Construction of hot tumor classification models in gastrointestinal cancers

Gastrointestinal (GI) cancers account for more than one-third of cancer-related mortality, and the prognosis for late-stage patients remains poor. Immunotherapy has been proven to extend the survival of patients at advanced stages; however, challenges persist in patient selection and overcoming drug resistance. Tumor-infiltrating lymphocytes (TILs) and tertiary lymphoid structures (TLS) in the tumor microenvironment (TME) have been found to be associated with anti-tumor immune responses. 'Hot tumors' with high levels of infiltration tend to respond better to immune checkpoint inhibitor (ICI) therapy, making them potential biomarkers for ICI treatment.

To explore potential biomarkers for predicting immunotherapy response and prognosis in GI cancers, we downloaded the gene expression profiles of seven GI cancers from The Cancer Genome Atlas (TCGA) database and characterized their TME, classifying the samples into hot/cold tumor subgroups. Furthermore, we developed a computational framework to construct cancer-specific hot tumor classification models with only a few genes. External independent datasets and qPCR experiments were used to verify the performance of our few-gene models.

We constructed cancer-specific few-gene models to identify hot tumors for GI cancers with only two to nine genes. The results showed that B cells are important for hot tumor determination, and the identified hot tumors are significantly associated with TLS. They not only overexpress TLS marker genes but are also associated with the presence of TLS in whole-slide images. Further, a two-gene qPCR model was developed to effectively distinguish between hot and cold tumor subgroups in cholangiocarcinoma, providing an opportunity for stratifying patients with hot tumors in clinical settings.

In conclusion, our established few-gene models, which can be easily integrated into clinical practice, can distinguish hot and cold tumor subgroups, and may serve as potential biomarkers for predicting ICI response.