Gene Expression, CTCs and Metastatic Breast Cancer Prognosis
Gene Expression, CTCs and Metastatic Breast Cancer Prognosis
Between February 2008 and January 2011, 130 MBC patients were included. CTC counts at baseline and FU were available for 130 and for 82 patients, respectively. Sixteen and 11 patients were excluded because of insufficient mRNA quantity and quality (QQ) and insufficient FU (i.e. <9 months and no treatment failure documented in that time), respectively, leaving 103 patients forming the first discovery set (supplementary Figure S1, available at Annals of Oncology online http://annonc.oxfordjournals.org/content/26/3/510/suppl/DC1). Between January 2011 and August 2012, another 78 patients with sufficient FU data were included under the same in- and exclusion criteria, forming the validation set. One patient was excluded because of insufficient QQ mRNA. Among these 77, the 11 patients excluded from the discovery set because of insufficient FU. Characteristics of all 197 included patients are depicted in supplementary Table S1, available at Annals of Oncology online http://annonc.oxfordjournals.org/content/26/3/510/suppl/DC1, and were similar between discovery and validation sets. However, when comparing the initial discovery cohort of 130 patients with the 103 QQ mRNA patients, the number of patients that had received adjuvant endocrine treatment was lower in the latter (Fisher's exact P 0.016).
TTF was chosen as the primary end point to reflect the heterogeneity in clinical decision making in MBC patients. TTF reliably reflects the benefit a patient derives from a treatment, as it captures the time gained by administering that treatment. Of note, apart from two patients who switched therapies because of toxicity, all patients started second-line treatment after disease progression. However, in contrast to PFS, TTF had not previously been correlated with CTC count. Therefore, we first verified whether CTC counts were associated with TTF in our cohort. Indeed, the 63 patients with ≥5 CTCs at baseline had a shorter TTF than the 67 patients with <5 CTCs [median TTF 10.2 versus 20 months, HR 2.92 (95% CI 1.71–4.95) P < 0.0001]. After 2–5 weeks of therapy, CTCs were again enumerated in 82 patients. Patients with ≥5 CTCs at FU had shorter TTF [9.2 versus 16.1 months, HR 2.83 (95% CI 1.39–5.76) P 0.004] than patients with <5 CTCs. Looking at change in CTC count during therapy, all patients with ≥5 CTCs at FU, regardless of CTC count at baseline, had a shorter TTF than patients with persistently low CTC counts [HR 3.83 (95% CI 1.71–8.86) P 0.001]. No statistically significant difference was seen between the TTF of patients with persistently low CTC counts versus those with a decline to <5 CTCs after a high baseline count (P 0.066, for Kaplan–Meier curves, see supplementary Figure S2, available at Annals of Oncology online http://annonc.oxfordjournals.org/content/26/3/510/suppl/DC1). In univariate analysis, besides CTC count, the presence of visceral metastases and a higher number of metastases were associated with shorter TTF. In multivariate analysis, only baseline CTC count was independently associated with TTF [HR 2.54 (95% CI 1.45–4.46) P 0.001 supplementary Table S2, available at Annals of Oncology online http://annonc.oxfordjournals.org/content/26/3/510/suppl/DC1]. CTC count at baseline [HR 2.44 (95% CI 1.27–4.69) P 0.007], at FU [HR 2.77 (95% CI 1.18–6.52) P 0.019], and CTC count change during therapy [HR 3.26 (95% CI 1.23–8.61) P 0.017] were also associated with OS.
To establish the prognostic value of CTC gene expression, we chose to base our analysis on the 55 mRNA genes that were previously established to be CTC-specific. While this 55-gene panel is based on its expression in patients with ≥5 CTCs, cell line spiking experiments showed the ability of this panel to detect epithelial signal in as little as one tumor cell spiked into 7.5 ml blood. We were therefore confident that this CTC-specific panel would be able to pick up CTC gene expression, if indeed present, in patients with <5 counted CTCs, and included all 103 patients in our discovery set, regardless of CTC count, for the subsequent analyses.
