4 Matrix and Tensor Factorization Methods …
69
are structurally analogous drugs. The drugs demonstrated an HSP response of the
cells as through up-regulation of key HSP genes. This pan-cancer response across all
three cancers is linked to the toxicity outcomes of the drugs. HSP90 is a molecular
chaperone protein that is essential for stabilization of a variety of other proteins [49],
and HSP90 inhibitors bind to the protein, resulting in its loss of function. HSP90
inhibitors have been evaluated for their therapeutic efficacy in multiple cancers [50,
51]. This component, therefore, presents a well-known HSP90 response of cancer
cells. For details on this and other components, see [41].
4.3.4 Predictive Toxicogenomic Space (PTGS)
Predictive toxicogenomics space (PTGS [4]) is a recent “big data compacting and
data fusion” methodology to model various adverse and toxic outcomes on cellular
and organism levels. A machine learning-based data summarization approach was
applied on a large transcriptomics data set. This methodology formed a predictive tool
termed PTGS that used as features over 1000 genes distributed over 14 overlapping
cytotoxicity-related gene space components, as described in [4]. Specifically, a LDA
matrix factorization-based method was applied to the gene profiles from the Connectivity Map data set, and the resulting summarized components were fused with
cytotoxicity data from the NCI-60 cancer cell line screens to generate the PTGS.
The PTGS tool was validated for predicting drug-induced liver injury (DILI) and
liver cytopathological changes by calculating PTGS component scores within three
liver-related subsets of the independent TG-GATEs database [52], being the largest
public toxicogenomics database. It was shown to successfully capture all the studied
liver pathological changes in rats, and in conjunction with human therapeutic drug
exposure levels (C max ), was able to facilitate the use of cell culture-derived toxicogenomics experiments with human and rat hepatocytes to predict DILI with greater
accuracy than other in vitro methods [4].
4.4 Discussion
Recent advances in machine learning methodologies have made it possible to perform
integrated analysis of the gene expression response data and toxicity profiles directly.
Such detailed analysis offers deeper insights by linking the activity patterns of the
genes directly with the toxicity responses, and hence enriching the factor components
with detailed interactions. As molecular responses of cancer cells are known to
depend on a multitude of factors, including drug MoA, cell type, and cellular states,
simultaneous modeling of these various factors is beginning to attract attention.
Specifically, in cancer, cells are known to be heterogeneous and respond selectively
to targeted drugs, making it valuable to systematically model the various factors and
segregate responses specific to a particular cancer-type from those which are generic.
69
are structurally analogous drugs. The drugs demonstrated an HSP response of the
cells as through up-regulation of key HSP genes. This pan-cancer response across all
three cancers is linked to the toxicity outcomes of the drugs. HSP90 is a molecular
chaperone protein that is essential for stabilization of a variety of other proteins [49],
and HSP90 inhibitors bind to the protein, resulting in its loss of function. HSP90
inhibitors have been evaluated for their therapeutic efficacy in multiple cancers [50,
51]. This component, therefore, presents a well-known HSP90 response of cancer
cells. For details on this and other components, see [41].
4.3.4 Predictive Toxicogenomic Space (PTGS)
Predictive toxicogenomics space (PTGS [4]) is a recent “big data compacting and
data fusion” methodology to model various adverse and toxic outcomes on cellular
and organism levels. A machine learning-based data summarization approach was
applied on a large transcriptomics data set. This methodology formed a predictive tool
termed PTGS that used as features over 1000 genes distributed over 14 overlapping
cytotoxicity-related gene space components, as described in [4]. Specifically, a LDA
matrix factorization-based method was applied to the gene profiles from the Connectivity Map data set, and the resulting summarized components were fused with
cytotoxicity data from the NCI-60 cancer cell line screens to generate the PTGS.
The PTGS tool was validated for predicting drug-induced liver injury (DILI) and
liver cytopathological changes by calculating PTGS component scores within three
liver-related subsets of the independent TG-GATEs database [52], being the largest
public toxicogenomics database. It was shown to successfully capture all the studied
liver pathological changes in rats, and in conjunction with human therapeutic drug
exposure levels (C max ), was able to facilitate the use of cell culture-derived toxicogenomics experiments with human and rat hepatocytes to predict DILI with greater
accuracy than other in vitro methods [4].
4.4 Discussion
Recent advances in machine learning methodologies have made it possible to perform
integrated analysis of the gene expression response data and toxicity profiles directly.
Such detailed analysis offers deeper insights by linking the activity patterns of the
genes directly with the toxicity responses, and hence enriching the factor components
with detailed interactions. As molecular responses of cancer cells are known to
depend on a multitude of factors, including drug MoA, cell type, and cellular states,
simultaneous modeling of these various factors is beginning to attract attention.
Specifically, in cancer, cells are known to be heterogeneous and respond selectively
to targeted drugs, making it valuable to systematically model the various factors and
segregate responses specific to a particular cancer-type from those which are generic.