Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2020 Nov 1.
Published in final edited form as: Lancet Oncol. 2019 Sep 30;20(11):1576–1586. doi: 10.1016/S1470-2045(19)30503-0

Time-to-progression after frontline chemoimmunotherapy for chronic lymphocytic leukemia: an independently validated retrospective cohort study

Carmen D Herling 1,*, Kevin R Coombes 2,*, Axel Benner 3, Johannes Bloehdorn 4, Lynn L Barron 5, Zachary B Abrams 6, Tadeusz Majewski 7, Jolanta E Bondaruk 8, Jasmin Bahlo 9, Kirsten Fischer 10, Michael Hallek 11, Stephan Stilgenbauer 12, Bogdan A Czerniak 13, Christopher Oakes 14, Alessandra Ferrajoli 15, Michael J Keating 16, Lynne V Abruzzo 17
PMCID: PMC7147008  NIHMSID: NIHMS1545880  PMID: 31582354

Abstract

Background

Because frontline treatment with FCR unequivocally improves progression-free and overall survival in CLL patients, it has become a “gold standard” for chemoimmunotherapy regimens. Follow-up studies show that a subset of generally young, physically-fit patients with mutated IGHV genes and low-risk cytogenetic abnormalities achieve long-term remissions by FCR. However, the question remains how to treat treatment-naïve CLL patients with unmutated IGHV-unmutated CLL patients, particularly those who desire a short therapy course, or who cannot tolerate tyrosine kinase inhibitor therapy.

Methods

We performed transcriptional profiling on samples obtained from 101 treatment-naïve CLL patients at the M.D. Anderson Cancer Center using Illumina HumanHT-12 v4 BeadChips. To identify genes associated with time-to-progression, we performed univariate Cox proportional hazards analyses, and used these genes to cluster IGHV-unmutated samples into two groups. After using cross-validation to assess robustness, we applied the Lasso to standardized gene expression values in order to find a minimum-gene signature. We validated the signature against 109 clinically-similar IGHV-unmutated cases from the German CLL8 trial profiled on Affymetrix microarrays.

Findings

We developed a 17-gene expression signature that distinguished treatment-naïve IGHV-unmutated CLL patients who are likely to achieve a long-term remission following frontline FCR chemoimmunotherapy from those who may benefit from alternative front-line regimens (HR=3.83; 95%CI=1.94–7.59, p=0.00004). We validated this signature on patients enrolled in the CLL8 study, finding that patients classified as high-risk had a hazard ratio of 1.90 (95%CI=1.18–3.06, p=0.006; median progression 39 months) compared to low-risk patients (median progression 59 months).

Interpretation

We have developed a robust, reproducible 17-gene signature that identifies a subset of treatment-naïve IGHV-unmutated CLL patients who may substantially benefit from treatment with FCR chemoimmunotherapy. We would recommend testing the value of the 17-gene signature in a prospective study that compares treatment with FCR to newer alternative therapies as part of a randomized clinical trial.

INTRODUCTION

Chronic lymphocytic leukemia (CLL) has a variable clinical course. Over the past decade, two clinical trials, the FCR300 study in the United States and the CLL8 study in Germany, demonstrated unequivocally that chemoimmunotherapy with fludarabine, cyclophosphamide, and rituximab (FCR) is highly efficacious front-line therapy for chronic lymphocytic leukemia (CLL).1,2 FCR was the first regimen to improve progression-free survival (PFS) and overall survival (OS) in CLL, and has become a “gold standard” chemoimmunotherapy regimen in physically-fit patients. Since then, follow-up studies have shown that a subset of patients achieve long-term durable remissions.3–5 In general, these patients are relatively young, under 65 years of age, and fit, without co-morbidities. Their CLL cells are usually IGHV-mutated and lack high-risk cytogenetic abnormalities, del(11)(q22–23) and del(17)(p13.1). These findings raise the possibility that FCR chemoimmunotherapy effects durable remissions in a carefully selected subset of patients.

Our aim was to develop and independently validate a gene expression signature to identify patients likely to achieve durable remissions with FCR chemoimmunotherapy. We hypothesized that inherent differences in the expression of protein-coding genes in the CLL cells of treatment-naïve patients determine the response duration following FCR chemoimmunotherapy. Based on a cohort of 101 treatment-naïve CLL patients at the University of Texas M.D. Anderson Cancer Center, we developed a 17-gene expression signature that distinguishes treatment-naïve IGHV-unmutated CLL patients who are likely to achieve a long-term remissions following front-line FCR chemoimmunotherapy from those who may benefit from alternative regimens. We validated this signature on a clinically-similar cohort of patients enrolled as part of the German CLL8 study, which used a different gene-expression profiling platform. Thus, we have developed a robust, reproducible 17-gene signature that identifies treatment-naïve IGHV-unmutated CLL patients who may substantially benefit from treatment with FCR chemoimmunotherapy.

METHODS

Sample collection

Our goal was to discover and validate gene expression signatures of time-to-progression (TTP) in CLL patients treated with frontline FCR therapy. TTP is defined as the time from the start of therapy until progression; death is excluded. In our cohort, the cause of death was often not disease-related or was unknown. Inclusion criteria were: (i) disease fulfilled diagnostic iwCLL criteria, (ii) patient received at least 3 cycles of FCR, (iii) availability of clinical follow-up, (iv) availability of peripheral blood (PB) samples acquired within six months of start of therapy, and (v) informed consent for gene expression profiling (GEP). Patients were excluded if they failed any inclusion criteria or the RNA quality was insufficient (RIN≤7) (Appendix pp.1–3, Table ST1). The studies were approved by the Institutional Review Boards and conducted according to the principles expressed in the Declaration of Helsinki. A CONSORT diagram is provided (Figure 1). The discovery/training cohort consisted of PB samples collected from 101 treatment-naïve CLL patients at the University of Texas M.D. Anderson Cancer (MDACC) and processed as described previously.6–8 All samples were obtained within 6 months before the start of first treatment and met CLL diagnostic criteria.1 Clinical and routine laboratory data were obtained from the medical records. The immunoglobulin heavy chain variable region (IGHV) gene somatic mutation status and ZAP70 expression, measured by either flow cytometry or immunohistochemistry, were assessed on blood or bone marrow samples (Appendix p.4).8–10 Common CLL-associated abnormalities (del(11)(q22.3) including ATM; del(13)(q14.3) including the DLEU2/mir-15a/16–1 cluster; del(17)(p13.1) including TP53; trisomy 12), were assessed by array-based SNP genotyping,6,8 and grouped according to the Döhner hierarchy.11 The validation cohort was 109 treatment-naïve patients enrolled on the CLL8 trial of the German CLL Study Group (GCLLSG); PB was collected at the University of Ulm just before study entry. Complete clinical data and prognostic markers for this cohort have been published previously.1,3,12

Figure 1:

Figure 1:

Open in a new tab

CONSORT diagram for the MDACC discovery cohort and the CLL8 validation cohort.

