Non-invasive Mantle Cell Lymphoma risk stratification by inference of tumor proliferation from cfDNA

Abstract

Abstract Background: Variability in tumor proliferation is a key prognostic factor in Mantle Cell Lymphoma (MCL). However, its measurement by gene expression profiling of frozen tumor tissue specimens (Rosenwald et al. Cancer Cell) or Ki-67 index of fixed tumor tissues (Katzenberger et al.Blood) have each faced limitations. The MCL35 gene expression signature is a validated independent prognostic biomarker for risk stratification in MCL, relying on tumor RNA profiling to accurately quantify variability in tumor proliferation on routinely fixed tissues (Scott et al. JCO). However, its clinical use can be limited by the requirement for invasive lymph node or tumor tissue biopsies, their attendant risks and associated costs, as well as specific technical challenges when profiling bone marrow specimens and leukemic MCL samples. While liquid biopsies might offer a promising alternative, no approach has yet bridged the molecular gap between blood- versus tissue-derived prognostic signatures for faithfully measuring MCL-proliferation associated risk noninvasively. Here, we tackle this challenge. Methods: We profiled 161 total samples from 117 patients from the Phase 3 randomized LYMA trial (Le Gouill et al. NEJM) relying on R-DHAP/ASCT for first remission induction in treatment-naïve MCL. Cases were selected based on availability of baseline paired blood and/or tissue specimens at diagnosis. We profiled 88 matched-tumor plasma specimens from 44 patients; for 73 patients, either isolated tumor specimens (n=56) or plasma specimens (n=17) were available. Tumor specimens (n=100; 68% FFPE, 32% Frozen) were expression profiled by RNA-Seq. Plasma specimens were profiled by EPIC-Seq (Esfahani et al. Nat Biotech) using a customized lymphoma-specific panel including key proliferation associated MCL genes (Mutter et al. Blood). We developed a novel deep generative ML model to better infer transcriptomic expression from cfDNA fragmentomic features. Trained on both paired and unpaired tumor and plasma samples from MCL patients (60% training, 40% test), this new generative ML model (termed iEPIC) learns a domain transfer function aiming to reconstruct RNA-like profiles from raw blood plasma EPIC-Seq Promoter Fragmentation Entropy (PFE) measurements. We computed MCL35 scores from these predicted profiles using optimized coefficients. Model performance was benchmarked against ground-truth tumor-derived MCL35 scores and evaluated for clinical validity in survival stratification and prediction of early progression (POD24). Results: When considering tumor gene expression profiles by RNA-Seq (n=100), higher MCL35 scores were significantly associated with blastoid morphology (p<0.001) and tumor Ki67% index (p<0.001), as expected. Higher tumor MCL35 scores were also significantly associated with both inferior PFS and OS, whether as a continuous (p=0.007 [PFS], p=0.004 [OS]) or categorical (p=0.01 [PFS], p=0.01 [OS]) variable. Multivariate Cox models confirmed the superior prognostic value of MCL35 after adjusting for other key prognostic variables including MIPI and Ki67. When considering inferred expression profiles from plasma cfDNA using EPIC-Seq, the new iEPIC GEP model significantly improved single-gene level correlations between noninvasive plasma measurements from cfDNA and invasive tumor tissue measurements from RNA. This improvement included correlation gains in key mitotic proliferation genes, such as MKI67 (p<0.01), TOP2A (p<0.01), and FOXM1 (p<0.01). When extended to the full MCL35 proliferation signature, iEPIC achieved ~0.7 Pearson correlation against tumor RNA-Seq, allowing improvements in precision, recall, and F1 score of noninvasive measurements by ~23%. Our noninvasive iEPIC GEP model also significantly stratified patients across Ki67 groups (p < 0.02), similar to tumor RNA-Seq (p<0.002). For prediction of POD24 status, this noninvasive iEPIC GEP model achieved reasonably high performance (AUC=0.73), as compared with invasive MCL35 from tumor RNA-Seq (AUC=0.83). Conclusions: We describe a novel non-invasive approach for computing MCL tumor proliferation score directly from blood-based cfDNA, toward enabling risk stratification in MCL without tissue biopsies. This model maintains gene-level biological fidelity and could be deployed in settings where tissue is unavailable or serial monitoring is required. We expect that this innovation could yield a significant advance toward liquid biopsy-driven precision medicine for patients with MCL.