Gene expression of INPP5F as an independent prognostic marker in fludarabine-based therapy of chronic lymphocytic leukemia

Chronic lymphocytic leukemia (CLL) is a heterogeneous disease. Various disease-related and patient-related factors have been shown to influence the course of the disease. The aim of this study was to identify novel biomarkers of significant clinical relevance. Pretreatment CD19-separated lymphocytes (n=237; discovery set) and peripheral blood mononuclear cells (n=92; validation set) from the REACH trial, a randomized phase III trial in relapsed CLL comparing rituximab plus fludarabine plus cyclophosphamide with fludarabine plus cyclophosphamide alone, underwent gene expression profiling. By using Cox regression survival analysis on the discovery set, we identified inositol polyphosphate-5-phosphatase F (INPP5F) as a prognostic factor for progression-free survival (P<0.001; hazard ratio (HR), 1.63; 95% confidence interval (CI), 1.35–1.98) and overall survival (P<0.001; HR, 1.47; 95% CI, 1.18–1.84), regardless of adjusting for known prognostic factors. These findings were confirmed on the validation set, suggesting that INPP5F may serve as a novel, easy-to-assess future prognostic biomarker for fludarabine-based therapy in CLL.


INTRODUCTION
Chronic lymphocytic leukemia (CLL) is generally an incurable disease. However, the clinical course of the disease is heterogeneous. Although some patients experience rapid progression and rapid need of antileukemic therapy, others remain stable and are observed for many years. This heterogeneity is also reflected in the response to therapy and in long-term clinical outcomes such as progression-free survival (PFS) and overall survival (OS). Prognostic factors that influence patient outcomes consist of cytogenetic rearrangements such as del(17p) and del(11q), molecular factors such as the immunoglobulin variable region heavy chain (IGVH) mutational status, p53 mutational status and expression of cell surface markers such as CD38 and ZAP70. 1 Clinical factors such as age, stage of disease and lymphocyte doubling time have also been shown to have prognostic relevance. 2 For patients in need of therapy, chemoimmunotherapy such as the combination of fludarabine and cyclophosphamide with the anti-CD20 monoclonal antibody rituximab (R-FC) has been demonstrated to significantly prolong PFS and OS compared with the combination of fludarabine and cyclophosphamide alone (FC) in untreated (first-line) patients 3 as well as to improve PFS in previously treated (second-line) patients. 4 The objective of this study was to evaluate prognostic and/or predictive biomarkers for rituximab-based therapy from the REACH trial. Recently, we reported that higher expression levels of PTK2 mRNA were associated with improved PFS in patients treated with R-FC compared with FC alone, whereas no significant differences were observed between the two treatment arms in patients with lower PTK2 expression levels, 5 demonstrating predictive significance for rituximab-based chemoimmunotherapy. Here, we demonstrate the prognostic relevance of inositol polyphosphate-5-phosphatase F (INPP5F) gene expression, a novel, easily assessable biomarker in relapsed CLL.

REACH (NCT0090051)
was an open-label randomized (1:1) phase III trial in relapsed CLL comparing FC with R-FC. The primary end point of the study was to demonstrate prolonged PFS of R-FC compared with FC alone. The study protocol was approved by institutional review boards at participating centers and all patients gave written informed consent. Details on trial design and eligibility criteria and clinical outcome have been described previously. 4 Patients were selected based on the availability of sufficient RNA and had to provide written informed consent to participate in the additional studies.

Gene expression profiling
Pretreatment samples for molecular profiling analysis were available from 237 of 552 (55%) patients enrolled in the REACH trial, selected by the availability of adequate quality and quantity of RNA and on whether the patients gave their informed consent to further molecular analysis. Peripheral blood mononuclear cells (PBMCs) were obtained by Ficoll density gradient centrifugation. CLL samples were positively enriched by magnetic cell sorting using CD19 microbeads and MACS columns (Miltenyi Biotec GmbH, Bergisch Gladbach, Germany) and resuspended in RLT buffer (Qiagen, Crawley, West Sussex, UK). In addition, as a validation set, PBMC samples (n = 92) were processed into RLT buffer without CD19 enrichment. RNA from both sets was subsequently isolated using Ambion miRVANA total RNA extraction kit (Life Technologies, Carlsbad, CA, USA). Gene expression profiling was performed using Affymetrix U133 Plus 2.0 full transcriptome oligonucleotide arrays according to the manufacturerecommended protocol (Affymetrix, Santa Clara, CA, USA).

Expression-level computations
The probe intensities of mRNA HG-U133 Plus 2.0 array were backgroundsubtracted, quantile-normalized and summarized using the RMA method 6 within each sample type (CD19 + and PBMCs). Data were further normalized by an empirical Bayes approach (Combat) 7 to remove potential batch effects due to data acquisition. All data used in this work were log 2 transformed. Probe sets with very low expression (log 2 -transformed expression levels o4.0 in 495% of samples) were excluded and a total of 26 453 probe sets (out of 54 675) were considered for further analysis.

