Limits...
Intra- and interspecies gene expression models for predicting drug response in canine osteosarcoma.

Fowles JS, Brown KC, Hess AM, Duval DL, Gustafson DL - BMC Bioinformatics (2016)

Bottom Line: We evaluated the use of gene-expression based models built in an intra- or interspecies manner to predict chemosensitivity and treatment outcome in canine OS.Models were built using linear discriminant analysis via the misclassification penalized posterior algorithm.The best carboplatin model utilized canine lines for gene identification and model training, with canine OS tumor data for co-expression.

View Article: PubMed Central - PubMed

Affiliation: Cell and Molecular Biology Program, Department of Clinical Sciences, Colorado State University, Fort Collins, CO, USA.

ABSTRACT

Background: Genomics-based predictors of drug response have the potential to improve outcomes associated with cancer therapy. Osteosarcoma (OS), the most common primary bone cancer in dogs, is commonly treated with adjuvant doxorubicin or carboplatin following amputation of the affected limb. We evaluated the use of gene-expression based models built in an intra- or interspecies manner to predict chemosensitivity and treatment outcome in canine OS. Models were built and evaluated using microarray gene expression and drug sensitivity data from human and canine cancer cell lines, and canine OS tumor datasets. The "COXEN" method was utilized to filter gene signatures between human and dog datasets based on strong co-expression patterns. Models were built using linear discriminant analysis via the misclassification penalized posterior algorithm.

Results: The best doxorubicin model involved genes identified in human lines that were co-expressed and trained on canine OS tumor data, which accurately predicted clinical outcome in 73 % of dogs (p = 0.0262, binomial). The best carboplatin model utilized canine lines for gene identification and model training, with canine OS tumor data for co-expression. Dogs whose treatment matched our predictions had significantly better clinical outcomes than those that didn't (p = 0.0006, Log Rank), and this predictor significantly associated with longer disease free intervals in a Cox multivariate analysis (hazard ratio = 0.3102, p = 0.0124).

Conclusions: Our data show that intra- and interspecies gene expression models can successfully predict response in canine OS, which may improve outcome in dogs and serve as pre-clinical validation for similar methods in human cancer research.

No MeSH data available.


Related in: MedlinePlus

Cell line-trained models on clinical outcome in canine osteosarcoma patients treated with doxorubicin and/or carboplatin. a & b Analysis comparing the survival curves of COS33 patients predicted to respond or not respond to doxorubicin (n = 22) (a) or carboplatin (n = 25) (b) from a NCI60-trained model with the COS16 tumor panel used as the co-expression set. c Survival analysis of predicted responders and non-responders in the COS33 to doxorubicin from a model trained on the osteosarcoma cell line subset of the FACC panel, with the COS16 used for co-expression. d Survival analysis of predicted responders and non-responders in the COS33 to carboplatin from a FACC-trained model co-expressed with the COS16. Significant difference in disease free interval between predicted groups was determined by Log Rank test
© Copyright Policy - OpenAccess
Related In: Results  -  Collection

License 1 - License 2
getmorefigures.php?uid=PMC4759767&req=5

Fig3: Cell line-trained models on clinical outcome in canine osteosarcoma patients treated with doxorubicin and/or carboplatin. a & b Analysis comparing the survival curves of COS33 patients predicted to respond or not respond to doxorubicin (n = 22) (a) or carboplatin (n = 25) (b) from a NCI60-trained model with the COS16 tumor panel used as the co-expression set. c Survival analysis of predicted responders and non-responders in the COS33 to doxorubicin from a model trained on the osteosarcoma cell line subset of the FACC panel, with the COS16 used for co-expression. d Survival analysis of predicted responders and non-responders in the COS33 to carboplatin from a FACC-trained model co-expressed with the COS16. Significant difference in disease free interval between predicted groups was determined by Log Rank test

Mentions: Our next step was to compare the ability of cell line-trained prediction models to accurately predict clinical outcome in an independent canine osteosarcoma tumor dataset (COS33). Sample information for the COS33 tumor panel is provide in Additional file 4: Table S3. Individual reference sets used in the different models included human and canine cancer cell line panels containing multiple tumor types (NCI60, FACC) as well as osteosarcoma-only subsets from the GDSC and FACC panels (GDSCosteo, FACCosteo). After DEGs for DOX or CARBO were identified in the reference set,they were further filtered based on co-expression analysis with a canine osteosarcoma tumor panel COS16 (Table 1). Models were then trained on the corresponding reference set and tested independently on the COS33 tumor panel. Data from a historic cohort study of 470 dogs treated for osteosarcoma [37] were used to determine cutoffs between “responders” and “non-responders” in both canine osteosarcoma tumor datasets based on median disease free interval of dogs that received doxorubicin (276 days) or carboplatin (296 days). The original cutoffs for the COS16 panel reported by O’Donoghue et al. fit within our definition of “responders” and “non-responders”, so no adjustments to group classification needed to be made [31]. All of our modeling results are reported in Additional file 5: Table S4, with error rates based on external validation in the test set. The NCI60 model had an error rate of 0.3182 compared to 0.3043 for the FACC model (p = 0.0669 and 0.0466, binomial) (Additional file 5: Table S4). However, the NCI60 model resulted in better curve separation of predicted responders and non-responders compared to the FACC model in the survival curve analysis (Additional file 5: Table S4, Fig. 3a).Fig. 3


Intra- and interspecies gene expression models for predicting drug response in canine osteosarcoma.

