A Phase 2 Study of the Efficacy and Safety of Oral Selinexor in Recurrent Glioblastoma

In a post-hoc exploratory analysis of the KING trial to seek molecular markers of outcome, RNAseq was performed on resected tumor specimens at the time of diagnostic surgery before the recurrence from 57 study patients from all arms with adequate selinexor exposure and evidence of either clinical benefit or resistance defined above. RNAseq data were used to infer the activity for 6,203 master regulator proteins using the VIPER algorithm. Overall design: The sequenced specimens were split into a discovery set of 7 clear responders (BOD of CR or PR) compared to 23 resistors (BOR of PD) and a validation set of 11 with other evidence of clinical benefit (BOR of durable SD >140 days) and 11 non-responders (BOR of PD, or BOR non-durable SD <100 days). An ensemble of five different machine learning algorithms was used to generate an integrated predictive model for selinexor response in GBM. This model was based on the VIPER-inferred activity for three proteins that were activated in the responders compared to the non-responders in the discovery set, ZC3H12A (false discovery rate P-value [FDR]=6.45 x 10^-11), RAB43 (FDR=3.81 x 10^-10), and SOCS3 (FDR=3.16 x 10^-9). The model achieved an integrated area under the receiver operating characteristic (ROC) curve of 0.88 (p<0.05, permutation test) for a Leave-one-out cross-validation analysis in the discovery set and correctly predicted 9 of 11 patients classified as experiencing clinical benefit and 7 of 11 patients classified as non-responders in the validation set (ROC-AUC = 0.67).