Statistical analysis
To test for a possible selection bias on the baseline patient characteristics for each mRNA subset (CD19+ and PBMCs) as compared with the overall REACH population, the χ 2 test or Wilcoxon-Mann-Whitney test was used respectively for binary and continuous variables. Similar statistical tests were used to check whether the baseline patient characteristics were balanced between FC and R-FC within each mRNA subset. The predictive and/or prognostic utility of the mRNA was assessed using the following approach. First, a log-likelihood ratio test (with two degrees of freedom (2 d.f.)) was used to compare the Cox proportional hazards model of PFS including treatment, mRNA (as continuous) and treatment-mRNA interaction as covariates (full model) against the same model but with the interaction term being absent (reduced model). Use of a join test (2 d.f.) allowed simultaneous searching for both prognostic and predictive markers, significantly increasing the power to detect predictive markers compared with a 1-d.f. treatment-mRNA pure interaction test (based on simulations; data not shown). The q-values (false discovery rate) 8 were calculated from raw log-likelihood ratio test P-values to account for multiple hypothesis testing and only mRNAs with a q-value o1% were considered significant. At this stage, the full model was considered again and the treatment-mRNA interaction term was tested to decide whether a significant marker was predictive or prognostic. In addition, to assess whether candidate mRNAs provide predictive or prognostic information independently of known prognostic factors, the mRNA was also evaluated in the context of an expanded, multivariate Cox proportional hazards model that included a parsimonious set of known prognostic factors chosen (using a forward stepwise selection procedure) among the following variables: age; Binet stage; IGVH mutational status; chromosome 17p, 11q and 13q deletions; trisomy 12; β2-microgloblulin; lymphocytic count; and Eastern Cooperative Oncology Group (ECOG) performance status.

Patient characteristics
Gene expression profiling data were available from 237 CD19enriched samples (CD19 + discovery set: 115 within the FC arm; 122 within the R-FC arm) and 92 PBMC samples that were not enriched for CD19 (PBMC validation set: 46 within the FC arm; 46 within the R-FC arm) of 552 patients enrolled in the REACH trial. The baseline patient demographics and tumor characteristics are shown in Table 1 for the overall REACH population, as well as for the CD19 + and PBMC subpopulations with available mRNA (mRNA study population). Summaries from the table suggest that the mRNA study population was representative of the REACH overall study population 4 (formal statistical tests for each variable in Table 1 did not show any statistical difference between either CD19+ or PBMC mRNA subsets as compared with the overall population) and that the two treatment arms were well balanced with respect to related risk factors, such as age, stage, high-risk cytogenetics and IGVH mutational status (again there were no statistically significant differences between FC and R-FC arm within each mRNA subset  Figure 1a). The prognostic value of INPP5F expression on PFS was confirmed in the validation set (PBMC samples) with the mRNA term P o 0.001 in a Cox regression model regardless of adjusting for age, Binet stage, IGVH mutational status, del(17p) and ECOG performance status (HR, 1.73; 95% CI, 1.33-2.24 without adjustment; HR, 1.83; 95% CI, 1.3-2.6 with adjustment). Again, the treatment-mRNA interaction term was not significant (P = 0.9/0.9 without/with adjustment). The Kaplan-Meier curves for patients grouped into low and high INPP5F expression based on the median expression level are shown in Figure 1b (median PFS, 33 vs 17.6 months, respectively).
In addition, INPP5F expression measured by Affymetrix HG-U133 Plus 2.0 microarray from a set of 107 patients with newly diagnosed CLL 9 available in the Gene Expression Omnibus database (http://www.ncbi.nlm.nih.gov/geo/; experiment-ID = E-GEOD-22762) was analyzed for OS. In this additional, independent data set, a significant correlation between OS and INPP5F (P = 0.006; HR, 2.0; 95% CI, 1.2-3.3) was again observed in a univariate Cox regression analysis. INPP5F remained significant (P = 0.02; HR, 1.7; 95% CI, 1.1-2.7) after adjusting for del(17p) and trisomy 12 (a stepwise variable selection did exclude del(11q) and del(13q) from the model; no other predictors were made publicly available). The Kaplan-Meier curves of OS for patients grouped into low and high INPP5F expression are provided in Figure 3.
Correlation of INPP5F expression to BCL-2, IKKβ (IKKb) and IKBα (IKBa) In the discovery set (CD19 + ), there was a significant correlation of INPP5F expression to the expression level of BCL-2 (r = 0.4; P o0.001). A trend toward statistical significance was observed when INPP5F and BCL-2 expression were correlated in the  Table 1).