Gene expression profiling

We extracted total RNA from CD19-positive PB B cells enriched using immunomagnetic beads.8 For the MDACC cohort, transcriptional profiling was performed using Illumina HumanHT-12 v4 Expression BeadChips (San Diego, CA), according to the manufacturer’s instructions. Data were background corrected, normalized, and log-transformed using the lumi R package, version 2.20.1.13 Standard exploratory plots, hierarchical clustering, and principal component analyses (PCA) were prepared to assess the intrinsic quality of individual arrays and exclude technological or processing artifacts. To account for multiple testing, we bounded the false discovery rate (FDR), which was estimated using a beta-uniform mixture model.14 For the CLL8 cohort, transcriptional profiling using Affymetrix GeneChip® Human Exon 1.0 ST Array (Santa Clara, CA, USA) and expression data processing were performed as described previously.15

Identification of genes with prognostic value for clinical outcome

For both cohorts, response to treatment and progression were classified according to the National Cancer Institute working group criteria.16 Using the R survival package (version 2.14–3), we fit univariate per-gene Cox proportional hazards models to find genes that were statistically associated with TTP as a continuous variable. Genes with a p-value <0.0085 (FDR=33%) were used in hierarchical clustering with Pearson correlation and Ward’s linkage to define subtypes based on GEP. The GEP/TTP subtypes were verified by PCA and investigated for interactions with known CLL prognostic markers (Fisher’s exact/chi-square testing; Appendix pp.5–6, Table ST2). The GEP/TTP subtypes were included as variables in multivariable Cox proportional hazards models competing with classical predictors of clinical prognosis. We used the Akaike Information Criterion (AIC) to stepwise add or subtract variables and optimize models with the strongest prognostic value.17 We thoroughly cross-validated the clustering into subtypes and the resulting models (Appendix pp.7–10, Figures SF1–2). To identify samples with a high posterior probability of being reliably assigned to subtypes, we applied linear discriminant analysis (LDA) using the five most significant principal components as predictors. Using those reliable samples, we performed preliminary feature selection based on gene-by-gene t-tests (with p<0.0026, FDR=15%) between the groups. We then applied the “Lasso” method to find a minimum-gene signature.18

Independent validation of the model

After using the Lasso to finalize a model on the discovery cohort, details of the algorithm were sent to the biostatistician (AB) on the CLL8 trial. Although progression-free-survival (PFS) was the primary outcome in previous reports on the CLL8 cohort, these data were reanalyzed to compute TTP. Using samples from 109 IGHV-unmutated FCR-treated CLL8 patients, data from each of the 17 genes in the final prognostic model were standardized to have mean zero and variance one. A continuous score was computed as the linear predictor of the standardized gene expression levels using the coefficients from the prognostic model as weights. This score was dichotomized into a binary predictor of low risk (≤0) and high risk (>0) categories. We evaluated the model’s performance on the independent validation set using a proportional cause-specific hazards model of TTP, using either the continuous score or the binary predictor. Cumulative incidence of TTP was estimated using the Aalen-Johansen estimator. Prediction error curves were used to assess the predictive accuracy of the models,19 where prediction error is defined using Brier’s score as a function of time.20 Clinical models whose variables and coefficients were learned from the MDACC cohort were also validated on the CLL8 cohort.

Statistical methods

All statistical computations were performed using version 3.4.0 of the R statistical programming environment, including version 2.41–3 of the survival package, version 2.0–10 of glmnet, version 3.1.4 of ClassComparison, and version 3.3.5 of ClassDiscovery. The significance of time-to-event models was assessed using the p-values from a score (logrank) test based on Cox proportional hazards models.

Gene network, functional annotation and pathway evaluation

Network construction, canonical pathway analyses, and functional annotations were performed using Ingenuity Pathway Analysis (IPA; Ingenuity Systems, Redwood City, CA, USA; www.ingenuity.com), ToppGene (https://toppgene.cchmc.org/), and DAVID 6.8 (http://david.abcc.ncifcrf.gov/). Separate gene files for each of the identified gene clusters (cyan, red, purple, green) were analyzed using the default settings for each tool. Gene sets were evaluated by a modified Fisher’s exact test with a Benjamini-Hochberg correction for multiple testing.

Role of the Funding Source

The study sponsors had no role in the study design, collection, analysis, data interpretation, or manuscript preparation, nor in the decision to submit the manuscript for publication. The corresponding author had full access to all data and accepts final responsibility for the decision to submit this manuscript for publication.

RESULTS

The MDACC discovery cohort consisted of 101 CLL patients; survivors were followed for a median of 146.5 months (quartiles= 137.7–162.3) after treatment (Table 1). Most patients were at low or intermediate clinical stage (Rai stage 0–2) at sample collection time. Established markers of CLL prognosis, including IGHV somatic mutation status, ZAP70 expression, and serum beta-2-microglobulin (β2M) levels, were represented at expected frequencies (Table 1). All patients received at least 3 cycles of frontline therapy, between 2000–10-02 and 2006–10-26, with either standard FCR (n=61)2,21 or rituximab-intensified FCR (3 instead of one rituximab dose(s) per cycle) (n=40).22 Most patients (n=79 (78.2%) of 101) completed 6 cycles of chemoimmunotherapy and responded to treatment (n=98 (97%) of 101, with 84 (83%) complete remissions, 14 (14%) partial remissions, 2 (2%) no response, and 1 (1%) unknown). At final follow up, 49 (48.5%) of 101 patients progressed, 42 (41.6%) died, and 63 (62.4%) either progressed or died. The median overall survival (OS) from the start of FCR therapy was not reached at the end of follow-up. The median TTP was 92.4 months (quartiles=44.1–infinity) and the median PFS was 66.6 months (quartiles=40.5– infinity).

