Abstract
Background: Heterogeneity within triple-negative breast cancer (TNBC) makes it challenging to target therapeutically. Despite recent advances in the standard of care treatment of stage II-III TNBC, which now includes multiple chemotherapeutic agents plus an immune checkpoint inhibitor, the cure rate for these patients remains suboptimal. There is still a critical need for improved molecular characterization of TNBC and the development of accurate prognostic tools to assist in making treatment escalation and de-escalation decisions. Here, we present an analysis of the DNA and RNA landscape in a set of patients treated on a clinical trial and validation of molecular and prognostic features using three additional data sets.
Methods: CALGB 40603 (Alliance) is a randomized phase II trial that evaluated the impact of adding carboplatin and/or bevacizumab to standard anthracycline/taxane neoadjuvant chemotherapy in stage II-III TNBC. Targeted panel DNA sequencing of pre-treatment tumor samples from 238 patients enrolled in this study was performed and analyzed alongside matched pre-treatment RNAseq data. These results were then combined with clinical outcomes data to develop a prognostic multi-variable elastic net model of overall survival. We subsequently evaluated this model on samples from 380 stage II-III TNBC patients from three independent publicly available datasets.
Results: Similar to other TNBC data sets, DNAseq results of 40603 identified TP53 (86%) as the most frequent somatically mutated gene, followed by MT-ND5 (16%), MT-ND4 (12%), CSMD3 (8%), and PIK3CA (7%). BRCA1 (germline 8%, somatic 4%), BRCA2 (germline 2%, somatic 0.4%), and PALB2 (germline 1%) gene mutations were also detected, and when combined into a single “homologous recombination deficiency” category totaled 15%. The DNA copy number landscape of TNBC was very similar across the four data sets, and largely mirrored the landscape seen for basal-like breast cancer characterized in TCGA. Importantly, only one somatically mutated gene (PIK3R2), and no DNA copy number altered regions, were associated with pathologic complete response or survival at a false discovery rate < 0.05. We next sought to use these multi-platform data to train a prognostic Cox elastic net regression model using CALGB 40603 data and a diverse set of features, including tumor stage, 804 RNA expression signatures, 727 genes with somatic mutations, and 534 chromosomal segment-level copy number alterations. This yielded a 31-feature prognostic model with C-index performance values (95% confidence interval) of 0.68 (0.66-0.69), 0.61 (0.59-0.63), and 0.83 (0.82-0.84) using FUSCC (PMID 30853353), METABRIC (PMID 22522925), and TCGA (PMID 23000897) test sets. The most heavily weighted features in this model were tumor stage and multiple immune signatures including an IgG signature, and this model had higher C-index values than a model trained using only tumor stage.
Conclusions: This study provides a comprehensive and integrated characterization of the DNA- and RNA-based landscape of a large stage II-III TNBC patient data set. Through the development of a multi-omic elastic net model of TNBC survival, we show that we may improve the prognostic accuracy for overall survival beyond stage alone by incorporating machine learning selected molecular features. Furthermore, the model’s ability to identify good overall survival outcome patients when given neoadjuvant chemotherapy illustrates its potential utility for informing treatment escalation/de-escalation decisions, which could be clinically valuable if further validated.
Support: U10CA180821, U10CA180882; U24CA176171; U10CA180888 (SWOG); P50-CA058223; Genentech; https://acknowledgments.alliancefound.org. Clinicaltrials.gov Id#: NCT00861705
Methods: CALGB 40603 (Alliance) is a randomized phase II trial that evaluated the impact of adding carboplatin and/or bevacizumab to standard anthracycline/taxane neoadjuvant chemotherapy in stage II-III TNBC. Targeted panel DNA sequencing of pre-treatment tumor samples from 238 patients enrolled in this study was performed and analyzed alongside matched pre-treatment RNAseq data. These results were then combined with clinical outcomes data to develop a prognostic multi-variable elastic net model of overall survival. We subsequently evaluated this model on samples from 380 stage II-III TNBC patients from three independent publicly available datasets.
