Computational analysis of whole slide images predicts PD-L1 expression and progression-free survival in immunotherapy-treated non-small cell lung cancer patients

Abstract

Background: Immune checkpoint inhibitors (ICIs) have revolutionized cancer treatment by significantly improving the efficacy of treatments and tolerability for patients with non-small cell lung cancer (NSCLC). However, even after meticulous selection based on molecular criteria, only 20-30% of the patients respond to ICIs. This highlights the urgent clinical need to develop more precise biomarkers to better identify individuals who will benefit from these expensive therapies.

Methods: Data from NSCLC patients treated with immunotherapy were collected from two institutions. From the histological images of tumors, pathomics features were extracted. We employed six machine learning models and seven feature selection methods to predict expression of the programmed death-ligand 1 (PD-L1), a current biomarker used to select patients for immunotherapy, and progression-free survival (PFS). The association between pathomics features and biological pathways was explored to validate pathomics-based signatures. We performed gene set enrichment analysis to identify the pathways enriched with the predictive signatures.

Results: Handcrafted histological features were extracted from the whole slide images (WSI). The Support Vector Machines model with the SurfStar feature selection method, offered the best results, with an area under the curve (AUC) of around 0.66 for both the training and validation sets to predict PD-L1. For the prediction of PFS, the most effective model was linear discriminant analysis using the Multi Surf feature selection method with an AUC of 0.71 for the training set and 0.62 for the validation set. We found immune pathways to be upregulated in the high PD-L1 and high PFS groups, confirming the utility of image analysis for predicting clinical endpoints in patients treated with immunotherapy.

Conclusion: Our models, based on the analysis of histological images, can serve as predictive biomarkers for PD-L1 and PFS. This approach, focused on histological images, enables the distinction of patients based on treatment response, thus providing clinicians with a valuable tool for patient management. With further validation on external cohorts, these models could enhance clinical decision-making through analysis of routine medical images.

Fig. 1

The workflow for the development of the model to predict PD-L1 expression and PFS from patients with NSCLC. (A) Whole Slide image pre-processing and feature extraction. (B) Model development and biological validation

Fig. 2

Heatmaps depicting the efficacy of various machine learning models (on the y-axis) across Different Feature Selection Techniques (on the x-axis) for Predicting PD-L1. (A) AUC Scores During Cross-Validation and (B) AUC Scores in the External Validation Phase

Fig. 3

Heatmaps depicting the efficacy of various machine learning models (on the y-axis) across different feature selection techniques (on the x-axis) for predicting PFS. (A) AUC scores during cross-validation and (B) AUC scores in the external validation Phase

Fig. 4

Median performance of machine learning methods to predict (A) PD-L1 and (B) PFS on the validation dataset. The color bars correspond to different machine learning models: RF (blue), DT (orange), SVC (brown), and LD (brown, only in panel B). The error bars represent the standard deviation across multiple feature selection techniques

Fig. 5

Biological validation - Volcano plot highlighting pathway enrichment analysis using (A) PD-L1 and (B) PFS. Significantly upregulated signaling pathways are highlighted in green, while significantly downregulated pathways are indicated in blue, and non-significant pathways are represented in gray. The frame lines represent the threshold of significance, corresponding to the log10FDR with a cutoff of FDR = 0.1