Table 1.

Patient characteristics

MDACC GCLLSG/CLL8
All (n = 101) Unmutated (n = 66) Unmutated (n = 109)
Median age at diagnosis in years (quartiles) 56 (49–62) 56 (49–63) 60 (53–65)
Gender Male 76 (75.2%) 52 (78.8%) 86 (78.9%)
Female 25 (24.8%) 14 (21.2%) 23 (21.1%)
Rai stage 0–2 77 (76.2%) 52 (78.8%) 63 (69.2%)
≥3 24 (23.8%) 14 (21.2%) 28 (30.8%)
Not available 0 0 18
IGHV status Mutated 35 (34.7%) 0 (0%) 0 (0%)
Unmutated 66 (65.3%) 66 (100%) 109 (100%)
β2M ≤4 mg/L 68 (67.3%) 42 (63.6%) 85 (78.7%)
>4 mg/L 33 (32.7%) 24 (36.4%) 23 (21.3%)
Not available 0 0 1
WBC ≤150×109/L 83 (82.2%) 50 (75.8%) 84 (77.1%)
>150×109/L 18 (17.8%) 16 (24.2%) 25 (22.9%)
Not available 0 0 1
ZAP70 expression Positive 49 (55.7%) 43 (74.1%) 31 (47.7%)
Negative 39 (44.3%) 15 (25.9%) 34 (52.3%)
Not available 13 8 44
CD38 expression <30% 75 (74.3%) 45 (68.2%) 59 (58%)
≥30% 26 (25.7%) 21 (31.8%) 42 (42%)
Not available 0 0 8
Cytogenetics
del17p 1 (1.0%) 1 (1.5%) 13 (11.9%)
del11q 17 (16.8%) 16 (24.2%) 42 (48.5%)
+12 17 (16.8%) 10 (15.2%) 3 (2.8%)
FISH normal 31 (30.7%) 24 (36.4%) 22 (20.2%)
del13q 35 (34.7%) 15 (22.7%) 29 (26.6%)
Time to event parameter in months (median, range)
Diagnosis to sample 21 (0–120) NA
Diagnosis to FCR treatment 21 (1–120) NA
Sample to FCR treatment 0 (0–6) NA
FCR treatment to final follow-up* 146 (40–184) 73 (0.3–93)
Open in a new tab
*

Based on the method of Schemper and Smith,35 based on patients still alive at last follow-up (MDACC, n=59; CLL8, n=63).

The CLL8 validation cohort consisted of 109 patients, treated between 2003–07-21 and 2006–04-04, selected because they were IGHV-unmutated with available gene expression profiling data. When compared to the IGHV-unmutated subset of the MDACC cohort, CLL8 patients were significantly less likely to be ZAP70-positive, but had higher CD38 expression and unfavorable cytogenetics (Table 1). Median follow up was 73.4 months (quartiles=68.2–82.9), median OS was 80.8 (quartiles=58.3–infinity) months, median TTP was 46.5 months (quartiles=23.1–83.0), and median PFS was 42.4 months (quartiles=21.4–72.0) months. At final follow up, 77 (70.6%) of 109 patients progressed, 46 (42.2%) died, and 83 (76.1%) either progressed or died.

Fitting univariate Cox proportional hazards models to GEP data for the MDACC cohort (n=101) at FDR=0.33 (p≤0.0085), we identified 1,136 probes as significantly associated with TTP (Appendix, p.11, Table ST3). Hierarchical clustering using these probes divided patients into three subsets, driven by the gene expression patterns in four subsets (Figure 2A). Logrank test and Cox proportional hazards analysis demonstrated that the three patient subsets had statistically significantly different TTP (log-rank p=5.9 × 10−10) (Figure 2B). Patients assigned to the “intermediate” (group B; magenta) or “unfavorable” (group C; orange) prognosis subsets had an approximately 5 times (HR=4.64, 95% CI 1.06–20.18) or 18 times (HR=18.36, 95% CI 4.37–77.18) higher risk of progression, respectively, compared to patients in the “favorable” (group A; blue) subset.

Figure 2:

Figure 2:

Open in a new tab

Hierarchical gene clustering identifies gene sets that divide patients into prognostic groups. Heat map illustrating hierarchically-clustered gene sets that separate three patient subgroups with different clinical outcomes (TTP) after frontline FCR therapy. 1,136 genes associated with TTP in univariate analysis were included in the supervised clustering process. Each row depicts a single gene, each column a single patient of 101 CLL cases included in this microarray study. Gene sets clustering together are color coded on the y-axis (cyan: 401 probes, green: 70 probes, purple: 424 probes, red: 241 probes); patient subsets (TTP) are colored in blue (group A), orange (group B) and magenta (group C) on the x-axis (upper margin of figure). B: Kaplan-Meier survival functions for time to progression

Pathway analysis with IPA, ToppGene, and DAVID using the 1,136 genes produced similar results (FDR=0.05; Appendix p.12, Table ST4). The purple cluster contains 420 genes that are highly expressed in unfavorable cases with shorter TTP (Figure 2). This cluster is significantly enriched for genes associated with metabolic pathways, including oxidative phosphorylation and ribonucleoside metabolism. ToppGene analysis demonstrates that the most significant Gene Ontology (GO) categories are oxidoreductase activity (GO:0016491, FDR q=1.01×10−8); glycosyl compound, nucleoside and ribonucleoside metabolic processes (GO:1901657, GO:0009116, GO:0009119, all with FDR q=1.22×10−11); and the mitochondrion (GO:0005739, FDR q=9.79×10−16). The most significant pathways are metabolic pathways (KEGG:132956, FDR q=3.04×10−13) and oxidative phosphorylation (KEGG:82942, FDR q=4.31×10−7). DAVID Analysis yielded similar results. These functions suggest that differences in metabolic activity of CLL cells distinguish patients with different outcomes after FCR.

