Machine learning identifies immune-based biomarkers that predict efficacy of anti-angiogenesis-based therapies in advanced lung cancer.

Abstract

Background: The anti-angiogenic drugs showed remarkable efficacy in the treatment of lung cancer. Nonetheless, the potential roles of the intra-tumoral immune cell abundances and peripheral blood immunological features in prognosis prediction of patients with advanced lung cancer receiving anti-angiogenesis-based therapies remain unknown. In this study, we aimed to develop an immune-based model for early identification of patients with advanced lung cancer who would benefit from anti-angiogenesis-based therapies.

Methods: We assembled the real-world cohort of 1058 stage III-IV lung cancer patients receiving the anti-angiogenesis-based therapies. We comprehensively evaluated the tumor immune microenvironment characterizations (CD4, CD8, CD68, FOXP3, and PD-L1) by multiplex immunofluorescence (mIF), as well as calculated the systemic inflammatory index by flow cytometry and medical record review. Based on the light gradient boosting machine (LightGBM) algorithm, a machine-learning model with meaningful parameters was developed and validated in real-world populations.

Results: In the first-line anti-angiogenic therapy plus chemotherapy cohort (n = 385), the intra-tumoral proportion of CD68 + Macrophages and several circulating inflammatory indexes were significantly related to drug response (p < 0.05). Further, neutrophil to lymphocyte ratio (NLR), monocyte to lymphocyte ratio (MLR), the systemic inflammation response index (SIRI), and myeloid to lymphoid ratio (M:L) were identified to construct the non-invasive prediction model with high predictive performance (AUC: 0.799 for treatment response and 0.7006-0.915 for progression-free survival (PFS)). Additionally, based on the unsupervised hierarchical clustering results, the circulating cluster 3 with the highest levels of NLR, MLR, SIRI, and M: L had the worst PFS with the first-line anti-angiogenic therapy plus chemotherapy compared to other circulating clusters (2.5 months, 95 % confidence interval 2.3-2.7 vs. 6.0-9.7 months, 95 % confidence interval 4.9-11.1, p < 0.01). The predictive power of the machine-learning model in PFS was also validated in the anti-angiogenic therapy plus immunotherapy cohort (n = 103), the anti-angiogenic monotherapy cohort (n = 284), and the second-line anti-angiogenic therapy plus chemotherapy cohort (n = 286).

Conclusions: Integrating pre-treatment circulating inflammatory biomarkers could non-invasively and early forecast clinical outcomes for anti-angiogenic response in lung cancer. The immune-based prognostic model is a promising tool to reflect systemic inflammatory status and predict clinical prognosis for anti-angiogenic treatment in patients with stage III-IV lung cancer.