Machine learning identifies clinical tumor mutation landscape pathways of resistance to checkpoint inhibitor therapy in NSCLC.

Background: Immune checkpoint inhibitors (CPIs) have revolutionized cancer therapy for several tumor indications. However, a substantial fraction of patients treated with CPIs derive no benefit or have short-lived responses to CPI therapy. Identifying patients who are most likely to benefit from CPIs and deciphering resistance mechanisms is therefore essential for developing adjunct treatments that can abrogate tumor resistance.

Patients and methods: In this study, we used a machine learning approach that used the US-based nationwide de-identified Flatiron Health and Foundation Medicine non-small cell lung carcinoma (NSCLC) clinico-genomic database to identify genomic markers that predict clinical responses to CPI therapy. In total, we analyzed data from 4,433 patients with NSCLC.

Results: Analysis of pretreatment genomic data from 1,511 patients with NSCLC identified. Of the 36 genomic signatures identified, 33 exhibited strong predictive capacity for CPI response (n=1150) compared with chemotherapy response (n=361), while three signatures were prognostic. These 36 genetic signatures had in common a core set of four genes (BRAF, BRIP1, FGF10, and FLT1). Interestingly, we observed that some (n=19) of the genes in the signatures (eg, TP53, EZH2, KEAP1 and FGFR2) had alternative mutations with contrasting clinical outcomes to CPI therapy. Finally, the genetic signatures revealed multiple biological pathways involved in CPI response, including MAPK, PDGF, IL-6 and EGFR signaling.

Conclusions: In summary, we found several genomic markers and pathways that provide insight into biological mechanisms affecting response to CPI therapy. The analyses identified novel targets and biomarkers that have the potential to provide candidates for combination therapies or patient enrichment strategies, which could increase response rates to CPI therapy in patients with NSCLC.