Machine learning based biomarkers to predict CD8 infiltration in non-small cell lung cancers using CT imaging.

Abstract

Listen

Translate article

e21145 Background: Immune-checkpoint inhibitors (ICIs), specifically monoclonal antibodies targeting the programmed cell death protein-1 and programmed death-ligand 1 (PD-1 and PD-L1) have shown promise in the treatment of non-small cell lung cancer (NSCLC). Unfortunately, patients response to ICI is difficult to predict, with cancer stage, treatment line and regimen all affecting response rates. Identification of patients who will be “responsive” to ICI is of great importance to increase treatment efficacy. Current techniques used to identify responsive patients rely on PD-L1/PD-1 expression, microsatellite instability, tumor mutagenic burden and less commonly CD8+ infiltration, all of which are invasive and costly procedures. In this study we aim to develop a machine learning derived CT imaging biomarker that predicts CD8 status of tumors, allowing for non-invasive determination of ICI treatment efficacy. Methods: To study CD8 infiltration we extracted patients that included CT scans and CD8 transcript data from publicly available cohorts (TCGA-LUSC, TCGA-LUAD, NSCLC-Radiogenomics) to produce the Lung3 cohort (104 patients). Using this cohort, we trained (80 patients) and tested (24 patients) an AI model to predict tumor CD8 infiltration using patient CT scans as the sole input. CD8 expression (high and low) was characterized using transcriptomic quantification and served as surrogate for CD8+ infiltration, following previously described methods. Supervised and unsupervised models previously trained to identify lung features were trained to identify CD8 status using the CD8 expression data provided in the Lung3 dataset. Algorithm performance was tested on a randomly chosen subset of the cohort (24 patients) and compared against transcriptomic confirmed CD8 expression. AUC were calculated for supervised and unsupervised models. Results: AUC for the supervised and unsupervised models generated using the Lung3 cohort are shown below. Our results suggest that the derived supervised and unsupervised models could identify the CD8 infiltration status using only patient CT scans in the test dataset. Conclusions: The machine learning based models described here show encouraging results suggesting that CT imaging of NSCLC lung cancers contain sufficient information to predict the CD8 infiltration status of tumors. Although these results are just a proof of concept, the promise of a CT based non-invasive marker for tumor immune status is encouraging to help advance treatment efficacy in the ever-growing field of ICI therapy.