Of the 103 patients, 42 patients were classified as poor prognosis and 61 patients as good prognosis based on a cutoff at 9 months TTF. This cutoff, chosen based on the median PFS in first-line MBC patients, was deemed valid as the median TTF in our cohort was 8.9 months (95% CI 7.3–10.2).
A predictor was built based on the 55 CTC-specific genes. In univariate analysis, nine genes were at a P < 0.05 (supplementary Table S3, available at Annals of Oncology online http://annonc.oxfordjournals.org/content/26/3/510/suppl/DC1) and 16 genes at a P < 0.1 differentially expressed between the good and poor prognosis group. LOOCV was carried out with these latter 16 genes, and a CCP was calculated for each sample, for which the receiver operating characteristic (ROC) curve is depicted in Figure 1A. At an area under the curve (AUC) of 0.69 (95% CI 0.59–0.80, P 0.0001), the 16-gene CTC profile carried out similarly to the CTC count [AUC 0.62 (95% CI 0.51–0.73) P 0.0145] (Figure 1B). Because our primary interest was correctly predicting patients with early therapy failure, we aimed for our CTC profile to identify poor prognosis patients with 90% sensitivity. At this cutoff, 76 patients had an unfavorable profile, half of whom failed treatment before 9 months (positive predictive value, PPV 50%). Twenty-seven patients had a favorable profile, of whom 23 indeed experienced no treatment failure (negative predictive value, NPV 85%). Test characteristics of both the 16-gene CTC profile and count are depicted in Figure 1C and D.
(Enlarge Image)
Figure 1.
Receiver operating characteristic (ROC) curves (A and B), test performance (C) and test characteristics (D) of CTC count and 16-gene CTC profile in 103 patients. AUC, area under the curve; PPV, positive predictive value; NPV, negative predictive value.
The Kaplan–Meier curves for the 16-gene CTC profile, CTC count and for the combination are shown in Figure 2. Panel A shows that a CTC count ≥5 identifies poor prognosis patients among the 103 in whom the CTC profile was generated (log-rank P < 0.001). In Figure 2B, an early and clear distinction into a poor and good prognosis group is seen when separating patients according to the CTC profile (log-rank P < 0.001). The added value of the profile appears to lie mainly in its ability to further classify patients with <5 CTCs (Figure 2C and D, log-rank for trend P < 0.001), while this CTC profile does not identify prognostic groups among patients with ≥5 CTCs.
(Enlarge Image)
Figure 2.
Kaplan–Meier plots for patient subgroups as defined by CTC count (A), the 16-gene CTC profile (B), and the combination of CTC count and 16-gene CTC profile (C). Only one patient had more than five CTCs and a favorable 16-gene CTC profile; therefore, no curve is depicted for this subgroup. Panel D combines panels A and C and shows that the 16-gene CTC panel is able to distinguish a truly good (dashed green line) and an intermediate prognosis group (dotted green line) among patients with <5 CTCs (solid green line), while no added value is seen in patients with ≥5 CTCs (blue lines; no line depicted for patients with ≥5 CTCs and a favorable profile as there was only one such patient). u.p, unfavorable profile; f.p., favorable profile.
In univariate analysis, the 16-gene CTC profile was significantly associated with TTF [HR 4.57 (95% CI 2.20–9.50) P < 0.0001, Table 1], as were the number of metastases [HR 1.39 (95% CI 1.13–1.72) P 0.002], presence of visceral metastases [HR 1.84 (95% CI 1.05–3.23) P 0.035] and CTC count at baseline [HR 3.00 (95% CI 1.73–5.19) P < 0.001]. Other known prognostic factors such as triple-negative status were not associated with TTF. Among patients with <5 CTCs, those with an unfavorable CTC profile had a shorter TTF [HR 4.23 (95% CI 1.57–11.42) P 0.004] than those with a favorable profile. Among patients with ≥5 CTCs, only one patient had an unfavorable CTC profile, so the CTC profile did not have added value. In multivariate analysis including all these prognostic factors, only the 16-gene CTC profile was an independent predictor of TTF [HR 3.15 (95% CI 1.35–7.33) P 0.008].
