Weakly Supervised Deep Learning Predicts Immunotherapy Response in Solid Tumors Based on PD-L1 Expression

Area of Science:

• Computational pathology
• Biomarker discovery
• Cancer immunotherapy

Background:

• Programmed death-ligand 1 (PD-L1) immunohistochemistry (IHC) is crucial for predicting response to cancer immunotherapy.
• Current quantification methods (manual and computer-assisted) have limitations in reproducibility and predictive performance.
• Accurate PD-L1 assessment is vital for effective patient stratification.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for direct, end-to-end prediction of PD-L1 status from raw IHC image data.
• To assess the model's ability to predict response to immune checkpoint inhibitors (ICI).
• To compare the DL model's performance with traditional PD-L1 quantification scores (Tumor Proportion Score [TPS] and Combined Positive Score [CPS]).

Main Methods:

• A weakly supervised DL model was trained on PD-L1 stained non-small cell lung cancer (NSCLC) slides (MSK cohort).
• The model was validated on a pan-cancer cohort (VHIO cohort).
• The model predicted PD-L1 expression and response to ICIs, with performance compared against TPS and CPS.

Main Results:

• The DL model demonstrated strong performance in predicting PD-L1 expression (AUC 0.88 in NSCLC, 0.80 in pan-cancer).
• Predicted PD-L1 status showed a significantly improved association with ICI response (HR 1.5, P=0.049) compared to TPS (HR 1.4, P=0.082) and CPS (HR 1.2, P=0.386).
• Explainability analysis revealed the model integrates both staining intensity and morphological factors.

Conclusions:

• End-to-end weakly supervised DL offers a robust method for PD-L1 quantification from IHC images.
• This approach has the potential to enhance patient stratification for cancer immunotherapy.
• The DL model's holistic integration of morphology and staining intensity surpasses traditional assessment methods.