Deep Learning-Powered Spatial Transcriptomics Analysis to Predict Immune Checkpoint Inhibitor Response from Pre-Treatment Digital Pathology Slides

Abstract

Response to immune checkpoint inhibitors (ICIs) is governed in part by the spatial organization of the tumour microenvironment, yet most deep learning models applied to digital pathology treat tissue regions independently and therefore do not represent the neighbourhood interactions that shape immune activation. This study develops and retrospectively evaluates a Neighborhood-aware Masked Graph Spiking Neural Network (NAMG-SNN) that classifies ICI responders and non-responders from pre-treatment haematoxylin and eosin whole-slide images. Whole-slide images are stain-normalized, tiled and encoded into patch-level embeddings that combine statistical, textural, morphological and deep semantic descriptors. A spatial graph is then constructed over the retained patches and pruned by an adaptive similarity mask that suppresses weak or artefactual neighbourhood edges. The masked graph is propagated through a leaky integrate-and-fire spiking network whose accumulated firing rates form a spatial transcriptomic representation, which is finally classified into responder and non-responder categories. The model was evaluated on an integrated cohort assembled from The Cancer Genome Atlas (TCGA), the Panoptic segmentation of nUclei and tissue in Melanoma pathology Challenge (PUMA challenge) and the Phoenix virtual spatial transcriptomics resource, using patient-level splitting. On the integrated test data the framework reached 97.31% accuracy, 96.94% precision, 97.82% recall, 97.38% F1-score and 98.87% area under the receiver operating characteristic curve (AUC), with a Matthews correlation coefficient of 0.947 and Cohen’s kappa of 0.946, exceeding convolutional, transformer and graph baselines trained under identical conditions. Transcriptomic reconstruction reached a Pearson correlation of 0.962, mean absolute error of 0.051, root mean square error of 0.081 and structural similarity index (SSIM) of 0.966. Ablation confirmed that neighbourhood-aware graph construction, adaptive masking and spike-based learning each contributed to performance. These results are retrospective and computational.