Inferring PDL-1 status from H&E images using digital pathology to identify patients responsive to anti-PD(L)-1 immuno-oncology therapy for bladder cancer trials.

4579 Background: Programmed death ligand-1 (PDL-1) expression has been established as a biomarker for response to immune checkpoint inhibitors in bladder cancer. Targeted therapies that inhibit the PDL-1 pathway improve therapeutic response, but immunohistochemistry (IHC) testing is required to assess PDL-1 expression. In order to ease identification of patients who have high potential to respond to anti-PD(L)-1 therapies, we explore the use of a computer vision algorithm to infer PDL-1 expression from patients routine H&E images. Methods: We pretrained a Foundational Model on ~55k unlabeled WSIs from various sources (multiple scanners, hospitals, diseases, tissues..). We then finetuned it on a train set (n=1546) of PDL-1 labeled H&E images to infer expression. Labels were defined as PDL1-High or PDL1-Low based on the most sensible threshold for the various included antibodies (e.g., TPS>=10% on 22C3 antibody, CPS>=1% on 28-8 antibody, etc..). We applied this model to WSIs from 1) a holdout split (n=388) and 2) an independent dataset (n=93) to evaluate the performance at inferring PDL1-High vs. PDL1-Low, quantified by the Area Under ROC Curve (AUC). We then applied the model to a third dataset (n=27) containing WSIs from bladder cancer patients taken prior to start of anti-PD(L)-1 treatment (i.e., pembrolizumab) and evaluated treatment response (quantified by overall survival) for the two stratified groups based on the model outputs (PDL1-High vs. PDL1-Low). Results were compared to patients stratified based on IHC pathology readouts (i.e., TPS>=10% vs. TPS<10%). Results: For PD-L1 status inference, the model achieved AUC=0.82 and AUC=0.8 on holdout and independent datasets respectively. The table below shows survival analysis results upon pembrolizumab treatment (n=27) when stratifying patients based on the algorithm (AI-based PDL1-High vs. PDL1-Low) compared to IHC-based stratification. Note that the survival probability for PDL1-High patients is higher than PDL1-Low patients. Conclusions: We developed a computer vision algorithm to infer PDL-1 expression from routine H&E images. The algorithm shows strong performance at classifying PDL-1 status and proficiency in inferring outcomes to anti-PD(L)-1 therapy in a small bladder cancer data set. This tool represents a novel, rapid means to assess whether a patient may respond to anti-PD(L)-1 therapies from common H&E-stained images and has the potential to guide patient care decisions in clinic. Table: see text