The cyan cluster contains 401 genes that are highly expressed in favorable or intermediate cases with longer TTP. ToppGene analysis demonstrates that the most significant GO categories are ATP binding (GO:0005524, FDR q=2.38×10−3) and purine ribonucleoside triphosphate binding (GO:0035639, FDR q=2.38×10−3). DAVID analysis demonstrates that in the most significant annotation cluster (E=3.79), the most significant GO categories are nucleic acid binding (GO:0003676, p=6.20×10−4), DNA-templated transcription (GO:0006351, p=5.28×10−6), and the nucleus (GO:0005634, p=2.17×10−8). Significant UniProt keywords and sequence features show enrichment for zinc finger transcription factors (p=4.78×10−7 and p=1.47×10−5, respectively). Expression of both the “red” (n=241) and “green” (n=70) groups of genes are variably expressed within the GEP/TTP subtypes (Figure 2), and neither group shows statistically significant enrichment with GO categories (Appendix p.12, Table ST4).

Multivariate analysis of clinical variables in the MDACC cohort (n=100, excluding the del(17p) sample) using AIC (Appendix p.13) selected age at diagnosis, IGHV mutation status, serum β2M, and cytogenetics as independent markers of TTP (Table 2); exploratory univariate analyses were also performed (Appendix pp.14–15, Table ST5). Only these clinical variables were included in additional analyses for model development or validation. When we added the GEP-based categorization of patients (favorable/intermediate/unfavorable) as a competing variable, age, serum β2M, cytogenetics, and the GEP subtypes were retained as significant predictors of TTP (Table 2). When we repeated this analysis on the IGHV-unmutated subset, the only difference was that IGHV status could no longer be included as a factor. Thus, in both the full data set and the IGHV-unmutated subset, GEP was a significant predictor of early versus delayed progression after FCR, independent of clinical variables.

Table 2:

Multivariate models to predict TTP

MDACC GCLLSG/CLL8
All Samples (n = 100) Unmutated (n = 65) Unmutated (n = 109)
Clinical Only p-value HR (95% CI) p-value HR (95% CI) p-value HR (95% CI)
Age at diagnosis* 0.0247 0.97 (0.94–1.00) 0.0311 0.97 (0.94–1.00) 0.108 0.97 (0.95–1.01)
Unmutated IGHV 0.0003 4.51 (1.99–10.21) NA NA NA NA
β2M >4 mg/L 0.0007 3.00 (1.59–5.67) 0.0039 2.75 (1.38–5.46) 0.186 0.68 (0.38–1.21)
Cytogenetics FISH normal 0.7182 1.22 (0.42–3.56) 0.9958 1.00 (0.33–3.00) 0.79 0.92 (0.52–1.65)
del17p <0.0001 4.16 (2.16–8.02)
del11q 0.7096 1.23 (0.41–3.69) 0.8707 1.10 (0.37–3.27) 0.247 1.31 (0.83–2.05)
del13q 0.0270 3.18 (1.14–8.87) 0.0658 2.73 (0.94–7.95) 0.203 0.75 (0.48–1.17)
Clinical Plus GEP Subtype
Age at diagnosis 0.0010 0.94 (0.91–0.98) 0.0033 0.94 (0.91–0.98)
β2M >4 mg/L 0.0004 3.54 (1.75–7.17) 0.0023 3.24 (1.42–6.90)
Cytogenetics FISH normal 0.5667 0.72 (0.24–2.21) 0.2305 0.47 (0.14–1.61)
del11q 0.7304 1.22 (0.40–3.69) 0.9484 0.96 (0.32–2.92)
del13q 0.0557 2.77 (0.98–7.89) 0.2982 1.80 (0.23–5.48)
Subtype§ Intermediate (magenta) 0.0105 7.09 (1.58–31.82)
Unfavorable (orange) <0.0001 28.47 (6.48–125.05) 0.0002 4.30 (2.01–9.18)
Clinical Plus 17-gene signature
Age at diagnosis 0.0159 0.96 (0.93–0.99) 0.0357 0.97 (0.94–1.00) 0.060 0.97 (0.95–1.00)
Unmutated IGHV 0.0129 2.93 (1.26–6.83) NA NA NA NA
β2M >4 mg/L 0.0067 2.59 (1.30–5.15) 0.0722 1.85 (0.95–3.62) 0.149 0.62 (0.33–1.18)
Cytogenetics FISH normal 0.9046 0.94 (0.32–2.76) NS NS 0.589 1.31 (0.50–3.44)
del17p <0.001 8.23 (3.26–20.8)
del11q 0.6980 1.25 (0.41–3.79) NS NS 0.069 1.80 (0.96–3.39)
del13q 0.0805 2.53 (0.89–7.17) NS NS 0.442 1.28 (0.68–2.39)
High 17-gene signature 0.0014 2.88 (1.51–5.52) 0.0006 3.40 (1.70–6.82) 0.005 2.05 (1.24–3.40)
Open in a new tab
*

Treated as a continuous variable. Thus, the hazard ratio is the change associated with one year increase in age.

§

In comparison with all samples, “blue” is used as the baseline. In comparison within unmutated samples, “magenta” is used as the baseline.

Abbreviations: HR = hazard ratio, CI = confidence interval, NA = Not applicable, NS = Not selected.

