The variational autoencoder (VAE) has become a prominent tool in cancer omics research, valued for its ability to learn structured embeddings of high-dimensional, heterogeneous biological data. Yet a systematic understanding of how VAEs and related deep representation learning (DRL) methods engage with the temporal aspects of cancer in omics data to capture longitudinal molecular dynamics of tumor progression has been lacking. We present such an assessment. Screening 440 publications from 2014 to 2024, 21 directly relevant studies were identified. DRL methods are predominantly applied to cancer omics data for subtyping, diagnosis, and prognosis. However, these tasks do not explicitly leverage the longitudinal molecular dimension of disease progression. Longitudinal omics studies of primary cancer are scarce, constrained by practical, ethical, and biological barriers, including the destructive nature of sequencing technologies and inter-patient heterogeneity. Temporal dimension in the omics data are most commonly represented through pseudo-time inference with single-cell data, not necessarily reflecting longitudinal molecular dynamics. Alternatively, cancer stages may be used as a proxy time axis. Against this landscape, the VAE emerges as uniquely suited to bridge the gap between available cross-sectional omics data and the need for longitudinal molecular modeling. Its latent space is directly operable: sample representations can be interpolated and fed to the decoder to produce realistic intermediate omics profiles. We discuss the methodological requirements needed to unlock this potential, and briefly present an application focusing on generating and forecasting synthetic cancer stage trajectories from transcriptomic data in renal cell carcinoma.

Co-authors: Davide Cirillo, Alfonso Valencia