Having identified a highly prognostic 16-gene CTC profile, its external validity was assessed in 77 patients who were included under the same study protocol, but whose FU completed at a later time point. All known prognostic factors were similar between the original discovery set and the validation set (supplementary Table S1, available at Annals of Oncology online http://annonc.oxfordjournals.org/content/26/3/510/suppl/DC1).
In the validation set, the 16-gene profile correctly identified 18 of 22 patients with early treatment failure (82% sensitivity). Specificity was lower at 20% (Figure 3). The 16-gene profile predicted 62 patients to have a poor prognosis, of whom 18 experienced early treatment failure (PPV 29%). Eleven of 15 patients were correctly predicted to have a good prognosis (NPV 73%).
(Enlarge Image)
Figure 3.
Test performance (A) and test characteristics (C) of the 16-gene and the 8-gene CTC profile in the validation sets consisting of 77 and 90 patients, respectively. The ROC curve for the 8-gene panel in the discovery set is depicted in (B). AUC, area under the curve; PPV, positive predictive value; NPV, negative predictive value.
To explain the lack of statistically significant validation of our 16-gene CTC profile in the independent patient set, we hypothesized that—while the discovery and validation set were balanced for known prognostic factors—additional unknown prognostic factors could have interfered with the validation. To overcome this potential problem, we randomly divided the total group of 180 patients into two groups of 90. These two groups too were well balanced for known prognostic factors (supplementary Table S1, available at Annals of Oncology online http://annonc.oxfordjournals.org/content/26/3/510/suppl/DC1). Using LOOCV, a new 8-gene CTC profile was generated (supplementary Table S3, available at Annals of Oncology online). At an AUC of 0.77 [(95% CI 0.67–0.87) P < 0.0001], the 8-gene CTC profile carried out similarly to the 16-gene profile. This 8-gene profile was then validated in the other 90 patients. Sixty-seven patients were predicted to have a poor prognosis, 28 of whom failed treatment before 9 months (PPV 42%). Twenty-three patients had a predicted favorable prognosis, of whom 18 experienced no or late treatment failure (NPV 78% P 0.0851).
Results
Patient Characteristics
Between February 2008 and January 2011, 130 MBC patients were included. CTC counts at baseline and FU were available for 130 and for 82 patients, respectively. Sixteen and 11 patients were excluded because of insufficient mRNA quantity and quality (QQ) and insufficient FU (i.e. <9 months and no treatment failure documented in that time), respectively, leaving 103 patients forming the first discovery set (supplementary Figure S1, available at Annals of Oncology online http://annonc.oxfordjournals.org/content/26/3/510/suppl/DC1). Between January 2011 and August 2012, another 78 patients with sufficient FU data were included under the same in- and exclusion criteria, forming the validation set. One patient was excluded because of insufficient QQ mRNA. Among these 77, the 11 patients excluded from the discovery set because of insufficient FU. Characteristics of all 197 included patients are depicted in supplementary Table S1, available at Annals of Oncology online http://annonc.oxfordjournals.org/content/26/3/510/suppl/DC1, and were similar between discovery and validation sets. However, when comparing the initial discovery cohort of 130 patients with the 103 QQ mRNA patients, the number of patients that had received adjuvant endocrine treatment was lower in the latter (Fisher's exact P 0.016).
