A multi-omics deep learning model for hypoxia phenotype to predict tumor aggressiveness and prognosis in uveal melanoma for rationalized hypoxia-targeted therapy

Abstract

Uveal melanoma (UM) represents the most common primary intraocular malignancy in adults and is characterized by aggressive behaviors and a lack of targeted therapies. Hypoxia-targeted therapy has become a promising new therapeutic strategy in tumors. Therefore, a better understanding of the tumor hypoxia microenvironment is critical to improve the treatment efficacy of UM. In this study, we conducted an extensive multi-omics analysis to explore the heterogeneity and prognostic significance of the hypoxia microenvironment. We found that UM revealed the most significant degree of intertumoral heterogeneity in hypoxia by quantifying tumor hypoxia compared with other solid tumor types. Then we systematically correlated the hypoxia phenotypes with clinicopathological features and found that hypoxic UM tumors were associated with an increased risk of metastasis, more aggressive phenotypes, and unfavorable clinical outcomes. Integrative multi-omics analyses identified multidimensional molecular alterations related to hypoxia phenotypes, including elevated genome instability, co-occurring of 8q arm gains and loss of chromosome 3, and BAP1 mutations. Furthermore, hypoxic UM tumors could be characterized by increased CD8+ T cell infiltration and decreased naïve B cell and dysregulated metabolic pathways. Finally, we introduced DNN2HM, an interpretable deep neural network model to decode hypoxia phenotypes from multi-omics data. We showed that the DNN2HM improves hypoxia phenotype prediction and robustly predicts tumor aggressiveness and prognosis in different multi-center datasets. In conclusion, our study provides novel insight into UM tumor microenvironment, which may have clinical implications for future rationalized hypoxia-targeted therapy.

Keywords

• Deep neural network
• Hypoxia microenvironment
• Uveal melanoma
• Multi-omics data