Results: Similar to other TNBC data sets, DNAseq results of 40603 identified TP53 (86%) as the most frequent somatically mutated gene, followed by MT-ND5 (16%), MT-ND4 (12%), CSMD3 (8%), and PIK3CA (7%). BRCA1 (germline 8%, somatic 4%), BRCA2 (germline 2%, somatic 0.4%), and PALB2 (germline 1%) gene mutations were also detected, and when combined into a single “homologous recombination deficiency” category totaled 15%. The DNA copy number landscape of TNBC was very similar across the four data sets, and largely mirrored the landscape seen for basal-like breast cancer characterized in TCGA. Importantly, only one somatically mutated gene (PIK3R2), and no DNA copy number altered regions, were associated with pathologic complete response or survival at a false discovery rate < 0.05. We next sought to use these multi-platform data to train a prognostic Cox elastic net regression model using CALGB 40603 data and a diverse set of features, including tumor stage, 804 RNA expression signatures, 727 genes with somatic mutations, and 534 chromosomal segment-level copy number alterations. This yielded a 31-feature prognostic model with C-index performance values (95% confidence interval) of 0.68 (0.66-0.69), 0.61 (0.59-0.63), and 0.83 (0.82-0.84) using FUSCC (PMID 30853353), METABRIC (PMID 22522925), and TCGA (PMID 23000897) test sets. The most heavily weighted features in this model were tumor stage and multiple immune signatures including an IgG signature, and this model had higher C-index values than a model trained using only tumor stage.
Conclusions: This study provides a comprehensive and integrated characterization of the DNA- and RNA-based landscape of a large stage II-III TNBC patient data set. Through the development of a multi-omic elastic net model of TNBC survival, we show that we may improve the prognostic accuracy for overall survival beyond stage alone by incorporating machine learning selected molecular features. Furthermore, the model’s ability to identify good overall survival outcome patients when given neoadjuvant chemotherapy illustrates its potential utility for informing treatment escalation/de-escalation decisions, which could be clinically valuable if further validated.
Support: U10CA180821, U10CA180882; U24CA176171; U10CA180888 (SWOG); P50-CA058223; Genentech; https://acknowledgments.alliancefound.org. Clinicaltrials.gov Id#: NCT00861705
Background: Heterogeneity within triple-negative breast cancer (TNBC) makes it challenging to target therapeutically. Despite recent advances in the standard of care treatment of stage II-III TNBC, which now includes multiple chemotherapeutic agents plus an immune checkpoint inhibitor, the cure rate for these patients remains suboptimal. There is still a critical need for improved molecular characterization of TNBC and the development of accurate prognostic tools to assist in making treatment escalation and de-escalation decisions. Here, we present an analysis of the DNA and RNA landscape in a set of patients treated on a clinical trial and validation of molecular and prognostic features using three additional data sets.
Methods: CALGB 40603 (Alliance) is a randomized phase II trial that evaluated the impact of adding carboplatin and/or bevacizumab to standard anthracycline/taxane neoadjuvant chemotherapy in stage II-III TNBC. Targeted panel DNA sequencing of pre-treatment tumor samples from 238 patients enrolled in this study was performed and analyzed alongside matched pre-treatment RNAseq data. These results were then combined with clinical outcomes data to develop a prognostic multi-variable elastic net model of overall survival. We subsequently evaluated this model on samples from 380 stage II-III TNBC patients from three independent publicly available datasets.
Results: Similar to other TNBC data sets, DNAseq results of 40603 identified TP53 (86%) as the most frequent somatically mutated gene, followed by MT-ND5 (16%), MT-ND4 (12%), CSMD3 (8%), and PIK3CA (7%). BRCA1 (germline 8%, somatic 4%), BRCA2 (germline 2%, somatic 0.4%), and PALB2 (germline 1%) gene mutations were also detected, and when combined into a single “homologous recombination deficiency” category totaled 15%. The DNA copy number landscape of TNBC was very similar across the four data sets, and largely mirrored the landscape seen for basal-like breast cancer characterized in TCGA. Importantly, only one somatically mutated gene (PIK3R2), and no DNA copy number altered regions, were associated with pathologic complete response or survival at a false discovery rate < 0.05. We next sought to use these multi-platform data to train a prognostic Cox elastic net regression model using CALGB 40603 data and a diverse set of features, including tumor stage, 804 RNA expression signatures, 727 genes with somatic mutations, and 534 chromosomal segment-level copy number alterations. This yielded a 31-feature prognostic model with C-index performance values (95% confidence interval) of 0.68 (0.66-0.69), 0.61 (0.59-0.63), and 0.83 (0.82-0.84) using FUSCC (PMID 30853353), METABRIC (PMID 22522925), and TCGA (PMID 23000897) test sets. The most heavily weighted features in this model were tumor stage and multiple immune signatures including an IgG signature, and this model had higher C-index values than a model trained using only tumor stage.
