The ecological and evolutionary perspectives of tumorigenesis can be characterized as a process of microevolution in tumor cells that altered the tumor microenvironment and further induced tumor cell proliferation, metastasis, and the death of tumor patients. Here, we introduced XgeneVAE, an interpretable unsupervised deep learning framework that quantified the semantic changes in multi-omics embedding space for characterizing the microevolution processes and fitness effects of heterogeneous tumor samples. We then validated that the scales of the latent embedding variances can reflect the differences in the overall survival of tumor patients, as well as their applications in uncovering the driving genomic alternations in different cancer types. These results confirmed that the XgeneVAE model can better represent the heterogeneity in distinct cancer types and as an interpretable model for understanding the fitness effects in tumorigenesis and their association with clinical outcomes.