Deep Learning-Derived Pathomic Features Predict NCIT Efficacy in Resectable Locally Advanced ESCC: Clinical Utility and Mechanistic Insights

TL;DR

A deep learning model predicts treatment response in esophageal cancer patients and identifies a potential new target to improve immunotherapy effectiveness.

Contribution

A novel AI-driven pathomic model for predicting neoadjuvant chemoimmunotherapy response in ESCC, with mechanistic insights into immunotherapy resistance.

Findings

The integrated model achieved high accuracy (AUC = 0.897) in predicting treatment response using pathomic features and clinical variables.

High ECiT scores correlate with immune activation, while low scores are linked to ER stress and UPR activation, suggesting a role in immunotherapy resistance.

EIF2S3 is identified as a key mediator of UPR activation and poor prognosis, offering a potential therapeutic target.

Abstract

Esophageal squamous cell carcinoma exhibits high mortality and limited therapeutic options. While immune checkpoint inhibitors improve outcomes, identifying non-responders to neoadjuvant chemoimmunotherapy remains urgent. This study developed a predictive model for treatment efficacy using deep learning and real-world cohort data, with mechanism exploration via TCGA datasets. Integrating histopathological images and clinical variables, the model demonstrated a robust performance and revealed associations between treatment response, immune activation, and specific cellular processes. These findings offer insights that may inform personalized therapeutic strategies and improve the understanding of potential mechanisms underlying immunotherapy resistance.

Background: Esophageal squamous cell carcinoma (ESCC) is the predominant subtype of esophageal cancer, with poor outcomes following neoadjuvant chemoradiotherapy (NCRT). Neoadjuvant chemoimmunotherapy (NCIT) has emerged as a promising strategy, but reliable predictive biomarkers remain lacking. This study aimed to develop an AI-driven pathomic model for NCIT response prediction and explore its biological mechanisms. Methods: We analyzed 269 H&E-stained whole-slide images (WSIs) from 198 ESCC patients (104 from Tongji Hospital, 94 from TCGA). Using ResNet152, we segmented WSIs into four tissue categories (tumor cells, stroma, lymphocytes, and necrosis), extracted spatially weighted pathomic features, and constructed the ECiT score via logistic regression. An integrated model combining the ECiT score with clinical variables (T stage, P53 status) was developed. Mechanistic analyses were performed using TCGA-ESCA and GSE160269 datasets. Results: The integrated model achieved AUCs of 0.897 (training) and 0.809 (temporal validation), outperforming clinical (AUC = 0.624) and pathomic-only (AUC = 0.751) models. Mechanistically, a high ECiT score correlated with enhanced immune activation (elevated CD4+ memory T cell infiltration), while low scores were linked to endoplasmic reticulum (ER) stress-unfolded protein response (UPR) activation. EIF2S3 was identified as a key molecular mediator, correlating with three pathomic features, UPR activation, and poor prognosis. Conclusions: This study may offer a preliminary indicator that could assist in personalized clinical decision-making. Correlative evidence suggests that the EIF2S3-mediated ER stress–UPR axis represents a potential candidate therapeutic target to overcome NCIT resistance, generating testable hypotheses to advance precision oncology for resectable locally advanced ESCC.

Figures