A Machine Learning Approach Identifies a 30-Gene Model That Predicts Sensitivity to Selinexor in Multiple Myeloma

Selinexor is the first FDA-approved drug for penta-refractory multiple myeloma (MM) patients, i.e. patients refractory to at least two proteasome inhibitors, two immunomodulatory drugs, and an anti-CD38 monoclonal antibody. Selinexor is an oral selective inhibitor of nuclear export (SINE), which specifically targets XPO1 (Exportin 1)-mediated nuclear export, leading to increased nuclear accumulation and activation of major tumor suppressor proteins, inhibition of NF-kB, and inducing selective apoptosis in cancer cells.

Several phase I and II clinical trials demonstrated evidence of anti-cancer activity of Selinexor in solid tumors, non-Hodgkin lymphoma, acute myeloid leukemia and MM (PMIDs: 29487219, 27458288, 28647672, 28468797, 29304833, 29381435). In the pivotal phase 2b STORM (Selinexor Treatment of Refractory Myeloma) trial, the combination of Selinexor with dexamethasone in patients with MM treated with prior bortezomib, carfilzomib, lenalidomide, pomalidomide and daratumumab (penta-exposed) and triple class refractory MM, has shown an overall response rate (ORR) of 26.2% and a clinical benefit rate of 39.3%, with a median progression-free survival (PFS) of 3.7 months and median overall survival (OR) of 8.6 months (Chari et al, NEJM in press 2019).

In order to identify biomarkers for selection of patients at higher likelihood of clinical benefit from Selinexor therapy and understand mechanisms of Selinexor resistance, we analyzed the transcriptome of CD138+ cells from bone marrow aspirates obtained prior to treatment from 54 patients enrolled in STORM. The raw data (fastq) was mapped using the tool STAR and gene-level annotated using the tool featureCounts. Gene expression counts were normalized using the variance stabilization transformation (VST) method implemented in the tool DESeq2, including potential confounding variables such as sequencing batch as covariates. Patients were split in two groups based on PFS = 120 days. We trained a linear Support Vector Machine (SVM) with L1 regularization to select features discriminating between responders (PFS > 120, n = 17) and non-responders (PFS < 120, n = 37). The L1 regularization reduces data dimensionality in large dataset and enables selection of the most relevant features. We trained the SVM using 80% (43) of the samples for feature selection and training with leave-one-out cross validation, and the remaining 20% (11) for model testing. The SVM achieved an AUC (Area Under ROC Curve) value of 0.71 on the test set. The model consisted of 30 genes which identified 3 patient clusters with significantly different PFS (Fig 1a, 1b). The cluster with poorer PFS was characterized by up-regulation of MAGE-A1.

MAGE-A is a cancer testis antigen that is aberrantly expressed in MM (PMID: 21565982) and has recently been described to have a critical role in chemotherapy resistance via regulation of BIM and p53 (Chari et al, Blood Advances 2017; Cho et al, ASH 2018). To test whether MAGE-A may contribute mechanistically to Selinexor resistance, we treated MM cells (H929 and RPMI8226) with Selinexor after depletion of MAGE-A by RNA interference. Our results show decreased viability in MAGE-A depleted cells as compared to non-target siRNA, suggesting that MAGE-A depletion increases sensitivity to Selinexor.

Taken together, our results provide a signature for selecting patients that may benefit from Selinexor therapy and provide insights into mechanisms of Selinexor resistance. We are currently performing whole-exome sequencing profiling of the patients in this study and will present the results at ASH 2019.

View largeDownload slide

View largeDownload slide

Close modal