Robust Prediction of Immune Checkpoint Inhibition Therapy for Non-Small Cell Lung Cancer

Abstract

BackgroundThe development of immune checkpoint inhibitors (ICIs) is a revolutionary milestone in the field of immune-oncology. However, the low response rate is the major problem of ICI treatment. The recent studies showed that response rate to single-agent programmed cell death protein 1 (PD-1)/programmed cell death-ligand 1 (PD-L1) inhibition in unselected non-small cell lung cancer (NSCLC) patients is 25% so that researchers defined several biomarkers to predict the response of immunotherapy in ICIs treatment. Common biomarkers like tumor mutational burden (TMB) and PD-L1 expression have several limitations, such as low accuracy and inadequately validated cutoff value.MethodsTwo published and an unpublished ICIs treatment NSCLC cohorts with 129 patients were collected and divided into a training cohort (n = 53), a validation cohort (n = 22), and two independent test cohorts (n = 34 and n = 20). We identified six immune-related pathways whose mutational status was significantly associated with overall survival after ICIs treatment. Then these pathways mutational status combined with TMB, PD-L1 expression and intratumor heterogeneity were incorporated to build a Bayesian-regularization neural networks (BRNN) model to predict the ICIs treatment response.ResultsWe firstly proved that TMB, PD-L1, and mutant-allele tumor heterogeneity (MATH) were independent biomarkers. The survival analysis of six immune-related pathways revealed the mutational status could distinguish overall survival after ICIs treatment. When predicting immunotherapy efficacy, the overall accuracy of area under curve (AUC) in validation cohort reaches 0.85, outperforming previous predictors in either sensitivity or specificity. And the AUC in two independent test cohorts reach 0.74 and 0.80.ConclusionWe developed a pathway-model that could predict the efficacy of ICIs in NSCLC patients. Our study made a significant contribution to solving the low prediction accuracy of immunotherapy of single biomarker. With the accumulation of larger data sets, further studies are warranted to refine the predictive performance of the approach.

Systematic pan-cancer analysis showed that RAD51AP1 was associated with immune microenvironment, tumor stemness, and prognosis

“…This hypothesis may also explain another finding in our study that RAD51AP1 expression was positively correlated with TMB in various tumors ( Figure 11B ). Moreover, TMB and MATH are effective biomarkers of immune checkpoint inhibitors (ICIs) treatment, and their increment in the tumor may indicate better ICIs therapy efficacy ( Gao et al, 2020 ; Jiang et al, 2021 ). Interestingly, we also found that cancer with higher RAD51AP1 expression often presented higher MSI ( Figure 11A ), which is positively correlated to ICIs therapy efficacy in CRC and other cancers ( Dudley et al, 2016 ).…”

Impact of pharmacodynamic biomarkers in immuno-oncology phase 1 clinical trials

“…Given the limited impact of current single-plexed PD assays and the complexity of tumour cell-host immunity interactions, bioinformatic models which integrate different tissue-, blood-and imaging-PD results may be needed to overcome the performance of single analytes and improve dose-finding strategies for specific IO agents [29]. Recent studies utilising an integrated biomarker approach to predict tumour response to IO agents have demonstrated improved performance over single independent biomarkers [30,31]. The future of PD analyses will likely rely on the joint use of different PD biomarkers depending on the expected biologic activity of the IO agent.…”

Bayesian network enables interpretable and state-of-the-art prediction of immunotherapy responses in cancer patients

“…Most prior attempts to predict immunotherapy responses have used ML-based approaches [14][15][16] , which are complex "black-box" systems that cannot handle missing data. As input, they require all of the necessary clinical and molecular information to be provided.…”

“…In recent years, machine learning (ML)-based methods have been applied to unravel the factors determining the efficacy of ICI treatment for NSCLC. For one example, the AUC of a neural network model that integrated several factors (TMB, PD-L1 score, mutant-allele tumor heterogeneity, and immune-related pathways) was as high as 0.80 in a test cohort 14 . Another study integrated PD-L1 score and CT images achieved an AUC of 0.76 15 .…”