Concordance of multispectral immunofluorescence (mIF) with programmed death ligand 1 (PD-L1) and stromal tumor infiltrating lymphocyte (sTILs) clinical assays in early-stage breast cancer (BC).

Abstract

e12589 Background: There is growing need for surrogate immune-based biomarkers in BC to enhance prediction accuracy and to facilitate comparisons of therapeutic activity in the context of clinical trials. We report mIF as a high-dimensional biomarker that is highly concordant with both PD-L1 (SP142) and sTIL clinical assays (H&E), but with greater precision to quantify on-treatment pharmacodynamic effects. We also demonstrate how statistical modeling can be employed to improve precision of PD-L1/sTIL estimates, and reduce sample size requirements for biomarkers-driven comparative clinical trials. Methods: Samples were obtained from a recently-reported pre-surgical cytokine-based phase Ib immunotherapy (IRX-2) clinical trial (NCT02950259). Pre-treatment and resection tissues (n = 30) were analyzed for sTIL score, PD-L1 status, and mIF (PerkinElmer Vectra) using antibodies to quantify BC cells (cytokeratin), immune cells (CD3, CD8, CD163, FOXP3), and per-cell PD-L1 expression. Mixed effects regression models were employed to correct for intratumoral heterogeneity and enhance statistical precision. Monte Carlo simulations were employed to evaluate whether regression modeling could reduce sample size required to detect treatment-related increases in sTIL/PD-L1 in pre-surgical clinical trials. Results: Results: mIF sTIL estimates were highly correlated with H&E sTIL score (pre-treatment samples: r = 0.59, p > 0.001; excisional samples: r = 0.63, p > 0. 001). mIF PD-L1 estimates were highly correlated with the PDL1 SP142 assay, with average densities increasing concordantly according to SP142 IC category (IC0 v. IC1, p = 0.04; IC0 v. IC2/3, p = 0.003). Our regression modeling approach dramatically improved power to demonstrate treatment-related changes in sTIL and PD-L1. The estimated sample size requirement to show the observed effect on sTIL score (1.80 fold change, CV 0.83) was n = 25 with conventional t-testing of means, versus n = 13 with regression modeling. Conclusions: mIF is concordant with SP142 and sTIL prognostic assays, but with increased precision to quantify treatment related changes. Regression modeling can dramatically improve the utility of mIF as a surrogate immune-based biomarker. Our approach is being utilized in ongoing phase II studies to compare the activity of pembrolizumab +/- locoregional cytokines (IRX-2) in triple negative BC, and paclitaxel/trastuzumab +/- pembrolizumab +/- pertuzumab in HER2-positive BC.