Limits...
IDEA: Integrated Drug Expression Analysis-Integration of Gene Expression and Clinical Data for the Identification of Therapeutic Candidates.

Ung MH, Varn FS, Cheng C - CPT Pharmacometrics Syst Pharmacol (2015)

Bottom Line: Cancer drug discovery is an involved process spanning efforts from several fields of study and typically requires years of research and development.However, the advent of high-throughput genomic technologies has allowed for the use of in silico, genomics-based methods to screen drug libraries and accelerate drug discovery.We were able to predict drug candidates that have been known and those that have not been known in the literature to exhibit anticancer effects.

View Article: PubMed Central - PubMed

Affiliation: Department of Genetics, Geisel School of Medicine at Dartmouth Hanover, New Hampshire, USA.

ABSTRACT
Cancer drug discovery is an involved process spanning efforts from several fields of study and typically requires years of research and development. However, the advent of high-throughput genomic technologies has allowed for the use of in silico, genomics-based methods to screen drug libraries and accelerate drug discovery. Here we present a novel approach to computationally identify drug candidates for the treatment of breast cancer. In particular, we developed a Drug Regulatory Score similarity metric to evaluate gene expression profile similarity, in the context of drug treatment, and incorporated time-to-event patient survival information to develop an integrated analysis pipeline: Integrated Drug Expression Analysis (IDEA). We were able to predict drug candidates that have been known and those that have not been known in the literature to exhibit anticancer effects. Overall, our method enables quick preclinical screening of drug candidates for breast cancer and other diseases by using the most important indicator of drug efficacy: survival.

No MeSH data available.


Related in: MedlinePlus

Use of DRS profiles to predict metastasis. (a) Hierarchical clustering of DRS profiles. Magenta sample labels indicate metastatic tumors and aqua sample labels indicate nonmetastatic tumors. (b) Receiver operating characteristic curve for random forest machine learning classifier, using 84 most significant DTPs.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4544055&req=5

fig07: Use of DRS profiles to predict metastasis. (a) Hierarchical clustering of DRS profiles. Magenta sample labels indicate metastatic tumors and aqua sample labels indicate nonmetastatic tumors. (b) Receiver operating characteristic curve for random forest machine learning classifier, using 84 most significant DTPs.

Mentions: Since metastasis is an important driver of cancer progression, we postulated that there would be a relationship between metastatic tumors and their DRS profiles. If DRS does indeed capture information regarding metastatic tendency, then the distribution of DRS should be different between metastatic and nonmetastatic tumors. Thus, we also aimed to evaluate the predictive power of DRS in stratifying patients based on whether their tumors became metastatic by applying IDEA to an independent breast cancer dataset from van de Vijver et al.19 We first identified 84 DRS profiles that exhibited the greatest difference between metastatic and nonmetastatic tumor samples and then clustered samples based on these DRS profiles to derive three apparent clusters corresponding to the metastatic identity of the samples (see Methods). The first cluster had 48 metastatic samples out of 103 (47%), the second cluster had 27 metastatic samples out of 99 (27%), and the third cluster had eight metastatic clusters out of 58 (14%) (Figure7a). This indicates that DRS profiles can distinguish differences in metastatic potential between breast cancer tumors, and implies that treatment with a subset (cluster) of these drugs may predispose a tumor to take on a prometastatic molecular identity. Conversely, another subset may reverse a tumor's tendency to metastasize. To further evaluate this idea, we trained a random forest classifier with the same DRS profile features to determine if it could accurately stratify metastatic and nonmetastatic tumors, and evaluated its performance using 10-fold cross-validation. The model achieved an AUC of 0.71 calculated from the ROC curve, suggesting that DRS is an informative predictor when classifying tumors based on a known survival-associated phenotype (Figure7b). Furthermore, we implemented IDEA in the van de Vijver et al. dataset and were able to validate 28 drugs (P < 0.05, Wald test) including wortmannin, LY-294002, and etoposide (Supplementary Table S9). Additionally, we were able to validate 20 of the 25 top candidates in an independent meta-dataset compiled and normalized by Ur-Rehman et al.,21 indicating that our methodology is robust across different datasets. (P < 0.05, Supplementary Table S10).


