Integrating multi-omics data for enhanced prognosis prediction in gastric cancer post-neoadjuvant therapy

Neoadjuvant therapy is crucial for locally advanced gastric cancer (LAGC), yet response varies significantly. Traditional models based on clinicopathological features often lack precision. This study aimed to develop and validate a comprehensive prognostic model integrating deep learning features from CT images and immune scores to improve risk assessment.

A total of 179 LAGC patients who received neoadjuvant therapy between 2019 and 2022 were divided into a development cohort (DC, n = 125) and an internal validation cohort (IVC, n = 54). Additionally, an external validation cohort (EVC) of 29 patients was included. Pre-treatment abdominal enhanced CT images were analyzed using a ResNet18-based deep learning model to extract features and generate a DeepScore via univariate Cox and LASSO regression. ImmuneScore was calculated from postoperative transcriptome data using the ESTIMATE algorithm. A multi-omics nomogram combining DeepScore, ImmuneScore, and ypTNM stage was constructed, calibrated in the development cohort, and validated.

In the DC, 3-year DFS rates for high, medium, and low DeepScore groups were 83.3%, 71.4%, and 29.3% (P < 0.0001); in the IVC, they were 92.0%, 66.7%, and 35.7% (P = 0.0011). The integrated nomogram achieved AUCs of 0.858, 0.843, and 0.839 (1-, 2-, 3-year DFS) in the DC, and 0.844, 0.825, and 0.833 in the IVC. In the EVC, the nomogram achieved AUCs of 0.786 and 0.785 for 1- and 2-year DFS, respectively. Low-risk patients showed significantly higher 3-year DFS and OS than high-risk patients in both DC and IVC cohorts (all P < 0.001). ssGSEA revealed higher immune infiltration in the low-risk group, and GSEA indicated enrichment in immune-related pathways.

The integrated model combining deep learning and immune scores enhances prognostic accuracy for LAGC patients after neoadjuvant therapy, offering valuable support for clinical decision-making.