DISCUSSION
INPP5F is one of the several polyphosphoinositide phosphatases whose role has been partially elucidated. 11 INPP5F degrades PIP2 (phosphatidylinositol 4,5-bisphosphonate) and PIP3 (phosphatidylinositol 3,4,5-trisphosphonate) regulating AKT/phosphatidylinositol 3-kinase (PI3K) signaling. [12][13][14][15] INPP5F is predicted to reduce PIP3 levels and subsequently reduce AKT/PI3K signaling, thereby attenuating expression of antiapoptotic molecules and leading to increased chemotherapy sensitivity. Because of this, our data showing that lower INPP5F levels are associated with improved outcome and higher INPP5F levels with poorer outcome in CLL may seem contradictory. However, it has been demonstrated that INPP5F acts as a regulator of PIP3 levels. Stimulation of the AKT/PI3K pathway with insulin-like growth factor in INPP5F knockout mice led to a substantial increase in PIP3 levels compared with INPP5F wild-type mice, 16 suggesting a feedback loop of INPP5F when the ATK/PI3K pathway is stimulated. In addition, unlike PTEN, which degrades PIP3 to PI(4,5)P2, INPP5F degrades PIP3 to PI (3,4)P2, which can function as a second messenger. Furthermore, PI(3,4)P2 activity has been shown to correlate with AKT activity. 15 In CLL, the NF-κB signaling pathway has been shown to be constitutively active, 17 resulting in downstream overexpression of antiapoptotic BCL-2 family members. 18,19 BCL-2 and BCL-2 family members have been shown to be upregulated in CLL and are also shown to be associated with an unfavorable prognosis [20][21][22][23][24][25][26][27] and the connection between BCL-2 and NF-κB has been reported for various hematological malignancies including CLL. 17,28 Rituximab has been shown to exert its antitumor activity in vitro by attenuating constitutively active AKT and subsequent modulation expression of the antiapoptotic proteins of the BCL-2 family. 29,30 Moreover, this rituximab-mediated downregulation of active AKT resulted in sensitization to chemotherapeutics. Low levels of BCLxl have been shown to be associated with sensitivity to chemotherapy and CD20-targeted therapy. 29 In addition, increased levels of BCL-2 family protein have been associated with poorer response to fludarabine and a shorter time to first treatment interval. 23 We also detected significant correlation of INPP5F expression with gene expression of BCL-2 (r = 0.4; Po 0.001) and a significant correlation of INPP5F to IKKb expression (r = 0.45; P o0.001), whereas for the expression levels of IKBa, a significant inverse correlation to the expression of INPP5F was observed (r = -0.26; P o 0.001). Interestingly, the INPP5F expression level correlated significantly with the expression of IKKb/IKBKB, an activator of NF-κB, and was inversely correlated to expression of IKBa, an inhibitor of NF-κB, 31,32 suggesting an association between high expression of INPP5F and the activated NF-κB pathway and an agonistic AKT/PI3K function. However, none of the latter genes demonstrated significant prognostic relevance in the present study.   To our knowledge, this is the first study describing INPP5F expression level as a novel prognostic biomarker for PFS as well as OS in a large set of patients treated within a randomized trial. Its prognostic relevance was found in a set of CD19-enriched PBMC samples, and subsequently confirmed in a validation set of unselected PBMCs. The robustness of gene expression profiling for biomarker research in peripheral blood has been demonstrated in multiple previous reports, showing a significant correlation of expression levels between array-based gene expression analysis and reverse transcription-PCR, such as the MAQC Consortium 33 and also previous publication of outcomes with this same set of samples. 5,34 Along these lines, Herold et al. 9,10 showed significant correlation of MZB-1 expression in CLL samples when comparing array-based expression profiling with reverse transcription-PCR. In addition, MZB-1 showed significant correlation to the expression of INPP5F, a result that could be confirmed in our discovery set (r = 0.44, P o 0.001; Supplementary Materials) and validation set (r = 0.36, P o 0.001; Supplementary Materials), further emphasizing the robustness of the reported array data. Further evidence of the prognostic relevance of INPP5F levels stems from a subset of untreated (first-line) CLL patients, in whom high expression levels of INPP5F mRNA were associated with resistance to R-FC and FC therapy, 35 as well as a set of previously treated and untreated CLL patients 9 showing that INPP5F expression levels were an adverse prognostic factor for OS. Further research is warranted to understand the biological function of INPP5F on survival signaling. INPP5F may be a novel, easily assessable biomarker in CLL, because its prognostic significance of INPP5F expression remained detectable in the set of unselected PBMCs, and this may support the clinical feasibility of this biomarker as the analytical workflow may not require separation of B cells as for ZAP70 or laborious assessment of the IGVH mutational status. To determine the significance of INPP5F for prognostic and therapeutic decision making, INPP5F cutoff levels need to be defined, and the prognostic relevance of INPP5F expression needs to be analyzed in the context of other therapies used for treatment of CLL such as bendamustine. 36,37 In addition, our data may also support the strategy of combining NF-κB and/or BCL-2 inhibitors to standard therapy in CLL.

CONFLICT OF INTEREST
GP, DM, MB, HS, GD-N, TN, R-FY and MW are employees of F Hoffmann-La Roche Ltd. TR received research funding and honoraria from F Hoffmann-La Roche Ltd. AD and DD declare no conflict of interest.