Prediction of Early Disease Progression at 24 Months (POD24) Using Pre-Treatment Biopsies from Patients with Follicular Lymphoma (FL) Treated with Immunochemotherapy (IC)

Background: Gene expression profiling has been successfully used to interrogate FL biology related to overall survival alone or in combination with clinical parameters and mutation profiles such as the M7-FLIPI. However, none of these has ever been specifically trained against the most reproducible and pertinent predictor of excess mortality: early progression of disease by 24 months (POD24). Nor has any assay been validated on a clinical platform for testing patients in a CAP-CLIA environment. We therefore undertook a discovery analysis and also investigated published gene signatures to determine the best method for upfront prediction of POD24 with clinical progression (not transformation).

Methods: For model generation, we assembled and harmonized gene expression data from frozen tissues from 377 total FL cases treated with immunochemotherapy (IC) from the Lymphoma/Leukemia Molecular Profiling Project (LLMPP; N=80, unpublished, Affymetrix), British Columbia (N=137, Silva et al, 2018, DASL Illumina), and PRIMA cohorts (N=160, S. Huet et al, 2018, Affymetrix) and performed in silico analysis to identify the most promising candidate genes to predict POD24. In addition to the identified genes, we also included all “IR1/IR2” genes (S. Dave 2004) and “PRIMA23” (Huet) genes and other published genes of interest totaling 144 genes related to tumor, microenvironment, and housekeeping functions. We used these 144 genes as the basis for an nCounter code set (NanoString, Seattle, WA), and then evaluated their expression in formalin-fixed, paraffin-embedded (FFPE) tissues from a training set of 57 matching cases from the LLMPP cohort and an independent set of 59 independent samples from the University of Iowa/Mayo Clinic SPORE Molecular Epidemiology Resource (MER). Comparing the FFPE NanoString results from the LLMPP training set to their matched fresh frozen Affymetrix results identified, 48 of the 144 candidate genes for which the two platforms showed agreement with correlation > 0.5). Multiple prediction architectures were compared using these 48 genes and optimized via cross-validation on this training set to develop a final locked down POD24 predictor (“CCP-32”) based on a linear combination of the expression of 32 selected informative genes and an additional 19 housekeeping genes. Additionally, for comparison, the Huet profiling scores were calculated on the LLMPP training samples and a cut-point, optimized to select patients with POD24 as a dichotomous endpoint. These two predictors were then applied to an independent set of 174 FFPE samples from the MER population-based cohort. Samples with very low signal in the normalization genes were excluded (8/116 training and 17/174 validation). Samples from patients that were event free until they experienced transformation or were lost to follow-up in less than 24 months after IC were treated as censored and were thus excluded (2/116 training, 6/174 validation).

Results: When applied to the MER FFPE validation set, the CCP-32 predictor placed 30% in a poor prognosis group with an observed POD24 rate of 56% while the remaining 70% samples had an observed 18% POD24 – an enrichment that was statistically significant (p=2.9x10^-5). A training-optimized cut of the Huet profiling score performed similarly on the validation set with 44% placed in the poor prognosis group, with an observed POD24 rate of 48%, while the remaining 56% had an observed POD24 of 15% (p=1.5x10^-5)

Conclusions: Starting with an agnostic in silico analysis of data from frozen FL tissue to identify predictive genes, also including known gene sets pertinent to FL, we first identified genes that correlated best in FFPE tissues using digital expression profiling to develop an algorithm specifically trained to detect POD24 due to progression (not transformation) on a clinical platform using routinely processed samples. A novel predictor and the Huet profiling score were each able to detect a biological subset that was heavily enriched in the approximately 30-44% of FL patients at risk for POD24. Additional cohort analyses are underway. Integration of gene mutations and other genetic factors will also be explored to determine whether these can improve the predictive value with a long-term goal of creating a combined assay that could be suitable for use in a clinical laboratory to select high risk patients for clinical trials.