Conclusions: This study provides a comprehensive and integrated characterization of the DNA- and RNA-based landscape of a large stage II-III TNBC patient data set. Through the development of a multi-omic elastic net model of TNBC survival, we show that we may improve the prognostic accuracy for overall survival beyond stage alone by incorporating machine learning selected molecular features. Furthermore, the model’s ability to identify good overall survival outcome patients when given neoadjuvant chemotherapy illustrates its potential utility for informing treatment escalation/de-escalation decisions, which could be clinically valuable if further validated.
Support: U10CA180821, U10CA180882; U24CA176171; U10CA180888 (SWOG); P50-CA058223; Genentech; https://acknowledgments.alliancefound.org. Clinicaltrials.gov Id#: NCT00861705
Methods: CALGB 40603 (Alliance) is a randomized phase II trial that evaluated the impact of adding carboplatin and/or bevacizumab to standard anthracycline/taxane neoadjuvant chemotherapy in stage II-III TNBC. Targeted panel DNA sequencing of pre-treatment tumor samples from 238 patients enrolled in this study was performed and analyzed alongside matched pre-treatment RNAseq data. These results were then combined with clinical outcomes data to develop a prognostic multi-variable elastic net model of overall survival. We subsequently evaluated this model on samples from 380 stage II-III TNBC patients from three independent publicly available datasets.
Results: Similar to other TNBC data sets, DNAseq results of 40603 identified TP53 (86%) as the most frequent somatically mutated gene, followed by MT-ND5 (16%), MT-ND4 (12%), CSMD3 (8%), and PIK3CA (7%). BRCA1 (germline 8%, somatic 4%), BRCA2 (germline 2%, somatic 0.4%), and PALB2 (germline 1%) gene mutations were also detected, and when combined into a single “homologous recombination deficiency” category totaled 15%. The DNA copy number landscape of TNBC was very similar across the four data sets, and largely mirrored the landscape seen for basal-like breast cancer characterized in TCGA. Importantly, only one somatically mutated gene (PIK3R2), and no DNA copy number altered regions, were associated with pathologic complete response or survival at a false discovery rate < 0.05. We next sought to use these multi-platform data to train a prognostic Cox elastic net regression model using CALGB 40603 data and a diverse set of features, including tumor stage, 804 RNA expression signatures, 727 genes with somatic mutations, and 534 chromosomal segment-level copy number alterations. This yielded a 31-feature prognostic model with C-index performance values (95% confidence interval) of 0.68 (0.66-0.69), 0.61 (0.59-0.63), and 0.83 (0.82-0.84) using FUSCC (PMID 30853353), METABRIC (PMID 22522925), and TCGA (PMID 23000897) test sets. The most heavily weighted features in this model were tumor stage and multiple immune signatures including an IgG signature, and this model had higher C-index values than a model trained using only tumor stage.
Conclusions: This study provides a comprehensive and integrated characterization of the DNA- and RNA-based landscape of a large stage II-III TNBC patient data set. Through the development of a multi-omic elastic net model of TNBC survival, we show that we may improve the prognostic accuracy for overall survival beyond stage alone by incorporating machine learning selected molecular features. Furthermore, the model’s ability to identify good overall survival outcome patients when given neoadjuvant chemotherapy illustrates its potential utility for informing treatment escalation/de-escalation decisions, which could be clinically valuable if further validated.
