Abstract Background: Treatment-induced anti-tumor immune response may be reflected in peripheral blood leukocyte composition and shifts. The relative abundance of immune cell subsets captured by neutrophil-to-lymphocyte ratio (NLR) is a promising biomarker of response to immune checkpoint inhibition (ICI). However early dynamics and integration with circulating tumor DNA (ctDNA) kinetics during ICI have not been comprehensively studied. Methods: We evaluated the predictive role of early dynamic changes of peripheral immune cell subsets in 239 ICI treated metastatic non-small cell lung cancers (mNSCLC) and combined those with ctDNA dynamics in 18 early stage NSCLCs (esNSCLC) treated with neoadjuvant ICI. Blood cell counts were analyzed at baseline, 4 weeks on treatment, and at first radiographic follow up (mNSCLC) or preoperatively (esNSCLC). mNSCLC patients without progression at 6 months were considered to have durable clinical benefit (DCB). We employed eXtreme Gradient Boosting, a decision-tree based machine learning algorithm, to integrate baseline values and early changes in immune cell subsets. To reduce modeling overfitting, we trained an ensemble of models, incorporating 10-fold cross-validation that included feature selection (training n=171) and tested the model in an unseen independent set (n=68). For esNSCLCs, major pathologic response (MPR; ≤10% residual tumor; RT) was determined. We performed Targeted Error Correction next generation sequencing on 60 serial plasma and matched leukocyte DNA samples to assess ctDNA clonal dynamics. Changes in immune cell subsets were correlated with RT, MPR, progression-free (PFS) and overall survival (OS). Results: Our machine learning integrative model predicted DCB with an area under the ROC curve (AUC) of 0.96 for training, 0.72 for cross-validation testing and 0.74 for unseen testing datasets. Feature importance analysis revealed that NLR at 4 weeks, first radiographic follow up, and relative change in NLR from baseline were the strongest predictors of outcome together with relative eosinophil and lymphocyte count. Our model's performance was superior to TMB (AUC=0.52) or PD-L1 (AUC=0.54). For esNSCLCs, change in NLR at 4 weeks after ICI initiation was predictive of tumor regression (p=0.02), such that those with MPR showed significant decrease in NLR. This was also predictive of PFS (log rank p=0.004) and OS (log rank p=0.02). Furthermore, higher eosinophil count at 4 weeks was correlated with MPR (p=0.006) and decreased RT (p=0.006). Immune cell subset kinetics were concordant with ctDNA clearance in all but two patients and importantly NLR dynamics at 4 weeks were predictive of MPR even when ctDNA was undetectable. Conclusions: Early changes in peripheral immune cell subsets together with ctDNA are reflective of anti-tumor immune response during ICI and may more accurately predict ICI response than currently used biomarkers. Citation Format: Michael Hwang, Jenna Canzoniero, Samuel Rosner, Guangfan Zhang, Mara Lanis, Lamia Rhymee, Alexandria Curry, Gavin Pereira, Kristen Marrone, Joshua Reuss, Jarushka Naidoo, Christine Hann, Vincent Lam, Benjamin Levy, David Ettinger, Patrick Forde, Julie Brahmer, Victor Velculescu, Tanguy Seiwert, Valsamo Anagnostou. Early dynamics in peripheral blood immune cell subsets and ctDNA are predictive of outcome to immunotherapy [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 27.
Read full abstract