Data-driven simulation is extensively used to assess the performance of statistical methods in a large variety of conditions. Indeed, \textit{in vivo} or \textit{in vitro} experiments with ground truth in controlled conditions are too expensive to produce a large number of replicates, especially in genomics. Similarly, designing a genomic data challenge such as HADACA3, whose focus is on multi-omics cellular deconvolution, also requires to reproduce \textit{in silico} the complex variability of paired bulk tissues from many omics data types. Here we focused on two frequently used omic data types for cell deconvolution, RNA-seq gene expressions data and DNA methylation rates. Most stochastic models for such multivariate data assume specific distributions at the individual level, CpG's probes or genes, consistently accounting for the intrinsic nature of genomic measurements, including nonnegativity and overdispersion. However, introducing biologically sound patterns of dependence, across genes or, for a given gene, across different types of omic data, is more challenging, due to the complexity of this dependence and the huge number of genes or CpG sites. Ignoring this dependence in simulation setups may lead to a strong optimism bias in the evaluation of cell deconvolution methods. The gene regulatory network induces a particular dependence structure for both data types, which for this study is inferred from \textit{in vitro} benchmark gene expression and DNA-methylation data using an efficient Expectation-Maximisation algorithm \cite{friguet2009factor}. Here, we propose a data-driven methodology introducing a low-rank factor approximation of the conditional variance-covariance between genes. The factorial decomposition captures specific and shared variability of the data, with different levels of details possible controlled by the number of factors treated as a tuning parameter. Our study clearly demonstrates that the accuracy of most cell deconvolution algorithms strongly depends on the interplay between patterns of cellular heterogeneity and dependence across genes.