# A cell-specific computational framework reveals a pan-cancer hypoxia signature predicting overall survival and ICI response

## TL;DR

This study introduces a new method to assess hypoxia in tumors across many cancer types, linking it to survival and response to immunotherapy.

## Contribution

A pan-cancer hypoxia signature (HYP.SIG) is developed to predict survival and immunotherapy response at the single-cell level.

## Key findings

- HYP.SIG scores correlate with genetic instability and poor ICI response across multiple cancer types.
- HYP.SIG outperforms existing signatures in predicting immunotherapy outcomes.
- Four potential therapeutic targets (LDHA, SERF2, SLC2A1, NOP53) were identified using CRISPR data.

## Abstract

Hypoxia forms an immunosuppression environment and is involved in tumor immune escape, which may be the potential culprit of resistance to anticancer therapies. Nevertheless, there is still a lack of research that explores the characteristics of hypoxia in pan-cancer at the single-cell level and assesses the application of hypoxia in immune checkpoint inhibitor (ICI) efficacy and clinical outcomes. We delineated cell-specific hypoxia levels and developed a computational framework to generate a pan-cancer tumor hypoxia-related transcriptomic signature (HYP.SIG) using 38 scRNA-seq datasets encompassing 362 patients and 893,464 cells across 19 cancer types. We defined computational indicators of hypoxia levels as HYP.SIG scores to characterize the hypoxia status across 33 cancer types and 29 normal tissues within 18,901 samples. HYP.SIG scores exhibited cancer type-specific associations with genetic instability, and were linked to oncogenic signaling, poor response to ICI therapy, and impaired survival in multiple cancer types. Moreover, we established a predictive model for immunotherapy response utilizing six machine learning algorithms and 9 ICI cohorts (904 patients, four cancer types). HYP.SIG achieved better predictive performances in comparison to other previously established signatures. Subsequently, we applied three machine learning-based feature selection algorithms to filter HYP.SIG survival-related signatures and developed a prognostic model for predicting overall survival, incorporating clinical disease stages. Eventually, we screened four candidate therapeutic targets (LDHA, SERF2, SLC2A1, NOP53) for patients with tumors using 17 CRISPR cohorts and 1078 CRISPR cell lines. Overall, our study provides new ideas for survival prognostication, prediction of ICI response, and clinical therapeutic target development from the perspective of hypoxia.