Support: U10CA180821, U10CA180882; U24CA176171; U10CA180888 (SWOG); P50-CA058223; Genentech; https://acknowledgments.alliancefound.org. Clinicaltrials.gov Id#: NCT00861705
Paired DNA and RNA analysis of CALGB 40603 (Alliance) reveals insights into the molecular and prognostic landscape of stage II-III triple-negative breast cancer
Brooke Felsheim
Abstract
Background: Heterogeneity within triple-negative breast cancer (TNBC) makes it challenging to target therapeutically. Despite recent advances in the standard of care treatment of stage II-III TNBC, which now includes multiple chemotherapeutic agents plus an immune checkpoint inhibitor, the cure rate for these patients remains suboptimal. There is still a critical need for improved molecular characterization of TNBC and the development of accurate prognostic tools to assist in making treatment escalation and de-escalation decisions. Here, we present an analysis of the DNA and RNA landscape in a set of patients treated on a clinical trial and validation of molecular and prognostic features using three additional data sets.
Methods: CALGB 40603 (Alliance) is a randomized phase II trial that evaluated the impact of adding carboplatin and/or bevacizumab to standard anthracycline/taxane neoadjuvant chemotherapy in stage II-III TNBC. Targeted panel DNA sequencing of pre-treatment tumor samples from 238 patients enrolled in this study was performed and analyzed alongside matched pre-treatment RNAseq data. These results were then combined with clinical outcomes data to develop a prognostic multi-variable elastic net model of overall survival. We subsequently evaluated this model on samples from 380 stage II-III TNBC patients from three independent publicly available datasets.
Results: Similar to other TNBC data sets, DNAseq results of 40603 identified TP53 (86%) as the most frequent somatically mutated gene, followed by MT-ND5 (16%), MT-ND4 (12%), CSMD3 (8%), and PIK3CA (7%). BRCA1 (germline 8%, somatic 4%), BRCA2 (germline 2%, somatic 0.4%), and PALB2 (germline 1%) gene mutations were also detected, and when combined into a single “homologous recombination deficiency” category totaled 15%. The DNA copy number landscape of TNBC was very similar across the four data sets, and largely mirrored the landscape seen for basal-like breast cancer characterized in TCGA. Importantly, only one somatically mutated gene (PIK3R2), and no DNA copy number altered regions, were associated with pathologic complete response or survival at a false discovery rate < 0.05. We next sought to use these multi-platform data to train a prognostic Cox elastic net regression model using CALGB 40603 data and a diverse set of features, including tumor stage, 804 RNA expression signatures, 727 genes with somatic mutations, and 534 chromosomal segment-level copy number alterations. This yielded a 31-feature prognostic model with C-index performance values (95% confidence interval) of 0.68 (0.66-0.69), 0.61 (0.59-0.63), and 0.83 (0.82-0.84) using FUSCC (PMID 30853353), METABRIC (PMID 22522925), and TCGA (PMID 23000897) test sets. The most heavily weighted features in this model were tumor stage and multiple immune signatures including an IgG signature, and this model had higher C-index values than a model trained using only tumor stage.
Conclusions: This study provides a comprehensive and integrated characterization of the DNA- and RNA-based landscape of a large stage II-III TNBC patient data set. Through the development of a multi-omic elastic net model of TNBC survival, we show that we may improve the prognostic accuracy for overall survival beyond stage alone by incorporating machine learning selected molecular features. Furthermore, the model’s ability to identify good overall survival outcome patients when given neoadjuvant chemotherapy illustrates its potential utility for informing treatment escalation/de-escalation decisions, which could be clinically valuable if further validated.
Support: U10CA180821, U10CA180882; U24CA176171; U10CA180888 (SWOG); P50-CA058223; Genentech; https://acknowledgments.alliancefound.org. Clinicaltrials.gov Id#: NCT00861705
Methods: CALGB 40603 (Alliance) is a randomized phase II trial that evaluated the impact of adding carboplatin and/or bevacizumab to standard anthracycline/taxane neoadjuvant chemotherapy in stage II-III TNBC. Targeted panel DNA sequencing of pre-treatment tumor samples from 238 patients enrolled in this study was performed and analyzed alongside matched pre-treatment RNAseq data. These results were then combined with clinical outcomes data to develop a prognostic multi-variable elastic net model of overall survival. We subsequently evaluated this model on samples from 380 stage II-III TNBC patients from three independent publicly available datasets.
