Identification of biological markers of sensitivity to high-clinical-risk-adapted therapy for DLBCL patients

Diffuse large B-cell lymphoma (DLBCL) has striking clinical and molecular variability. Although a more precise identification of the multiple determinants of this variability is still under investigation, there is a consensus that high-clinical-risk DLBCL cases require a risk-adapted therapy, since intensification of chemotherapy with autologous stem-cell transplantation (ASCT) has been shown to improve the prognosis for high-risk patients in randomised clinical trials. In spite of this, the protocols used for these patients have a high morbidity, associated with ASCT and the use of multiple drugs. This makes it important to identify patients that may take benefit from risk-adapted therapies, through the recognition of biological markers that provide information about both the tumoral cells and the microenvironment. Unfortunately, many of the studies so far performed have relied on heterogeneous series of patients, staged or treated with different protocols. For instance, some of the variability in DLBCL arises from the fact that this diagnosis is applied to de novo and secondary tumours, nodal and extranodal, irrespective of clinical stage, patient age and associated infections. Additionally, DLBCL includes some specific variants, such as mediastinal DLBCL and T/HRBCL, with specific prognostic parameters. This could prevent the identification of potential predictive biomarkers, because the results of many studies show that the search for predictive biomarkers should be promoted in the context of samples of clinically homogeneous patients enrolled in clinical trials. A further source of variability is the dependence of some predictive markers on specific therapeutic approaches, as is the case for the Bcl6 expression in DLBCL, since Bcl-6+ cases have been shown not to benefit from the addition of R to CHOP. Here we have analysed a series of high-clinical-risk DLBCLs by a two-stage approach, first identifying functional signatures by expression analysis, then analysing surrogate biomarkers using tissue microarrays (TMAs). This eclectic approach could reveal new aspects of the relationship between the neoplastic cells and the microenvironment, leading to the identification of previously unknown prognostic markers. At the same time, the use of functional signatures to analyse expression-profiling data avoids the poor reproducibility of the data obtained from gene-by-gene analysis, and benefits from the existence of a growing body of data concerning the major pathways deregulated in DLBCL. To avoid the bias of semiquantitative scoring, in this study we have quantified the markers included in the multivariate analysis. Keywords: new biological variables, risk-adapted therapies Patients The study was performed in a series of 74 high-clinical-risk DLBCL patients included in a prospective trial from the GEL/TAMO Spanish Group (Grupo Español de Linfomas / Trasplante de Médula Ósea) between June 2001 and June 2005. High risk was defined as having either an age-adjusted IPI score >=2 or IPI <2 with a high β2 microglobulin level. All patients received therapy with MegaCHOP, after which, patients were again evaluated for response by CT and Ga 67S. Patients with CT response and negative Ga 67S received another MegaCHOP cycle followed by BEAM and ASCT. Those patients with positive Ga 67S or without CT response underwent salvage treatment with Ifosfamide and Etoposide (IFE) for two courses, followed by BEAM and ASCT whenever a response had been achieved. From June 2003 Rituximab was added to the schema, before administering ASCT. Those patients with less than a Partial Response (PR) or progressive disease after IFE were removed from the study. Diagnoses were revised by a panel of pathologists, following the World Health Organization (WHO)9 criteria. Only primary DLBCLs were included in this analysis. T-cell/histiocyte-rich large B-cell lymphomas (T/HRLBCLs) and mediastinal DLBCLs were excluded. Patients included in this biological analysis were consecutive, selected only on the basis of the availability of formalin-fixed paraffin-embedded or frozen tissue representative of the tumour at diagnosis. The final number of cases analysed was 50. Frozen diagnostic tissue was available from a subset of ten of these patients. Informed consent was obtained from the patients included in the trial under the supervision of the Local Ethical Committees. Identifying biological function surrogate markers for clinical prediction The expression profile was initially analysed in the ten samples from which frozen tissue was available. These included four patients who had died after treatment failure and six who were alive, with complete remission, and free of disease. All frozen tissue corresponded to a representative area of the tumour with more than 60% of viable tumoral cells present. We used a multistep gene-set-enrichment approach to identify functional pathways deregulated in DLBCLs resistant to the treatment strategy. The design of the analysis was as follows (details on the procedure are included as Supplementary information on material and Methods) : 1. Microarray analysis, using both cDNA and oligonucleotide microarrays in a set of 10 frozen diagnostic samples. 2. Identification of biological function representative of differentially expressed genes, using the Gene Ontology database (http://bioinformatics.weizmann.ac.il/cards) 3. Selection of surrogate immunohistochemical markers (representative of the biological function identified) 4. Immunohistochemical study on Tissue Microarrays containing diagnostic initial biopsy tissue in a set of 50 patients a. Semiquantitative scoring b. Quantification of the most relevant markers Identification of biological markers predicting treatment response The relations between the semiquantitative measurement of the 60 proteins evaluated and the clinical outcome were analysed. The clinical end-points were: a) CR after 3 MegaCHOP, b) CR after IFE and, c) absence of progression after consolidation with HDT/ ASCT. The chi-square statistic was used to assess associations between the protein-expression levels and the intermediate step of treatment. Significant relationships (p<0.05) and non-significant trends (p<0.1) were identified. Failure was defined as progression or death attributable to the tumour. Failure-free survival (FFS) curves were plotted using the Kaplan-Meier method. Statistical significance of associations between individual variables and FFS was determined using the log-rank test. Significant relationships (p<0.05) and non-significant trends (p<0.1) were identified. Genes were clustered on the basis of their biological similarities. The accumulations or alternations of marker alterations were analysed by the chi-square and Spearman correlation tests. Only significant relationships (p<0.05) were considered further. One surrogate marker for each altered main cellular function was selected. These markers were considered in order to develop a multivariate logistic regression model predicting failure of treatment. Proteins found to be potentially informative in this analysis were automatically scored. Various multivariate models were fitted using step-up (forward) and step-down (backwards) variable selection and other heuristic procedures. By comparing the fits of these models, a final predictive logistic regression model was obtained. The final model estimates the odds ratio (OR), 95% confidence interval (CI) and significance (P) of each variable. General applicability of the model was tested by leave-one-out cross-validation. Different predictor models were found when using leave-one-out cross validation but these showed only small variations in the weight of each marker. Accuracy was also tested by the Receiver Operating Characteristic (ROC) curve, which estimates the discriminating ability of the model. The fit of the final model was assessed by the Hosmer-Lemeshow test. To demonstrate the predictive capacity of the model, patients were ranked according to this score and then divided into quartiles. Statistical analyses were performed using SPSS and R tools.