We developed a 17-gene signature to distinguish IGHV-unmutated patients with intermediate versus unfavorable outcome (Appendix pp.16–20, Figures SF3–5). Clustering cannot, by itself, predict outcomes on new patient samples. Because the cluster with the best outcome was strongly associated with known markers of good outcome (IGHV-mutated, ZAP70-negative), we focused on developing a prognostic model with fewer genes to separate IGHV-unmutated cases with intermediate from those with unfavorable prognosis. Cross-validation suggested that the border between the two groups was imprecise. To increase our ability to distinguish them, we identified “reliable”, i.e., easily classified, samples and built our model using these samples (n=54; 24 intermediate, 30 unfavorable). We identified 726 differentially expressed genes (FDR=15%). Of these, 177 were in the set associated with TTP used for clustering. After omitting duplicate probes, pseudogenes, and genes of uncertain function, 139 candidate predictors remained (Appendix p.18, Figure SF4). Using the Lasso, we selected a minimal set of genes separating the two prognostic groups; the final optimal model contained 17 genes (Table 3). Using the predicted classes from the 17-gene model, we successfully separated the IGHV-unmutated patients into cases with intermediate and unfavorable risk of progression after front-line FCR therapy (Figure 3A). When we applied the prognostic model to the “unreliable” samples that were not used for training, none of the favorable outcome samples were assigned to the unfavorable group. Among IGHV-mutated samples, only 4 out of 35 were assigned to the unfavorable group (Appendix pp.19–20, Figure SF5).

Table 3.

Final 17-gene model to predict prognosis (TTP) in IGHV unmutated CLL patients

Rank* Symbol Standardized Coefficient$ Gene Name Entrez Gene ID Gene Cluster
1 OSBPL5 +0.633 Oxysterol binding protein like 5 114879 purple
2 MSI2 +0.234 Musashi RNA binding protein 2 124540 purple
3 KSR2 +0.219 Kinase suppressor of RAS2 283455 purple
4 NME1 +0.206 NME/NM23 nucleoside diphosphate kinase 1 4830 purple
5 SLC35A4 +0.199 Solute carrier family 35 member A4 113829 purple
6 TXN +0.188 Thioredoxin 7295 purple
7 LAG3 +0.187 Lymphocyte activating 3 3902 red
8 ZNHIT1 +0.162 Zinc finger HIT-type containing 1 10467 purple
9 PDE8A −0.159 Phosphodiesterase 8A 5151 cyan
10 RGS10 +0.150 Regulator of G-protein signaling 10 6001 purple
11 TSPO +0.145 Translocator protein 706 purple
12 CRLF3 −0.129 Cytokine receptor like factor 3 51379 cyan
13 DCAF12 +0.058 DDB1 and CUL4 associated factor 12 25853 purple
14 ADSL +0.040 Adenylosuccinate lyase 158 purple
15 AQP1 −0.037 Aquaporin 1 (Colton blood group) 358 cyan
16 GRN +0.025 Granulin 2896 purple
17 TTC38 +0.018 Tetratricopeptide repeat domain 38 55020 purple
Open in a new tab
*

Rank is the order in which terms were introduced to the model; lower rank genes are more relevant.

$

Standardized coefficient is the weight given to the standardized gene expression (subtracting the mean and dividing by the standard deviation of all IGHV unmutated samples). The full model also uses a constant “intercept” term equal to 0.178.

Figure 3:

Figure 3:

Open in a new tab

Cumulative incidence. A: Aalen-Johansen estimator of cumulative incidence plot for TTP after the start of FCR therapy in IGHV unmutated patients as a function of predicted risk class based on 17 genes in the final prognostic model, for 66 unmutated cases in the MDACC training set. Five year cumulative incidence: 0.25 for low risk and 0.65 for high risk; ten year cumulative incidence: 0.38 for low risk and 0.83 for high risk. B: Same plot for the 109 unmutated samples in the CLL8 independent validation subset. Five year cumulative incidence: 0.51 for low risk and 0.69 for high risk; ten year cumulative incidence, not available

We validated the 17-gene signature using 109 IGHV-unmutated patients from the CLL8 trial who underwent gene expression profiling at study entry. The score is given as a linear combination of standardized gene expression levels using model coefficients, which were learned from the training set (Table 3), as weights. Using a proportional hazards cause-specific Cox regression model, the estimated cause-specific hazard ratio (CSHR) for the continuous score was CSHR=1.41 (95% CI 1.13–1.76). For the binary score the CSHR (high risk vs. low risk) was 1.90 (95% CI 1.18–3.06) with a significant segregation between the two groups (p=0.008, Figure 3B). For multivariate analysis by cause-specific Cox regression, age at diagnosis, β2M >4 mg/L, and cytogenetics were considered as additional covariates available in CLL8. The effect of the continuous score and the binary risk score remained statistically significant (p=0.003 and p=0.006, respectively; Table 2).

DISCUSSION

We have developed and independently validated a robust and reproducible 17-gene signature that distinguishes treatment-naïve IGHV-unmutated CLL patients who are likely to achieve durable remissions following front-line FCR chemoimmunotherapy from those who may benefit from alternative regimens. Differences in the expression of genes involved in oxidative phosphorylation and purine metabolism appear to account, at least in part, for differences in therapy response in patients with IGHV-unmutated CLL.

Our final prognostic model uses only 17 out of the 1,136 genes that were related to TTP in univariate analyses. It is well-known that it is difficult to reproduce lists of prognostic genes selected from different data sets.23,24 The reasons for this instability include individual false positive genes, tumor heterogeneity, small sample sizes, and highly correlated genes.25,26 To address this, we cross-validated multiple steps in model construction and, more importantly, validated the model in an independent patient cohort treated on a separate clinical trial. However, the CLL8 validation cohort differed slightly from the initial MDACC test cohort, i.e., more advanced stage patients and patients with unfavorable cytogenetics. The MDACC training cohort contained only one case with del17p, which we excluded from the survival models. (TP53 mutation status and minimal residual disease testing were unavailable because the samples were acquired before these analyses was performed routinely.) Despite limited patient numbers in both cohorts, we believe that our validation process ensures that the gene signatures obtained are robust. A primary feature of the model is that 13 of the 17 genes come from the collection of “purple” genes; all 13 have positive coefficients indicating that increased expression corresponds to increased risk of progression. In contrast, 3 of the 17 genes come from the “cyan” collection and have negative coefficients, indicating that increased expression corresponds to decreased risk.