IDEA: Integrated Drug Expression Analysis-Integration of Gene Expression and Clinical Data for the Identification of Therapeutic Candidates.

Ung MH, Varn FS, Cheng C - CPT Pharmacometrics Syst Pharmacol (2015)

Use of DRS profiles to predict metastasis. (a) Hierarchical clustering of DRS profiles. Magenta sample labels indicate metastatic tumors and aqua sample labels indicate nonmetastatic tumors. (b) Receiver operating characteristic curve for random forest machine learning classifier, using 84 most significant DTPs.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4544055&req=5

fig07: Use of DRS profiles to predict metastasis. (a) Hierarchical clustering of DRS profiles. Magenta sample labels indicate metastatic tumors and aqua sample labels indicate nonmetastatic tumors. (b) Receiver operating characteristic curve for random forest machine learning classifier, using 84 most significant DTPs.
Mentions: Since metastasis is an important driver of cancer progression, we postulated that there would be a relationship between metastatic tumors and their DRS profiles. If DRS does indeed capture information regarding metastatic tendency, then the distribution of DRS should be different between metastatic and nonmetastatic tumors. Thus, we also aimed to evaluate the predictive power of DRS in stratifying patients based on whether their tumors became metastatic by applying IDEA to an independent breast cancer dataset from van de Vijver et al.19 We first identified 84 DRS profiles that exhibited the greatest difference between metastatic and nonmetastatic tumor samples and then clustered samples based on these DRS profiles to derive three apparent clusters corresponding to the metastatic identity of the samples (see Methods). The first cluster had 48 metastatic samples out of 103 (47%), the second cluster had 27 metastatic samples out of 99 (27%), and the third cluster had eight metastatic clusters out of 58 (14%) (Figure7a). This indicates that DRS profiles can distinguish differences in metastatic potential between breast cancer tumors, and implies that treatment with a subset (cluster) of these drugs may predispose a tumor to take on a prometastatic molecular identity. Conversely, another subset may reverse a tumor's tendency to metastasize. To further evaluate this idea, we trained a random forest classifier with the same DRS profile features to determine if it could accurately stratify metastatic and nonmetastatic tumors, and evaluated its performance using 10-fold cross-validation. The model achieved an AUC of 0.71 calculated from the ROC curve, suggesting that DRS is an informative predictor when classifying tumors based on a known survival-associated phenotype (Figure7b). Furthermore, we implemented IDEA in the van de Vijver et al. dataset and were able to validate 28 drugs (P < 0.05, Wald test) including wortmannin, LY-294002, and etoposide (Supplementary Table S9). Additionally, we were able to validate 20 of the 25 top candidates in an independent meta-dataset compiled and normalized by Ur-Rehman et al.,21 indicating that our methodology is robust across different datasets. (P < 0.05, Supplementary Table S10).

Bottom Line: Cancer drug discovery is an involved process spanning efforts from several fields of study and typically requires years of research and development.However, the advent of high-throughput genomic technologies has allowed for the use of in silico, genomics-based methods to screen drug libraries and accelerate drug discovery.We were able to predict drug candidates that have been known and those that have not been known in the literature to exhibit anticancer effects.

View Article: PubMed Central - PubMed

Affiliation: Department of Genetics, Geisel School of Medicine at Dartmouth Hanover, New Hampshire, USA.

ABSTRACT
Cancer drug discovery is an involved process spanning efforts from several fields of study and typically requires years of research and development. However, the advent of high-throughput genomic technologies has allowed for the use of in silico, genomics-based methods to screen drug libraries and accelerate drug discovery. Here we present a novel approach to computationally identify drug candidates for the treatment of breast cancer. In particular, we developed a Drug Regulatory Score similarity metric to evaluate gene expression profile similarity, in the context of drug treatment, and incorporated time-to-event patient survival information to develop an integrated analysis pipeline: Integrated Drug Expression Analysis (IDEA). We were able to predict drug candidates that have been known and those that have not been known in the literature to exhibit anticancer effects. Overall, our method enables quick preclinical screening of drug candidates for breast cancer and other diseases by using the most important indicator of drug efficacy: survival.

No MeSH data available.


Related in: MedlinePlus