- sensitivity learned from cancer cell lines. - Using trametinib as an example, we showed that our approach can faithfully recapitulate the known tumor specificity of the drug.. - GDSC also quantified the basal level gene expression of many of the cancer cell lines using microarray [1]. - Full list of author information is available at the end of the article. - Recently, the transcriptomes of the NCI-60 cancer cell lines were also analyzed using RNA-seq [9]. - they subsequently applied the models to gene-expression data from TCGA tumor samples to impute sensitivity of the tumor samples to 138 drugs.. - We predicted the IC 50 values of the can- cer cell lines in the testing set for each of the 453 drugs.. - indicating that the basal transcriptomes of the cancer cell lines can reasonably predict the sensitivity (IC 50 s) of the cell lines to most of the drugs. - Of the 453 drugs had both ρ P and ρ s testing-set correlations ≥0.4.. - Interestingly, for 34 (7.5%) of the drugs, the cancer cell lines’ transcriptomes had little or no predictive power for the cell lines’ sensitivities to the drug (either ρ P or ρ s. - for most of the others, however, data availability was not an issue.. - For each of the top 272 predictable drugs, we counted how many times each gene was selected into the 100 sets of the d ( d = 30) predictive genes. - 1 Schematic diagram of the work-flow. - We aimed to identify a 30-gene set whose gene expression levels are most predictive of the IC 50 values of the drug for the samples in the testing set. - The resulting model (a 30-gene set) was subsequently used to predict the IC 50 value of the TCGA/GTEx samples. - The predicted IC 50 values from the 100 runs were then averaged and taken as the predicted IC 50 value of the drug for the samples. - Table 1 Summary statistics of correlations between the observed and predicted ln (IC 50 ) of the 453 drugs in the test set Correlation Min. - Many of the drug-gene interactions were also identified by others [1, 2, 5].. - 0.3) with the IC 50 values of more than 100 drugs for those cell lines (additional file 3: Table S3), suggesting that higher expression of C19orf33 in cancer cell lines was positively associated with higher resistance of the cancer cell lines to the drugs. - Most of the positively correlated drugs are DNA synthe- sis inhibitors, microtubule assembly inhibitors, or cell cycle inhibitors. - Interestingly, C19orf33 expression in cancer cell lines showed a negative correlation with the IC 50 values of the kinase (MEK, ERK, SRC) inhibitors for the cell lines (additional file 3: Table S3), suggesting that cancer cell lines with higher C19orf33 expression are more sensitive to kinase inhibitors than those with lower C19orf33 expression. - For 14 drugs of diverse mechanisms of action, ABCB1 ap- peared in more than 20% of the predictive gene sets. - ABCB1 encodes ATP binding cassette subfamily B member 1, a member of the superfamily of ATP-binding cassette (ABC) transporters. - Many of the most frequently selected genes were also known targets of the drugs (additional file 2: Table S2).. - The drug nutlin-3a inhibits the interaction between p53 and MDM2, leading to activation of the p53 pathway [31]. - BCL2 was selected in almost 100% of the predictive gene sets for sensitivity to venetoclax. - AICAR prevents the production of the enzymes adenosine kinase (ADK) and adenosine deami- nase (ADA) [36].. - Our results suggest that expression levels of SPRY2 and ETV4 are likely indicative of the sensitivity of cancer cell lines to many MAP kinase inhibitors. - On the other hand, TRPM4 was selected in fewer than 5% of the sets for all other drugs, suggesting that the TRPM4 -acetalax interaction is spe- cific. - From the cell-line data, we observed that the ranks of the observed IC 50 values of trametinib in the 571 cancer cell lines were associated with the corresponding IC 50. - e.g., a few of the PAAD (pancreatic adenocarcinoma) samples. - For the description of the 33 TCGA tumor types, see supplementary data (additional file 1: Table S7A). - The solid line shows the median of the medians of the predicted IC 50 values for all 33 tumor types whereas the dashed line is one logarithmic unit below the solid line. - b, Violin plots of the predicted ln (IC 50 ) values of trametinib for COAD tumor (red) and normal (blue) samples from TCGA and for GTEx normal tissue samples from 15 major organs (green). - here the solid line shows the median of the medians of the predicted IC 50. - NRAS and KRAS are upstream molecules of the BRAF signaling pathway [50]. - Other top-ranked genes whose mutations were found to be associated with predicted