166
Z. Liu et al.
of ranked compound pairwise similarity list. Here, we used the 0.95 quantile
value as a cut-off to transform the ranked compound pairwise similarity list into
0 and 1.
(4) ROC curve analysis: finally, the PRank score between the two testing assay
platforms can be calculated based on ROC curve analysis. Consequently, the
area under the curve (AOC) value was considered as the PRank score to assess
the assay transferability quantitively.
This proposed PRank method aims to provide a novel framework to assess assay
transferability. Therefore, detailed strategies in each step could be modified and
updated. For example, alternative strategies to generate the compound gene signature
such as the fold change +p value criteria suggested by the MAQC consortium [39]
was also fit for PRank. Furthermore, significant genes could be domain-specific.
For instance, a lot of toxicogenomics annotation resources such as the Comparative
Toxicogenomics Database (CTD) could be employed to further limit the significant
gene list into different adverse related pathways (AOPs) or toxicity-related gene sets
[40]. In the current version, Dice’s correlation coefficient was used to calculate the
compound pairwise similarity. Other similarity measures such as Tanimoto similarity
or KL divergence based on topic similarity are also suggested for the future version.
Lastly, we used ROC curve analysis to calculate the PRank score to represent assay
transferability quantitatively. Other strategies for comparing two lists are also worth
testing.
In the conventional approach for querying assay transferability, the compound is
tested in the different assay testing systems. If the gene expression patterns of the
compound are similar in assay system A and B, we consider the compound could
reflect the same biology in both testing systems. However, this approach always
suffers from interior assay difference and batch effect in experimental design, which
fails to provide a global view on the assay transferability. In the PRank, “similar”
compounds have implied similar toxicity profiles in toxicogenomic space. A highly
similar pair of compounds should be observed in both testing TGx assays. In another
word, if the pairwise similarity of the two compounds is consistently ranked on the
top of all the compound pairs under the testing assays, these two compounds can
be a highly likely similar pair. The same concept can be extrapolated to assess the
transferability of any two testing systems where if two assays could produce the same
ranking resolution, we consider the two assays interchangeable.
9.4 Toxicogenomics Data and Annotation Resources
9.4.1 Open TG-GATEs
To explore transferability among TGx testing systems using the proposed PRank
method, a large scale of TGx data set from the Open TG-GATEs was employed
[31]. TG-GATEs is an acronym for Toxicogenomic Project—Genomics Assisted
Z. Liu et al.
of ranked compound pairwise similarity list. Here, we used the 0.95 quantile
value as a cut-off to transform the ranked compound pairwise similarity list into
0 and 1.
(4) ROC curve analysis: finally, the PRank score between the two testing assay
platforms can be calculated based on ROC curve analysis. Consequently, the
area under the curve (AOC) value was considered as the PRank score to assess
the assay transferability quantitively.
This proposed PRank method aims to provide a novel framework to assess assay
transferability. Therefore, detailed strategies in each step could be modified and
updated. For example, alternative strategies to generate the compound gene signature
such as the fold change +p value criteria suggested by the MAQC consortium [39]
was also fit for PRank. Furthermore, significant genes could be domain-specific.
For instance, a lot of toxicogenomics annotation resources such as the Comparative
Toxicogenomics Database (CTD) could be employed to further limit the significant
gene list into different adverse related pathways (AOPs) or toxicity-related gene sets
[40]. In the current version, Dice’s correlation coefficient was used to calculate the
compound pairwise similarity. Other similarity measures such as Tanimoto similarity
or KL divergence based on topic similarity are also suggested for the future version.
Lastly, we used ROC curve analysis to calculate the PRank score to represent assay
transferability quantitatively. Other strategies for comparing two lists are also worth
testing.
In the conventional approach for querying assay transferability, the compound is
tested in the different assay testing systems. If the gene expression patterns of the
compound are similar in assay system A and B, we consider the compound could
reflect the same biology in both testing systems. However, this approach always
suffers from interior assay difference and batch effect in experimental design, which
fails to provide a global view on the assay transferability. In the PRank, “similar”
compounds have implied similar toxicity profiles in toxicogenomic space. A highly
similar pair of compounds should be observed in both testing TGx assays. In another
word, if the pairwise similarity of the two compounds is consistently ranked on the
top of all the compound pairs under the testing assays, these two compounds can
be a highly likely similar pair. The same concept can be extrapolated to assess the
transferability of any two testing systems where if two assays could produce the same
ranking resolution, we consider the two assays interchangeable.
9.4 Toxicogenomics Data and Annotation Resources
9.4.1 Open TG-GATEs
To explore transferability among TGx testing systems using the proposed PRank
method, a large scale of TGx data set from the Open TG-GATEs was employed
[31]. TG-GATEs is an acronym for Toxicogenomic Project—Genomics Assisted