To understand the biological factors explaining the model, we applied gene set analysis in the context of the larger set of univariately important genes. Although not apparent from the 17 genes used to construct the model, gene set analysis shows that the “purple” collection, whose upregulation is associated with a poor outcome, is enriched for genes involved in metabolism and oxidative phosphorylation, including genes encoding proteins in complexes I, III, IV, and V, which participate in electron transport across the inner mitochondrial membrane (Appendix p.21, Figure SF6). Recent studies have demonstrated that, unlike many other highly proliferative lymphoid neoplasms and other cancers, CLL cells do not rely upon aerobic glycolysis to generate energy, the Warburg effect. Rather, their major energy source is oxidative phosphorylation. Enhanced oxidative phosphorylation in CLL cells is associated with poor prognostic features including unmutated IGHV genes, ZAP70 positivity, high Rai stage, increased serum β2M, and fludarabine resistance.27,28

The “purple” collection is also enriched for genes involved in purine metabolism. Purines, essential components of nucleic acids, also provide energy and cofactors required for cell survival and proliferation.29 Purine nucleotides are synthesized through two different pathways, the purine salvage pathway and the de novo purine biosynthesis pathway. One of the most highly expressed “purple” genes is adenylosuccinate lyase (ADSL), which participates in both pathways. In the purine salvage pathway, ADSL catalyzes the formation of adenylate (AMP) from adenylosuccinate (S-AMP) in the conversion of inosine monophosphate (IMP) into adenine nucleotides. In the de novo pathway, ADSL catalyzes the cleavage of succinyl groups to yield fumarate. Other highly expressed genes in this group include adenosine deaminase (ADA) and deoxyuridine triphosphate nucleotidohydrolase (DUT).

As expected, the vast majority of cases identified as having a favorable prognosis were IGHV-mutated; about two-thirds showed del13q as the sole abnormality. However, about one-third of intermediate or poor prognosis cases were also IGHV-mutated, and one-third of these showed del13q. Thus, mutated IGHV genes and del13q did not ensure a good prognosis. Other than the clinical outcome and the results of the 17-gene signature, we could not identify any clinical or laboratory features that would allow us to distinguish these cases from those with a good prognosis. A recent study showed that a subset of IGHV-mutated cases with isolated del13q demonstrate karyotypic complexity on stimulated chromosome banding analysis, and have a poor prognosis.30 We are unable to explore this possibility because our cohort predates routine stimulated chromosome banding analysis. In our multivariate models, del13q appears to have a deleterious effect. This artifact may be due to an interaction between IGHV mutation status and cytogenetics that we could not investigate because of the small sample size.

Tyrosine kinase inhibitors (TKIs), such as ibrutinib, a Bruton tyrosine kinase (BTK) inhibitor, and idelalisib, a PI3 kinase delta isoform (PI3Kδ) inhibitor, have revolutionized the treatment of patients with del17p and with relapsed, refractory disease. However, treatment with TKIs is associated with significant toxicities.31 Patients treated with ibrutinib are at increased risk for atrial fibrillation and hemorrhage; a small number of patients have developed ventricular arrhythmias or died suddenly. Patients treated with idelalisib are at increased risk for transaminitis, colitis, and pneumonitis.31,32 About 10% of patients will discontinue therapy due to side effects. Finally, it remains to be determined if disease will relapse if therapy is discontinued.

A major advantage of FCR chemoimmunotherapy with intent to cure is that the regimen is relatively brief, every four weeks for six cycles, and inexpensive. But treatment with FCR also has significant limitations.33 Up to 5% of patients treated with frontline FCR will develop a therapy-related myeloid neoplasm.21,34 Patients older than 65 or those with comorbidities often cannot tolerate a full course of therapy, and are at increased risk to develop myelosuppression and opportunistic infections.33 However, for patients who desire a short therapy course, or who cannot tolerate TKI therapy, FCR remains a viable option. We would recommend testing the value of the 17-gene signature in a prospective study that compares treatment with FCR to alternative therapies, such as ibrutinib, as part of a randomized clinical trial.

Supplementary Material

1
2
3

RESEARCH IN CONTEXT.

Evidence before this study

We searched PubMed for all studies that identified the subset of treatment-naïve CLL patients who may achieve durable remissions following chemoimmunotherapy with fludarabine, cyclophosphamide, and rituximab (FCR). We used the search terms “Gene expression profiling”, “CLL”, “fludarabine”, “FCR”, and “prognosis”. Several recent studies demonstrate that a subset of relatively young, fit, patients whose CLL cells contain mutated immunoglobulin heavy chain variable region (IGHV) genes and lack high-risk cytogenetic abnormalities achieve durable remissions following FCR chemoimmunotherapy. However, we found no studies that sought to identify patients with unmutated IGHV genes who might also achieve durable remissions following FCR chemoimmunotherapy.

Added value of this study

Using data from two seminal clinical trials that demonstrated the efficacy of frontline FCR chemoimmunotherapy in CLL, the FCR300 and CLL8 clinical trials, we developed a robust and reproducible 17-gene signature that distinguishes between treatment-naïve IGHV-unmutated CLL patients who are likely to achieve long-term remissions following front-line FCR chemoimmunotherapy from those who may benefit from alternative regimens.

Implications of all the available evidence

Our results suggest that differences in the expression of genes involved in oxidative phosphorylation and purine metabolism account, at least in part, for the differences in response to frontline FCR chemoimmunotherapy in patients with IGHV-unmutated CLL. Our study indicates that, using the 17-gene signature, it is possible to identify patients with IGHV-unmutated CLL who are likely to achieve long-term remissions with FCR. Thus, for this subset of patients who may desire a short therapy course, or who cannot tolerate therapy with tyrosine kinase inhibitors, FCR remains a viable option. We would recommend testing the value of the 17-gene signature in a prospective study that compares treatment with FCR to alternative therapies, such as ibrutinib, as part of a randomized clinical trial.

Acknowledgments

This work is supported by grants from the CLL Global Research Foundation and the National Institutes of Health/National Cancer Institute R01 CA182905 and P30 CA016058.

