The tumor microenvironment (TME) profoundly influences tumor progression and affects immunotherapy responses and resistance. Understanding its heterogeneity is the key for developing immunotherapy. However, the available methods can only partially portray the TME heterogeneity with a small number of cell types. Here, we developed a deep learning-based frame with a design visible, DCNet, that embeds the relationships between cells and their marker genes in the neural network, and can infer the cell landscape with more than 400 cell types based on bulk RNA-seq data. DCNet accurately recapitulated the cell landscape of multiple single cell RNA-seq datasets, which showed better robustness and stability. Based on the cell landscape of TCGA patients, which was built with DCNet, the patients were divided into two groups with significant differences in survival time and distinct cell-type populations. DCNet provides a foundation for decoding TME heterogeneity. The source code of DCNet can be found on GitHub: https://github.com/xindd/DCNet .

Cancer remains a formidable challenge in medicine due to its propensity for recurrence and metastasis, which can result in unfavorable treatment outcomes. This challenge is particularly acute for early-stage patients, who may experience recurrence and metastasis without timely detection. Here, we first analyzed the differences in clinical characteristics among the primary tumor, recurrent tumor, and metastatic tumor in different stages of cancer, which may be caused by the molecular level. Moreover, the importance of predicting early cancer recurrence and metastasis is emphasized by survival analyses. Next, we used a multi-omics approach to identify key molecular changes associated with early cancer recurrence and metastasis and discovered that early metastasis in cancer demonstrated a high degree of genomic and cellular heterogeneity. We performed statistical comparisons for each level of omics data including gene expression, mutation, copy number variation, immune cell infiltration, and cell status. Then, various analytical techniques, such as proportional hazard model and Fisher’s exact test, were used to identify specific genes or immune characteristics associated with early cancer recurrence and metastasis. For example, we observed that the overexpression of BPIFB1 and high initial B-cell infiltration levels are linked to early cancer recurrence, while the overexpression or amplification of ANKRD22 and LIPM, mutation of IGHA1 and MUC16, high fibroblast infiltration level, M1 polarization of macrophages, cellular status of DNA repair are all linked to early cancer metastasis. These findings have led us to construct classifiers, and the average area under the curve (AUC) of these classifiers was greater than 0.75 in The Cancer Genome Atlas (TCGA) cancer patients, confirming that the features we identified could be biomarkers for predicting recurrence and metastasis of early cancer. Finally, we identified specific early sensitive targets for targeted therapy and immune checkpoint inhibitor therapy. Once the biomarkers we identified changed, treatment-sensitive targets can be treated accordingly. Our study has comprehensively characterized the multi-omics characteristics and identified a panel of biomarkers of early cancer recurrence and metastasis. Overall, it provides a valuable resource for cancer recurrence and metastasis research and improves our understanding of the underlying mechanisms driving early cancer recurrence and metastasis.

Dysregulation of signaling pathways plays an essential role in cancer. However, there is not a comprehensive understanding on how oncogenic signaling pathways affect the occurrence and development with a common molecular mechanism of pan-cancer. Here, we investigated the oncogenic signaling pathway dysregulation by using multi-omics data on patients from TCGA from a pan-cancer perspective to identify commonalities across different cancer types. First, the pathway dysregulation profile was constructed by integrating typical oncogenic signaling pathways and the gene expression of TCGA samples, and four molecular subtypes with significant phenotypic and clinical differences induced by different oncogenic signaling pathways were identified: TGF-β+ subtype; cell cycle, MYC, and NF2− subtype; cell cycle and TP53+ subtype; and TGF-β and TP53− subtype. Patients in the TGF-β+ subtype have the best prognosis; meanwhile, the TGF-β+ subtype is associated with hypomethylation. Moreover, there is a higher level of immune cell infiltration but a slightly worse survival prognosis in the cell cycle, MYC, and NF2− subtype patients due to the effect of T-cell dysfunction. Then, the prognosis and subtype classifiers constructed by differential genes on a multi-omics level show great performance, indicating that these genes can be considered as biomarkers with potential therapeutic and prognostic significance for cancers. In summary, our study identified four oncogenic signaling pathway–driven patterns presented as molecular subtypes and their related potential prognostic biomarkers by integrating multiple omics data. Our discovery provides a perspective for understanding the role of oncogenic signaling pathways in pan-cancer.