
existing knowledge is not sufficient to clearly identify relevant AOPs; though it might
be possible to use transcriptomic data to construct biologically relevant pathways and
potential AOPs (see [30–32] and others for more discussion) while estimating PODs.
The alternative is to use an AOP-independent approach (e.g., relying on the most
sensitive pathway), which can give valuable information on PODs without detailed
knowledge of AOPs. This is especially true if the toxicity effects of the chemical are
broad and disturbing many different pathways.
Early efforts using the AOP-independent approach have focused on the most sensitive pathways, i.e., using the lowest BMD or BMDL from all pathways (after filtering)
as the transcriptomic POD. Though this approach has been shown to be successful
with a number of studies, other alternatives can also be considered. Farmahin et al.
[20] compared 11 different approaches for deriving transcriptomic PODs with different criteria for gene and pathway selection. They concluded that three approaches
have the best performance: (1) using the 20 significantly enriched pathways with the
lowest BMDs, (2) using the 20 genes with the largest fold changes relative to controls, and (3) using the 20 genes that contribute to the greatest number of enriched
pathways. Reassuringly, other approaches, including the most sensitive pathway
approach, usually also give reasonable results. From the author’s experience, the
most sensitive pathway approach is sometimes unstable when the number of time
points and sample size is small, and the tenth percentile of all pathways can give
good results. Thus, trying multiple approaches in determining the POD should be
beneficial.
15.3 PODs Based on In Vitro Assays
Discussions so far have focused on transcriptional profiling with the same tissues used
to derive apical endpoints. This provides the most direct comparison of transcriptional
PODs with the apical endpoint based counterpart. On the other hand, a significant
amount of data has become available using in vitro assays. The ToxCast project
(including Tox21 assays) generated data for a diverse array of chemicals using a dose
response format that usually involves a range of doses in triplets. ToxCast provides
a specialized software package (the R package tcpl, [33]) to fit dose response curves
with constant, Hill, or gain-loss models. It is therefore feasible to perform BMD-type