Declaration of Interests: CDH reports grants from Hoffmann-La Roche during the conduct of the study, and research funding and travel support from Roche in the context of other clinical trials. KRC reports grants from NIH/NCI during the conduct of the study, and grants from the NIH/NCI and NIH/NLM outside the submitted work. JBa reports honoraria and travel support from Roche during the conduct of the study. KF reports non-financial support from Roche during the conduct of the study, and personal fees from AbbVie outside the submitted work. MH reports grants and/or other support from Roche, AbbVie, Gilead, and Janssen, and Celgene outside the submitted work. SS reports grants, personal fees and non-financial support from Hoffmann La-Roche during the conduct of the study; grants, personal fees and non-financial support from AbbVie, Amgen, AstraZeneca, Celgene, Gilead, GSK, Hoffmann La-Roche, Janssen, Novartis, Pharmacyclics, and Sunesis outside the submitted work. MJK reports grants from AbbVie during the course of the study. LVA reports grants from NCI/NIH during the conduct of the study.

Footnotes

Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Contributor Information

Carmen D. Herling, Department I of Internal Medicine, Center for Integrated Oncology Aachen-Bonn-Cologne-Duesseldorf, Germany.

Kevin R. Coombes, Department of Biomedical Informatics, The Ohio State University, Columbus, Ohio, USA.

Axel Benner, Division of Biostatistics, German Cancer Research Center (DKFZ), Heidelberg, Germany.

Johannes Bloehdorn, Internal Medicine III, University Hospital Ulm, Ulm, Germany.

Lynn L. Barron, Department of Hematopathology, The University of Texas M.D. Anderson Cancer Center, Houston, Texas, USA

Zachary B. Abrams, Department of Biomedical Informatics, The Ohio State University, Columbus, Ohio, USA

Tadeusz Majewski, Department of Pathology and Laboratory Medicine, The University of Texas M.D. Anderson Cancer Center, Houston, Texas, USA.

Jolanta E. Bondaruk, Department of Pathology and Laboratory Medicine, The University of Texas M.D. Anderson Cancer Center, Houston, Texas, USA

Jasmin Bahlo, Department I of Internal Medicine, Center for Integrated Oncology Aachen-Bonn-Cologne-Duesseldorf, Germany.

Kirsten Fischer, Department I of Internal Medicine, Center for Integrated Oncology Aachen-Bonn-Cologne-Duesseldorf, Germany.

Michael Hallek, Department I of Internal Medicine, Center for Integrated Oncology Aachen-Bonn-Cologne-Duesseldorf, Germany.

Stephan Stilgenbauer, Internal Medicine III, University Hospital Ulm, Ulm, Germany.

Bogdan A. Czerniak, Department of Pathology and Laboratory Medicine, The University of Texas M.D. Anderson Cancer Center, Houston, Texas, USA

Christopher Oakes, Department of Internal Medicine, The Ohio State University, Columbus, Ohio, USA.

Alessandra Ferrajoli, Department of Leukemia, The University of Texas M.D. Anderson Cancer Center, Houston, Texas, USA.

Michael J. Keating, Department of Leukemia, The University of Texas M.D. Anderson Cancer Center, Houston, Texas, USA

Lynne V. Abruzzo, Department of Pathology, The Ohio State University, Columbus, Ohio, USA

