Predicting response to durvalumab and olaparib in mTNBC using a novel PD-L1 expression biomarker: A comparison with PD-L1 IHC.

Abstract

e14612 Background: Immunohistochemistry (IHC) of programmed death ligand-1 (PD-L1) expression is widely used to predict response to Immune Checkpoint Inhibition (ICI). Nevertheless, even after stratification with PD-L1, the clinical benefit of ICI treatment in metastatic Triple-Negative Breast Cancer (mTNBC) remains widely varied. We have previously described DeepPT - a deep learning model to infer gene expression from H&E slides [Nature Cancer, 2024]. Here we study the ability of the inferred expression of PD-L1 to predict response, comparing it to PD-L1 IHC and to biomarkers based on measured transcriptomics. Methods: We studied a cohort of 19 mTNBC patients treated with Olaparib and Durvalumab. For each case, RNAseq data and 1-3 H&E slides were available from a pre-treatment biopsy, along with the IHC-based PD-L1 measure. An IHC classification was determined by assigning positive status for PD-L1 > 10%/25% for 22C3/SP263, respectively. In addition, we used DeepPT to infer the whole transcriptome, and specifically PD-L1 expression from the H&E slides. We analyzed the ability of inferred, as well as RNAseq-based PD-L1 expression to predict objective response (RECIST 1.1) and compared these to the predictive value of PD-L1 IHC. Results: The overall response rate in the cohort was 58%. When examining PPV and sensitivity for predicting response using pre-treatment samples, the expression of PD-L1, as measured by RNAseq and as inferred by DeepPT were highly predictive: 80%/90% PPV and 80%/82% sensitivity, with F1 = 0.8 and 0.86 for measured PD-L1 expression and inferred PD-L1 expression, respectively. In comparison, the IHC PD-L1 status had 66.6% PPV and 18.2% sensitivity (F1 = 0.29). The PPV of the inferred PD-L1 expression is significantly higher than the PPV of IHC (p = 0.015 one sided proportion test), while the PPV of measured expression was not (p = 0.12). In addition, the inferred expression of PD-L1 has the desired characteristic of monotonicity, where partial responders have higher inferred PD-L1 values than those with stable disease, which in turn have higher values than those with progressive disease. This is not the case for IHC, where the percentage of patients with positive IHC does not correlate with RECIST responses. Conclusions: We show that PD-L1 mRNA expression, inferred from H&E slides, is a better and more accurate predictor for response to Durvalumab + Olaparib than PD-L1 IHC in the study cohort. While this result was obtained in a relatively small cohort, its further testing and validation could have significant clinical potential for advancing readily available and cost-effective response biomarkers, further democratizing precision oncology.