Charifa, Ahmad; Lam, Alfonso; Zhang, Hong; Ip, Andrew; Pecora, Andrew; Waintraub, Stanley; Graham, Deena; McNamara, Donna; Gutierrez, Martin; Jennis, Andrew; Sharma, Ipsa; Estella, Jeffrey; Ma, Wanlong; Goy, Andre; Albitar, Maher
Journal of Immunotherapy 47(1):p 10-15, January 2024. | DOI: 10.1097/CJI.0000000000000489
Abstract
Programmed death ligand-1 (PD-L1) immunohistochemistry (IHC) is routinely used to predict the clinical response to immune checkpoint inhibitors (ICIs); however, multiple assays and antibodies have been used. This study aimed to evaluate the potential of targeted transcriptome and artificial intelligence (AI) to determine PD-L1 RNA expression levels and predict the ICI response compared with traditional IHC. RNA from 396 solid tumors samples was sequenced using next-generation sequencing (NGS) with a targeted 1408-gene panel. RNA expression and PD-L1 IHC were assessed across a broad range of PD-L1 expression levels. AI was used to predict the PD-L1 status. PD-L1 RNA levels assessed by NGS demonstrated robust linearity across high and low expression ranges, and those assessed using NGS and IHC (tumor proportion score and tumor-infiltrating immune cells) had a similar pattern. RNA sequencing provided in-depth information on the tumor microenvironment and immune response, including CD19, CD22, CD8A, CTLA4, and PD-L2 expression status. Subanalyses showed a sustained correlation of mRNA expression with IHC (tumor proportion score and immune cells) across different solid tumor types. Machine learning showed high accuracy in predicting PD-L1 status, with the area under the curve varying between 0.83 and 0.91. Targeted transcriptome sequencing combined with AI is highly useful for predicting PD-L1 status. Measuring PD-L1 mRNA expression by NGS is comparable to measuring PD-L1 expression by IHC for predicting ICI response. RNA expression has the added advantages of being amenable to standardization and avoiding interpretation bias, along with an in-depth evaluation of the tumor microenvironment.
Immunohistochemistry (IHC) assays for the assessment of CD274 (programmed death ligand-1, PD-L1) expression in formalin-fixed paraffin-embedded (FFPE) tissues have been evaluated by multiple trials as a companion test with immune checkpoint inhibitors targeting PD-L1 and its main receptor (programmed cell death 1, PDCD1, or PD-1).1–5 These tests differ according to antibody clones, staining platforms, and scoring systems.6,7 These differences have led to an uncertainty in the general value of PD-L1 protein expression levels as a marker for different tumor types.8–10 Although a good concordance for 3 of the 5 antibody clones used in PD-L1 evaluation has been published,11,12 the interchanging assays and cutoffs could cause the misclassification of PD-L1 status in some cases. In addition, the use of FFPE tissue specimens with different fixation and storage methods may be a source of unexpected results for adequate PD-L1 antigen retrieval, potentially increasing the heterogeneity of the IHC intensity. All of these factors make the prediction of the patients’ clinical response to immune checkpoint inhibitors challenging using PD-L1 IHC.13,14
Recently, RNA profiling has become highly useful for providing information on the tumor microenvironment and immune response.15–23 The use of next-generation sequencing (NGS) to analyze RNA offers a high level of specificity and sensitivity for the simultaneous measurement of several targets to further elucidate the biology of tumors and predict the efficacy of various therapeutic approaches.15–19 Targeted RNA sequencing of various tissue samples has minimal input requirements, has the potential to be much more cost-effective than IHC methods,24 and allows sequencing at a deeper level for better quantification of low-level expressor genes. RNA sequencing and quantification using NGS are more reliable and reproducible than older techniques such as microarrays or PCR-based RNA quantification.17,18,25,26 The purpose of the paper is to demonstrate the reliability of RNAseq data to predict PD-L1 expression and the feasibility of developing a model system to predict specific PD-L1 cutoff points by using RNAseq data in machine learning systems.
KEYWORDS
immune checkpoint inhibitors, PD-L1, PD-L2, biomarker, machine learning, artificial intelligence