Artificial intelligence–based biomarkers of response to immunotherapy in patients with non–small-cell lung cancer considering previous lines of treatment.

Abstract

e14652 Background: Lung cancer is one of the most prevalent types of cancer worldwide: 250,000 new cases are diagnosed yearly in the US, amongst which 80% are non-small cell lung cancers (NSCLC). The optimal biomarker to select patients diagnosed with advanced disease and benefiting the most from immunotherapy is still to be identified. Herein, we used KEM (Knowledge Extraction and Management) explainable Artificial Intelligence (xAI) as a tool that systematically extracts and evaluates all association rules between all variables in a database, thus enabling the identification of subgroups of patients with advanced NSCLC treated with immunotherapy with higher chances of overall survival in the NIVOBIO cohort (Foy et al. Eur J Cancer 2023). We aimed to identify biomarkers linked to optimum response to immunotherapy considering previous lines of therapies. Methods: Data was retrieved from GEO warehouse (GSE161537) and aggregated into a consolidated database totaling 82 patients and 2,568 variables. A two-step analysis plan was then performed. First step relied on KEM xAI: rules between gene expression, number of previous treatment lines and survival were explored: 51,306 rules were generated. 19 rules involving 3 genes were retained using metrics such as Support (number of examples), Confidence (conditional probability) and Lift (relative probability) and focusing on genes with a consistent signal across rules. Second, Cox proportional hazards models were generated using these genes to test the interaction between their expression and the number of previous treatment lines the patient had undergone. Genes with a significant log-rank and interaction p-value were retained. Results: Our analysis identified 2 genes that were significantly associated with overall survival and previous lines of treatment: SOS2 (p < 0.001) and LIFR (p = 0.020) high expression was associated with improved survival among patients with at least two previous treatment lines, whereas it was associated with poor survival for other patients. Consistently, low expression values of these genes were associated with poor survival in patients with less than two previous treatment lines. SOS2 and LIFR gene expression was identified as having a significant interaction with the number of previous treatment lines in the Cox model with a hazard ratio of 0.08 and 0.39 respectively. Conclusions: Our analysis enabled the identification of biomarkers not obviously related to the immune microenvironment and associated with an improved or poor survival depending on when immunotherapy was administered. These findings are laying promising foundations for the development of dynamic biomarkers and their potential translation into therapeutic recommendations after further validation.