CTC Count Predicts TTF and OS
TTF was chosen as the primary end point to reflect the heterogeneity in clinical decision making in MBC patients. TTF reliably reflects the benefit a patient derives from a treatment, as it captures the time gained by administering that treatment. Of note, apart from two patients who switched therapies because of toxicity, all patients started second-line treatment after disease progression. However, in contrast to PFS, TTF had not previously been correlated with CTC count. Therefore, we first verified whether CTC counts were associated with TTF in our cohort. Indeed, the 63 patients with ≥5 CTCs at baseline had a shorter TTF than the 67 patients with <5 CTCs [median TTF 10.2 versus 20 months, HR 2.92 (95% CI 1.71–4.95) P < 0.0001]. After 2–5 weeks of therapy, CTCs were again enumerated in 82 patients. Patients with ≥5 CTCs at FU had shorter TTF [9.2 versus 16.1 months, HR 2.83 (95% CI 1.39–5.76) P 0.004] than patients with <5 CTCs. Looking at change in CTC count during therapy, all patients with ≥5 CTCs at FU, regardless of CTC count at baseline, had a shorter TTF than patients with persistently low CTC counts [HR 3.83 (95% CI 1.71–8.86) P 0.001]. No statistically significant difference was seen between the TTF of patients with persistently low CTC counts versus those with a decline to <5 CTCs after a high baseline count (P 0.066, for Kaplan–Meier curves, see supplementary Figure S2, available at Annals of Oncology online http://annonc.oxfordjournals.org/content/26/3/510/suppl/DC1). In univariate analysis, besides CTC count, the presence of visceral metastases and a higher number of metastases were associated with shorter TTF. In multivariate analysis, only baseline CTC count was independently associated with TTF [HR 2.54 (95% CI 1.45–4.46) P 0.001 supplementary Table S2, available at Annals of Oncology online http://annonc.oxfordjournals.org/content/26/3/510/suppl/DC1]. CTC count at baseline [HR 2.44 (95% CI 1.27–4.69) P 0.007], at FU [HR 2.77 (95% CI 1.18–6.52) P 0.019], and CTC count change during therapy [HR 3.26 (95% CI 1.23–8.61) P 0.017] were also associated with OS.
CTC Gene Expression
To establish the prognostic value of CTC gene expression, we chose to base our analysis on the 55 mRNA genes that were previously established to be CTC-specific. While this 55-gene panel is based on its expression in patients with ≥5 CTCs, cell line spiking experiments showed the ability of this panel to detect epithelial signal in as little as one tumor cell spiked into 7.5 ml blood. We were therefore confident that this CTC-specific panel would be able to pick up CTC gene expression, if indeed present, in patients with <5 counted CTCs, and included all 103 patients in our discovery set, regardless of CTC count, for the subsequent analyses.
16-gene CTC Profile Predicts for TTF
Of the 103 patients, 42 patients were classified as poor prognosis and 61 patients as good prognosis based on a cutoff at 9 months TTF. This cutoff, chosen based on the median PFS in first-line MBC patients, was deemed valid as the median TTF in our cohort was 8.9 months (95% CI 7.3–10.2).
A predictor was built based on the 55 CTC-specific genes. In univariate analysis, nine genes were at a P < 0.05 (supplementary Table S3, available at Annals of Oncology online http://annonc.oxfordjournals.org/content/26/3/510/suppl/DC1) and 16 genes at a P < 0.1 differentially expressed between the good and poor prognosis group. LOOCV was carried out with these latter 16 genes, and a CCP was calculated for each sample, for which the receiver operating characteristic (ROC) curve is depicted in Figure 1A. At an area under the curve (AUC) of 0.69 (95% CI 0.59–0.80, P 0.0001), the 16-gene CTC profile carried out similarly to the CTC count [AUC 0.62 (95% CI 0.51–0.73) P 0.0145] (Figure 1B). Because our primary interest was correctly predicting patients with early therapy failure, we aimed for our CTC profile to identify poor prognosis patients with 90% sensitivity. At this cutoff, 76 patients had an unfavorable profile, half of whom failed treatment before 9 months (positive predictive value, PPV 50%). Twenty-seven patients had a favorable profile, of whom 23 indeed experienced no treatment failure (negative predictive value, NPV 85%). Test characteristics of both the 16-gene CTC profile and count are depicted in Figure 1C and D.
(Enlarge Image)
Figure 1.
Receiver operating characteristic (ROC) curves (A and B), test performance (C) and test characteristics (D) of CTC count and 16-gene CTC profile in 103 patients. AUC, area under the curve; PPV, positive predictive value; NPV, negative predictive value.
The Kaplan–Meier curves for the 16-gene CTC profile, CTC count and for the combination are shown in Figure 2. Panel A shows that a CTC count ≥5 identifies poor prognosis patients among the 103 in whom the CTC profile was generated (log-rank P < 0.001). In Figure 2B, an early and clear distinction into a poor and good prognosis group is seen when separating patients according to the CTC profile (log-rank P < 0.001). The added value of the profile appears to lie mainly in its ability to further classify patients with <5 CTCs (Figure 2C and D, log-rank for trend P < 0.001), while this CTC profile does not identify prognostic groups among patients with ≥5 CTCs.