sensitivity to trametinib in four of the tumor types include ADAMTS2 , ANKRD5 , MYCBP2 , TTN , and VCAN. - Overall, most of the TCGA tumor samples were predicted to be highly sensitive (pan cancer median predicted ln (IC 50 ) <. - 0) to about 35 of the 272 drugs (additional file 5: Table S4). - Many of the drugs target DNA/protein synthesis, cell cycle, microtubules, and the mTOR pathway. - Most of the drugs were also predicted to be similarly cytotoxic to normal samples from TCGA (additional file 5: Table S4). - For each of the 272 drugs, we compared the median pre- dicted IC 50 of the drug for all tumor samples with the median predicted IC 50 value for all normal samples from the same tumor type from TCGA. - We identified eight drugs whose median predicted ln (IC 50 ) value for tumor samples was more than one loga- rithmic unit lower than that for corresponding normal samples in at least one of the 14 tumor types (additional file 4: Fig. - value among samples from one tumor type with the me- dian of the medians of the predicted IC 50 values from all 33 tumor types. - 2.7 times) lower than the median of the medians from all tumor types.. - Interestingly, 12 of the remaining 13 drugs that were predicted not specific for DLBC, THYM or both are kinase inhibitors, consistent with the notion that kinase inhibitors target specific cel- lular pathways. - The ln (IC 50 ) values of the drugs were predicted based on the RNA-seq data of the tumor and normal tissue samples from TCGA. - 2 legend for additional description of the violin plots. - of the 270 drugs, we divided the ~ 1100 TCGA BRCA sam- ples into five subgroups (basal-like, Her2-positive, luminal A, luminal B, and normal-like) based on the PAM50 classi- fication [53, 54]. - For each subtype, we compared the me- dian of the predicted IC 50 values of a drug for the samples of the subtype with the median of the medians of the pre- dicted IC 50 values for the five subtypes. - We also predicted that basal-like subtype breast cancer has higher sensitivity to five (bleomycin, daporinad, sepantronium bromide, etoposide, and ICL1100013) of the seven drugs, luminal B subtype breast cancer has higher sensitivity to ABT737 and navito- clax, both of which are BCL2 inhibitors.. - Violin plots of the predicted ln (IC 50 ) values of Acetalax a, Alisertib b, Dasatinib c, Debrafenib d, OSI-027 e, and Sapitinib f for TCGA tumor samples from 33 tumor types. - The solid line shows the median of the medians of the predicted IC 50 values for all 33 tumor types. - In this work, we began by investigating if a cancer cell line’s transcriptome [4, 5] can predict the IC 50 of a drug acting on that cell line for each of the 473 GDSC drugs and 1019 cell lines [1–3]. - We found that, for about half of the drugs, transcriptomes were reasonably predictive of the sensitivity of the cell lines to those drugs, i.e., that Spearman correlation between predicted and observed IC 50 values >. - Our results also revealed that SPRY2 expression is positively correlated with the sensitivity of the cancer cell lines to many MEK inhibitors from GDSC, suggesting that SPRY2 expression may be a predictive biomarker for the effectiveness of MEK kinase Table 2 Drugs to which breast cancer subtype(s) in TCGA samples were predicted to be sensitive. - Many of the putative biomarkers that we identified in this study may be ‘proxy’ markers for oncomutations.. - We have no direct evidence of a causal relationship between the expression of the predictive genes and the sensitivity of cell lines to the drugs. - Our analyses of the TCGA tumors revealed that trametinib has the highest specificity for melanoma, colorectal cancer, and rectal cancer among all 33 TCGA tumor types, consistent with the clinical application of trametinib. - We found that some of the drugs are highly cytotoxic to all tumors from all tumor types. - a, Predicted bleomycin sensitivity for the five subtypes of TCGA BRCA samples: violin plots of the predicted ln (IC 50 ) values of bleomycin for the five subtypes of breast tumors based on gene expression data and PAM50 classification of TCGA BRCA samples. - our analysis also found that those tumors were also more sensitive to trametinib than the normal tissues of the same origins.. - The predicted value of a sample is taken as the average of the values of its k -nearest neighbors. - Because of the averaging, the most extreme predicted values, either high or low, usually cannot be as extreme as the corresponding observed values. - Therefore, although the correlation between the predicted and observed values can be high, e.g., 0.8, the magnitude of the predicted values is generally