Biological sequences contain information about the function and regulation of different components within a biological system. Accurate modeling will yield a deeper understanding of the system, as well as better experimental designs. Nonetheless, computational constraints of the different available methods continue to prevent its use on large scale. Sliding sub-sampling of biological sequences can generate either vector-based or graph-based sequence encodings. Each one of them can be used to train generative models for an unsupervised representation learning task. Offering a suitable tool to analyze fast-evolving biological systems such as SARS-Cov2. Analysis of variational autoencoders bottleneck representation shows a distinguishable temporal component. Changes in 4-mer composition in the region that codes for the structural SARS-Cov2 proteins. Non-symmetrical changes in 4-mer composition drive the viral temporal adaptation process. Furthermore, mean nucleotide composition and encodings of SARS-Cov2 appear to be constrained by day length. Development and refinement of sequence analysis methods will lead to a better understanding of viral adaptation and evolution.