(Enlarge Image)
Figure 2.
Kaplan–Meier plots for patient subgroups as defined by CTC count (A), the 16-gene CTC profile (B), and the combination of CTC count and 16-gene CTC profile (C). Only one patient had more than five CTCs and a favorable 16-gene CTC profile; therefore, no curve is depicted for this subgroup. Panel D combines panels A and C and shows that the 16-gene CTC panel is able to distinguish a truly good (dashed green line) and an intermediate prognosis group (dotted green line) among patients with <5 CTCs (solid green line), while no added value is seen in patients with ≥5 CTCs (blue lines; no line depicted for patients with ≥5 CTCs and a favorable profile as there was only one such patient). u.p, unfavorable profile; f.p., favorable profile.
In univariate analysis, the 16-gene CTC profile was significantly associated with TTF [HR 4.57 (95% CI 2.20–9.50) P < 0.0001, Table 1], as were the number of metastases [HR 1.39 (95% CI 1.13–1.72) P 0.002], presence of visceral metastases [HR 1.84 (95% CI 1.05–3.23) P 0.035] and CTC count at baseline [HR 3.00 (95% CI 1.73–5.19) P < 0.001]. Other known prognostic factors such as triple-negative status were not associated with TTF. Among patients with <5 CTCs, those with an unfavorable CTC profile had a shorter TTF [HR 4.23 (95% CI 1.57–11.42) P 0.004] than those with a favorable profile. Among patients with ≥5 CTCs, only one patient had an unfavorable CTC profile, so the CTC profile did not have added value. In multivariate analysis including all these prognostic factors, only the 16-gene CTC profile was an independent predictor of TTF [HR 3.15 (95% CI 1.35–7.33) P 0.008].
Validation of the 16-gene Profile
Having identified a highly prognostic 16-gene CTC profile, its external validity was assessed in 77 patients who were included under the same study protocol, but whose FU completed at a later time point. All known prognostic factors were similar between the original discovery set and the validation set (supplementary Table S1, available at Annals of Oncology online http://annonc.oxfordjournals.org/content/26/3/510/suppl/DC1).
In the validation set, the 16-gene profile correctly identified 18 of 22 patients with early treatment failure (82% sensitivity). Specificity was lower at 20% (Figure 3). The 16-gene profile predicted 62 patients to have a poor prognosis, of whom 18 experienced early treatment failure (PPV 29%). Eleven of 15 patients were correctly predicted to have a good prognosis (NPV 73%).
(Enlarge Image)
Figure 3.
Test performance (A) and test characteristics (C) of the 16-gene and the 8-gene CTC profile in the validation sets consisting of 77 and 90 patients, respectively. The ROC curve for the 8-gene panel in the discovery set is depicted in (B). AUC, area under the curve; PPV, positive predictive value; NPV, negative predictive value.
To explain the lack of statistically significant validation of our 16-gene CTC profile in the independent patient set, we hypothesized that—while the discovery and validation set were balanced for known prognostic factors—additional unknown prognostic factors could have interfered with the validation. To overcome this potential problem, we randomly divided the total group of 180 patients into two groups of 90. These two groups too were well balanced for known prognostic factors (supplementary Table S1, available at Annals of Oncology online http://annonc.oxfordjournals.org/content/26/3/510/suppl/DC1). Using LOOCV, a new 8-gene CTC profile was generated (supplementary Table S3, available at Annals of Oncology online). At an AUC of 0.77 [(95% CI 0.67–0.87) P < 0.0001], the 8-gene CTC profile carried out similarly to the 16-gene profile. This 8-gene profile was then validated in the other 90 patients. Sixty-seven patients were predicted to have a poor prognosis, 28 of whom failed treatment before 9 months (PPV 42%). Twenty-three patients had a predicted favorable prognosis, of whom 18 experienced no or late treatment failure (NPV 78% P 0.0851).
Source...