Study Background
Patients with advanced non-small cell lung carcinoma (NSCLC) have benefited from improved first-line treatments, but ≥ 80% of patients who progress do not respond to subsequent lines of therapy, and the median progression free survival is only 2-4 months. Treatment options for those patients are limited, and drugs are being developed against new targets, such as KRAS. In a Phase 2 clinical trial, objective response to sotorasib (Amgen, Thousand Oaks, CA) was 37.1%, but co-occurring mutations in KEAP1 reduced the response rate to 14% in an exploratory analysis.
Approximately 20% of LUAD tumors have KEAP1 mutations. KEAP1MUT patients may therefore benefit from alternative treatments. Here we present machine-learning (ML) algorithms that can predict KEAP1MUT directly from routinely available hematoxylin and eosin (H&E) stained tissue samples. ML-based genomic assessment of the H&E samples that are routinely collected from each patient has potential to decrease the cost and turnaround time of genotyping because it requires no additional tissue collection or processing for genomic sequencing. This could increase the efficiency of clinical development program and may allow patients to receive the most effective therapies sooner.