Results: Similar to other TNBC data sets, DNAseq results of 40603 identified TP53 (86%) as the most frequent somatically mutated gene, followed by MT-ND5 (16%), MT-ND4 (12%), CSMD3 (8%), and PIK3CA (7%). BRCA1 (germline 8%, somatic 4%), BRCA2 (germline 2%, somatic 0.4%), and PALB2 (germline 1%) gene mutations were also detected, and when combined into a single “homologous recombination deficiency” category totaled 15%. The DNA copy number landscape of TNBC was very similar across the four data sets, and largely mirrored the landscape seen for basal-like breast cancer characterized in TCGA. Importantly, only one somatically mutated gene (PIK3R2), and no DNA copy number altered regions, were associated with pathologic complete response or survival at a false discovery rate < 0.05. We next sought to use these multi-platform data to train a prognostic Cox elastic net regression model using CALGB 40603 data and a diverse set of features, including tumor stage, 804 RNA expression signatures, 727 genes with somatic mutations, and 534 chromosomal segment-level copy number alterations. This yielded a 31-feature prognostic model with C-index performance values (95% confidence interval) of 0.68 (0.66-0.69), 0.61 (0.59-0.63), and 0.83 (0.82-0.84) using FUSCC (PMID 30853353), METABRIC (PMID 22522925), and TCGA (PMID 23000897) test sets. The most heavily weighted features in this model were tumor stage and multiple immune signatures including an IgG signature, and this model had higher C-index values than a model trained using only tumor stage.
Conclusions: This study provides a comprehensive and integrated characterization of the DNA- and RNA-based landscape of a large stage II-III TNBC patient data set. Through the development of a multi-omic elastic net model of TNBC survival, we show that we may improve the prognostic accuracy for overall survival beyond stage alone by incorporating machine learning selected molecular features. Furthermore, the model’s ability to identify good overall survival outcome patients when given neoadjuvant chemotherapy illustrates its potential utility for informing treatment escalation/de-escalation decisions, which could be clinically valuable if further validated.
Support: U10CA180821, U10CA180882; U24CA176171; U10CA180888 (SWOG); P50-CA058223; Genentech; https://acknowledgments.alliancefound.org. Clinicaltrials.gov Id#: NCT00861705
Background: Heterogeneity within triple-negative breast cancer (TNBC) makes it challenging to target therapeutically. Despite recent advances in the standard of care treatment of stage II-III TNBC, which now includes multiple chemotherapeutic agents plus an immune checkpoint inhibitor, the cure rate for these patients remains suboptimal. There is still a critical need for improved molecular characterization of TNBC and the development of accurate prognostic tools to assist in making treatment escalation and de-escalation decisions. Here, we present an analysis of the DNA and RNA landscape in a set of patients treated on a clinical trial and validation of molecular and prognostic features using three additional data sets.
Methods: CALGB 40603 (Alliance) is a randomized phase II trial that evaluated the impact of adding carboplatin and/or bevacizumab to standard anthracycline/taxane neoadjuvant chemotherapy in stage II-III TNBC. Targeted panel DNA sequencing of pre-treatment tumor samples from 238 patients enrolled in this study was performed and analyzed alongside matched pre-treatment RNAseq data. These results were then combined with clinical outcomes data to develop a prognostic multi-variable elastic net model of overall survival. We subsequently evaluated this model on samples from 380 stage II-III TNBC patients from three independent publicly available datasets.
Results: Similar to other TNBC data sets, DNAseq results of 40603 identified TP53 (86%) as the most frequent somatically mutated gene, followed by MT-ND5 (16%), MT-ND4 (12%), CSMD3 (8%), and PIK3CA (7%). BRCA1 (germline 8%, somatic 4%), BRCA2 (germline 2%, somatic 0.4%), and PALB2 (germline 1%) gene mutations were also detected, and when combined into a single “homologous recombination deficiency” category totaled 15%. The DNA copy number landscape of TNBC was very similar across the four data sets, and largely mirrored the landscape seen for basal-like breast cancer characterized in TCGA. Importantly, only one somatically mutated gene (PIK3R2), and no DNA copy number altered regions, were associated with pathologic complete response or survival at a false discovery rate < 0.05. We next sought to use these multi-platform data to train a prognostic Cox elastic net regression model using CALGB 40603 data and a diverse set of features, including tumor stage, 804 RNA expression signatures, 727 genes with somatic mutations, and 534 chromosomal segment-level copy number alterations. This yielded a 31-feature prognostic model with C-index performance values (95% confidence interval) of 0.68 (0.66-0.69), 0.61 (0.59-0.63), and 0.83 (0.82-0.84) using FUSCC (PMID 30853353), METABRIC (PMID 22522925), and TCGA (PMID 23000897) test sets. The most heavily weighted features in this model were tumor stage and multiple immune signatures including an IgG signature, and this model had higher C-index values than a model trained using only tumor stage.