pulled in from the extremes. - the trend of the predicted values among the samples, however, is usually preserved (Fig. - Overview of the GA/KNN algorithm. - In the present context, the main idea of the GA/KNN algorithm is to use an evolutionary algorithm to select many sets of d genes (see below) whose expression levels can accurately predict observed IC 50 values using the k -nearest-neighbors prediction rule.. - The prediction rule is simple: the predicted IC 50 value of a sample is defined as the average of the observed IC 50. - For prediction, a typical objective function being minimized is the sum of the squared deviations between the observed and predicted IC 50 values across all training samples (i.e., squared-error loss). - We used the training data to identify a set of d ( d = 30) genes whose expression levels were best pre- dictive of the IC 50 values of samples in the training set. - That set of d genes was subsequently used to predict the IC 50 values of the testing-set samples. - value of the k -nearest ( k = 3) training neighbors of a test- ing sample was taken as the predicted IC 50 value for the testing sample. - The final predicted value for a training-set sample was the average of the pre- dicted IC 50 values for that sample over the subset of the 100 independent training-testing partitions in which that sample appeared in a training set. - Because each training-testing partition provided a set of 30 genes as predictors, we used the frequency with which a gene was selected into the 100 sets of 30 predictor genes as a measure of the importance of that gene in prediction.. - For each of the 272 drugs, we repeated the same GA/KNN procedure applied to the cell-line data to both the tumor and the GTEx data. - Specifically, for each of the 272 drugs, we randomly partitioned the part of the tumor data from the CCLE cell lines into a training set (90%) and a testing set (10. - Drug sensitivity of cancer cell lines. - GDSC reports the ln (IC 50 ) for each combination of cell line and compound as a measure of the sensitivity of cell viability in that cell line to the compound. - Gene expression of cancer cell lines. - The distribution of the number of cancer cell lines per drug across the 573 drugs is summarized for each cancer type in Table 7A.. - Accordingly, for each of the 453 drugs from GDSC, we have CCLE gene expression profiles for a subset of the cell lines with IC 50 s for that drug. - We augmented each of the 453 matrices of cell-line data with columns of RNA-seq expression profiles for the tumor samples from the TCGA using the common genes between the two ( G = 19,163). - Similarly, we augmented each of the 453 matrices of cell-line data with columns of RNA-seq expression profiles for the normal tissue samples from GTEx using the common genes between the two ( G = 19163). - Hormone status of the TCGA breast invasive carcinoma (BRCA) tumor samples (file name: BRCA.clin.merged.txt) was downloaded from the Broad GDAC firehose (https://. - are the mean and standard deviation of the expression values for sample i. - Main parameters of the GA/KNN algorithm used for the analyses of all datasets. - Inverse correlation between SPRY2 expression (Z score) in cancer cell lines and the observed ln (IC 50 ) of the six MEK inhibitors for those cell lines. - Scatter plot of the counts of genes selected into the sets of 30 chro- mosomes from two independent runs with 100 runs and 1000 runs, respectively. - The PAM50 classification of the TCGA breast cancer tumor samples was kindly provided by Joel Parker (UNC). - Research Program of the National Institutes of Health, National Institute of Environmental Health Sciences (ES101765).. - This research was supported by Intramural Research Program of the National Institutes of Health, National Institute of Environmental Health Sciences (ES101765).. - Next-generation characterization of the Cancer Cell Line Encyclopedia. - RNA Sequencing of the NCI-60: Integration into CellMiner and CellMiner CDB.. - pharmacogenomics analyses of cancer cell lines. - In vivo activation of the p53 pathway by small-molecule antagonists of MDM2.. - A phase Ib dose-escalation study of the oral pan-PI3K inhibitor buparlisib (BKM120) in combination with the oral MEK1/2 inhibitor trametinib (GSK1120212) in patients with selected advanced solid tumors.. - Development and verification of the PAM50-based Prosigna breast cancer gene signature assay. - Cancer Cell. - Gene selection for sample classification based on gene expression data: study of sensitivity to choice of parameters of the GA/KNN method
Xem thử không khả dụng, vui lòng xem tại trang nguồn hoặc xem
Tóm tắt