REFERENCES

  • 1.Hallek M, Fischer K, Fingerle-Rowson G, et al. Addition of rituximab to fludarabine and cyclophosphamide in patients with chronic lymphocytic leukaemia: a randomised, open-label, phase 3 trial. Lancet 2010; 376(9747): 1164–74. [DOI] [PubMed] [Google Scholar]
  • 2.Keating MJ, O’Brien S, Albitar M, et al. Early results of a chemoimmunotherapy regimen of fludarabine, cyclophosphamide, and rituximab as initial therapy for chronic lymphocytic leukemia. J Clin Oncol 2005; 23(18): 4079–88. [DOI] [PubMed] [Google Scholar]
  • 3.Fischer K, Bahlo J, Fink AM, et al. Long-term remissions after FCR chemoimmunotherapy in previously untreated patients with CLL: updated results of the CLL8 trial. Blood 2016; 127(2): 208–15. [DOI] [PubMed] [Google Scholar]
  • 4.Thompson PA, Tam CS, O’Brien SM, et al. Fludarabine, cyclophosphamide, and rituximab treatment achieves long-term disease-free survival in IGHV-mutated chronic lymphocytic leukemia. Blood 2016; 127(3): 303–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Eichhorst B, Fink AM, Bahlo J, et al. First-line chemoimmunotherapy with bendamustine and rituximab versus fludarabine, cyclophosphamide, and rituximab in patients with advanced chronic lymphocytic leukaemia (CLL10): an international, open-label, randomised, phase 3, non-inferiority trial. Lancet Oncol 2016; 17(7): 928–42. [DOI] [PubMed] [Google Scholar]
  • 6.Duzkale H, Schweighofer CD, Coombes KR, et al. LDOC1 mRNA is differentially expressed in chronic lymphocytic leukemia and predicts overall survival in untreated patients. Blood 2011; 117(15): 4076–84. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.McCarthy H, Wierda WG, Barron LL, et al. High expression of activation-induced cytidine deaminase (AID) and splice variants is a distinctive feature of poor-prognosis chronic lymphocytic leukemia. Blood 2003; 101(12): 4903–8. [DOI] [PubMed] [Google Scholar]
  • 8.Schweighofer CD, Coombes KR, Majewski T, et al. Genomic variation by whole-genome SNP mapping arrays predicts time-to-event outcome in patients with chronic lymphocytic leukemia: a comparison of CLL and HapMap genotypes. J Mol Diagn 2013; 15(2): 196–209. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9.Admirand JH, Knoblock RJ, Coombes KR, et al. Immunohistochemical detection of ZAP70 in chronic lymphocytic leukemia predicts immunoglobulin heavy chain gene mutation status and time to progression. Mod Pathol 2010; 23(11): 1518–23. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Rassenti LZ, Huynh L, Toy TL, et al. ZAP-70 compared with immunoglobulin heavy-chain gene mutation status as a predictor of disease progression in chronic lymphocytic leukemia. N Engl J Med 2004; 351(9): 893–901. [DOI] [PubMed] [Google Scholar]
  • 11.Zenz T, Dohner H, Stilgenbauer S. Genetics and risk-stratified approach to therapy in chronic lymphocytic leukemia. Best Pract Res Clin Haematol 2007; 20(3): 439–53. [DOI] [PubMed] [Google Scholar]
  • 12.Stilgenbauer S, Schnaiter A, Paschka P, et al. Gene mutations and treatment outcome in chronic lymphocytic leukemia: results from the CLL8 trial. Blood 2014; 123(21): 3247–54. [DOI] [PubMed] [Google Scholar]
  • 13.Du P, Kibbe WA, Lin SM. lumi: a pipeline for processing Illumina microarray. Bioinformatics 2008; 24(13): 1547–8. [DOI] [PubMed] [Google Scholar]
  • 14.Pounds S, Morris SW. Estimating the occurrence of false positives and false negatives in microarray studies by approximating and partitioning the empirical distribution of p-values. Bioinformatics 2003; 19(10): 1236–42. [DOI] [PubMed] [Google Scholar]
  • 15.Vasyutina E, Boucas JM, Bloehdorn J, et al. The regulatory interaction of EVI1 with the TCL1A oncogene impacts cell survival and clinical outcome in CLL. Leukemia 2015; 29(10): 2003–14. [DOI] [PubMed] [Google Scholar]
  • 16.Cheson BD, Bennett JM, Grever M, et al. National Cancer Institute-sponsored Working Group guidelines for chronic lymphocytic leukemia: revised guidelines for diagnosis and treatment. Blood 1996; 87(12): 4990–7. [PubMed] [Google Scholar]
  • 17.Burnham K AD. Model selection and multimodel inference: A practical information-theoretic approach. New York, NY: Springer-Verlag; 2002. [Google Scholar]
  • 18.Tibshirani R The lasso method for variable selection in the Cox model. Stat Med 1997; 16(4): 385–95. [DOI] [PubMed] [Google Scholar]
  • 19.Graf E, Schmoor C, Sauerbrei W, Schumacher M. Assessment and comparison of prognostic classification schemes for survival data. Stat Med 1999; 18(17–18): 2529–45. [DOI] [PubMed] [Google Scholar]
  • 20.Brier G Verification of forecasts expressed in terms of probability. Monthly Weather Review 1950; 78: 1–3. [Google Scholar]
  • 21.Tam CS, O’Brien S, Wierda W, et al. Long-term results of the fludarabine, cyclophosphamide, and rituximab regimen as initial therapy of chronic lymphocytic leukemia. Blood 2008; 112(4): 975–80. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Short NJ, Keating MJ, Wierda WG, et al. Fludarabine, cyclophosphamide, and multiple-dose rituximab as frontline therapy for chronic lymphocytic leukemia. Cancer 2015; 121(21): 3869–76. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23.Ein-Dor L, Kela I, Getz G, Givol D, Domany E. Outcome signature genes in breast cancer: is there a unique set? Bioinformatics 2005; 21(2): 171–8. [DOI] [PubMed] [Google Scholar]
  • 24.Ntzani EE, Ioannidis JP. Predictive ability of DNA microarrays for cancer outcomes and correlates: an empirical assessment. Lancet 2003; 362(9394): 1439–44. [DOI] [PubMed] [Google Scholar]
  • 25.Michiels S, Koscielny S, Hill C. Interpretation of microarray data in cancer. Br J Cancer 2007; 96(8): 1155–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26.Xu JZ, Wong CW. Hunting for robust gene signature from cancer profiling data: sources of variability, different interpretations, and recent methodological developments. Cancer Lett 2010; 296(1): 9–16. [DOI] [PubMed] [Google Scholar]
  • 27.Sharma A, Janocha AJ, Hill BT, Smith MR, Erzurum SC, Almasan A. Targeting mTORC1-mediated metabolic addiction overcomes fludarabine resistance in malignant B cells. Mol Cancer Res 2014; 12(9): 1205–15. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28.Vangapandu HV, Ayres ML, Bristow CA, et al. The Stromal Microenvironment Modulates Mitochondrial Oxidative Phosphorylation in Chronic Lymphocytic Leukemia Cells. Neoplasia 2017; 19(10): 762–71. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29.Yin J, Ren W, Huang X, Deng J, Li T, Yin Y. Potential Mechanisms Connecting Purine Metabolism and Cancer Therapy. Front Immunol 2018; 9: 1697. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30.Baliakas P, Jeromin S, Iskas M, et al. Cytogenetic complexity in chronic lymphocytic leukemia: definitions, associations and clinical impact. Blood 2019. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Mauro FR, Caputo MD, Rosati S, et al. Balancing efficacy and toxicity of targeted agents currently used for the treatment of patients with chronic lymphocytic leukemia. Expert Rev Hematol 2018; 11(8): 601–11. [DOI] [PubMed] [Google Scholar]
  • 32.Brown JR, Kay NE. Chemoimmunotherapy Is Not Dead Yet in Chronic Lymphocytic Leukemia. J Clin Oncol 2017; 35(26): 2989–92. [DOI] [PubMed] [Google Scholar]
  • 33.Skarbnik AP, Faderl S. The role of combined fludarabine, cyclophosphamide and rituximab chemoimmunotherapy in chronic lymphocytic leukemia: current evidence and controversies. Ther Adv Hematol 2017; 8(3): 99–105. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34.Tam CS, Seymour JF, Prince HM, et al. Treatment-related myelodysplasia following fludarabine combination chemotherapy. Haematologica 2006; 91(11): 1546–50. [PubMed] [Google Scholar]
  • 35.Schemper M, Smith TL. A note on quantifying follow-up in studies of failure time. Controlled clinical trials 1996; 17(4): 343–6. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

1
NIHMS1545880-supplement-1.docx (981.5KB, docx)
2
NIHMS1545880-supplement-2.xlsx (182.4KB, xlsx)
3
NIHMS1545880-supplement-3.xlsx (50.4KB, xlsx)

RESOURCES