Conclusions: This study provides a comprehensive and integrated characterization of the DNA- and RNA-based landscape of a large stage II-III TNBC patient data set. Through the development of a multi-omic elastic net model of TNBC survival, we show that we may improve the prognostic accuracy for overall survival beyond stage alone by incorporating machine learning selected molecular features. Furthermore, the model’s ability to identify good overall survival outcome patients when given neoadjuvant chemotherapy illustrates its potential utility for informing treatment escalation/de-escalation decisions, which could be clinically valuable if further validated.
Support: U10CA180821, U10CA180882; U24CA176171; U10CA180888 (SWOG); P50-CA058223; Genentech; https://acknowledgments.alliancefound.org. Clinicaltrials.gov Id#: NCT00861705
Methods: CALGB 40603 (Alliance) is a randomized phase II trial that evaluated the impact of adding carboplatin and/or bevacizumab to standard anthracycline/taxane neoadjuvant chemotherapy in stage II-III TNBC. Targeted panel DNA sequencing of pre-treatment tumor samples from 238 patients enrolled in this study was performed and analyzed alongside matched pre-treatment RNAseq data. These results were then combined with clinical outcomes data to develop a prognostic multi-variable elastic net model of overall survival. We subsequently evaluated this model on samples from 380 stage II-III TNBC patients from three independent publicly available datasets.
Results: Similar to other TNBC data sets, DNAseq results of 40603 identified TP53 (86%) as the most frequent somatically mutated gene, followed by MT-ND5 (16%), MT-ND4 (12%), CSMD3 (8%), and PIK3CA (7%). BRCA1 (germline 8%, somatic 4%), BRCA2 (germline 2%, somatic 0.4%), and PALB2 (germline 1%) gene mutations were also detected, and when combined into a single “homologous recombination deficiency” category totaled 15%. The DNA copy number landscape of TNBC was very similar across the four data sets, and largely mirrored the landscape seen for basal-like breast cancer characterized in TCGA. Importantly, only one somatically mutated gene (PIK3R2), and no DNA copy number altered regions, were associated with pathologic complete response or survival at a false discovery rate < 0.05. We next sought to use these multi-platform data to train a prognostic Cox elastic net regression model using CALGB 40603 data and a diverse set of features, including tumor stage, 804 RNA expression signatures, 727 genes with somatic mutations, and 534 chromosomal segment-level copy number alterations. This yielded a 31-feature prognostic model with C-index performance values (95% confidence interval) of 0.68 (0.66-0.69), 0.61 (0.59-0.63), and 0.83 (0.82-0.84) using FUSCC (PMID 30853353), METABRIC (PMID 22522925), and TCGA (PMID 23000897) test sets. The most heavily weighted features in this model were tumor stage and multiple immune signatures including an IgG signature, and this model had higher C-index values than a model trained using only tumor stage.
Conclusions: This study provides a comprehensive and integrated characterization of the DNA- and RNA-based landscape of a large stage II-III TNBC patient data set. Through the development of a multi-omic elastic net model of TNBC survival, we show that we may improve the prognostic accuracy for overall survival beyond stage alone by incorporating machine learning selected molecular features. Furthermore, the model’s ability to identify good overall survival outcome patients when given neoadjuvant chemotherapy illustrates its potential utility for informing treatment escalation/de-escalation decisions, which could be clinically valuable if further validated.
Support: U10CA180821, U10CA180882; U24CA176171; U10CA180888 (SWOG); P50-CA058223; Genentech; https://acknowledgments.alliancefound.org. Clinicaltrials.gov Id#: NCT00861705
{{ help_message }}
{{filter}}