Fowles JS, Brown KC, Hess AM, Duval DL, Gustafson DL - BMC Bioinformatics (2016)

Cell line-trained models on clinical outcome in canine osteosarcoma patients treated with doxorubicin and/or carboplatin. a & b Analysis comparing the survival curves of COS33 patients predicted to respond or not respond to doxorubicin (n = 22) (a) or carboplatin (n = 25) (b) from a NCI60-trained model with the COS16 tumor panel used as the co-expression set. c Survival analysis of predicted responders and non-responders in the COS33 to doxorubicin from a model trained on the osteosarcoma cell line subset of the FACC panel, with the COS16 used for co-expression. d Survival analysis of predicted responders and non-responders in the COS33 to carboplatin from a FACC-trained model co-expressed with the COS16. Significant difference in disease free interval between predicted groups was determined by Log Rank test
© Copyright Policy - OpenAccess
Related In: Results  -  Collection

License 1 - License 2
Show All Figures
getmorefigures.php?uid=PMC4759767&req=5

Fig3: Cell line-trained models on clinical outcome in canine osteosarcoma patients treated with doxorubicin and/or carboplatin. a & b Analysis comparing the survival curves of COS33 patients predicted to respond or not respond to doxorubicin (n = 22) (a) or carboplatin (n = 25) (b) from a NCI60-trained model with the COS16 tumor panel used as the co-expression set. c Survival analysis of predicted responders and non-responders in the COS33 to doxorubicin from a model trained on the osteosarcoma cell line subset of the FACC panel, with the COS16 used for co-expression. d Survival analysis of predicted responders and non-responders in the COS33 to carboplatin from a FACC-trained model co-expressed with the COS16. Significant difference in disease free interval between predicted groups was determined by Log Rank test
Mentions: Our next step was to compare the ability of cell line-trained prediction models to accurately predict clinical outcome in an independent canine osteosarcoma tumor dataset (COS33). Sample information for the COS33 tumor panel is provide in Additional file 4: Table S3. Individual reference sets used in the different models included human and canine cancer cell line panels containing multiple tumor types (NCI60, FACC) as well as osteosarcoma-only subsets from the GDSC and FACC panels (GDSCosteo, FACCosteo). After DEGs for DOX or CARBO were identified in the reference set,they were further filtered based on co-expression analysis with a canine osteosarcoma tumor panel COS16 (Table 1). Models were then trained on the corresponding reference set and tested independently on the COS33 tumor panel. Data from a historic cohort study of 470 dogs treated for osteosarcoma [37] were used to determine cutoffs between “responders” and “non-responders” in both canine osteosarcoma tumor datasets based on median disease free interval of dogs that received doxorubicin (276 days) or carboplatin (296 days). The original cutoffs for the COS16 panel reported by O’Donoghue et al. fit within our definition of “responders” and “non-responders”, so no adjustments to group classification needed to be made [31]. All of our modeling results are reported in Additional file 5: Table S4, with error rates based on external validation in the test set. The NCI60 model had an error rate of 0.3182 compared to 0.3043 for the FACC model (p = 0.0669 and 0.0466, binomial) (Additional file 5: Table S4). However, the NCI60 model resulted in better curve separation of predicted responders and non-responders compared to the FACC model in the survival curve analysis (Additional file 5: Table S4, Fig. 3a).Fig. 3

Bottom Line: We evaluated the use of gene-expression based models built in an intra- or interspecies manner to predict chemosensitivity and treatment outcome in canine OS.Models were built using linear discriminant analysis via the misclassification penalized posterior algorithm.The best carboplatin model utilized canine lines for gene identification and model training, with canine OS tumor data for co-expression.

View Article: PubMed Central - PubMed

Affiliation: Cell and Molecular Biology Program, Department of Clinical Sciences, Colorado State University, Fort Collins, CO, USA.

ABSTRACT

Background: Genomics-based predictors of drug response have the potential to improve outcomes associated with cancer therapy. Osteosarcoma (OS), the most common primary bone cancer in dogs, is commonly treated with adjuvant doxorubicin or carboplatin following amputation of the affected limb. We evaluated the use of gene-expression based models built in an intra- or interspecies manner to predict chemosensitivity and treatment outcome in canine OS. Models were built and evaluated using microarray gene expression and drug sensitivity data from human and canine cancer cell lines, and canine OS tumor datasets. The "COXEN" method was utilized to filter gene signatures between human and dog datasets based on strong co-expression patterns. Models were built using linear discriminant analysis via the misclassification penalized posterior algorithm.

Results: The best doxorubicin model involved genes identified in human lines that were co-expressed and trained on canine OS tumor data, which accurately predicted clinical outcome in 73 % of dogs (p = 0.0262, binomial). The best carboplatin model utilized canine lines for gene identification and model training, with canine OS tumor data for co-expression. Dogs whose treatment matched our predictions had significantly better clinical outcomes than those that didn't (p = 0.0006, Log Rank), and this predictor significantly associated with longer disease free intervals in a Cox multivariate analysis (hazard ratio = 0.3102, p = 0.0124).

Conclusions: Our data show that intra- and interspecies gene expression models can successfully predict response in canine OS, which may improve outcome in dogs and serve as pre-clinical validation for similar methods in human cancer research.

No MeSH data available.


Related in: MedlinePlus