Interplatform and cross-cohort meta-analysis of tumor transcriptomics to predict non-small cell lung cancer c-Met protein overexpression.

Abstract

e15164 Background: Antibody-drug conjugate (ADC) antigen target expression is a key predictive biomarker of patients’ (pts) treatment response. Conventional immunohistochemistry (IHC) approaches are often challenged by assay tissue specificity and platform compatibility. The rapidly expanding tumor whole transcriptomics sequencing (WTS) in real-world oncology data (RWD) provides tremendous opportunities to accelerate identification of ADC target-positive pts. To demonstrate interplatform and cross-cohort robustness in identifying pts with high antigen expression, we present a meta-analysis across clinical and RWD non-small cell lung cancer (NSCLC) cohorts, focusing on c-Met protein as the ADC target. Methods: c-Met protein expression was evaluated by IHC (SP44 Ab clone [Roche]). Whole transcriptomics sequencing data from multiple pt cohorts were profiled by different sequencing platforms. A RW City of Hope NSCLC cohort (N = 230) and a LUMINOSITY trial (NCT03539536) cohort (N = 314) were profiled using the Personalis ImmunoID NeXT platform. A RW c-Met METPRO prognosis cohort was sequenced in 2 subgroups by Caris Tumor Seek Hybrid WTS (N = 258) and non-hybrid WTS (N = 242), correspondingly. Log2-normalized gene expression values from WTS were compared with IHC-defined cutoffs for c-Met protein overexpression (OE). A machine learning (ML) model was developed using differential expression features between c-Met protein– overexpressing and c-Met–low/normal pt samples. Results: Cross-cohort analysis of 1044 NSCLC cases shows MET RNA, measured as scaled MET log2 transcripts per million, significantly correlates with c-Met protein OE (IHC scoring of 3+ > 25%), with the area under the ROC (auROC) ranging from 0.802 to 0.952. The auROC is maintained when randomly subsampling the cohorts to 1/10 of the original size (range: 0.806, 0.968) confirming RNA is predictive. The robustness of RNA predictive performance across WTS platforms and in both large and small pt cohorts demonstrates the feasibility of using WTS to identify pts with c-Met protein OE in RWD clinical genomics database. To further improve performance we developed a regression-based ML model trained with City of Hope cohort, which achieved an auROC of 0.978. The ML model was directly validated across independent testing datasets to suggest strong predictiveness of c-Met OE including a cohort sequenced with the same WTS platform (LUMINOSITY, auROC = 0.86) and also on a cohort sequenced with a different WTS platform (Caris METPRO, auROC = 0.95). Conclusions: Our meta-analysis demonstrates the interplatform and cross-cohort stability and robust performance of predicting c-Met protein OE using RNA expression obtained from WTS. The results highlight the clinical practicality of using existing oncology RWD databases to accelerate the identification of antigen-positive pts.