A major issue with evaluating the large number of RNA-seq methods available is that, in real data, the form and strength of unwanted variation within these data is never known (see here for motivation about unwanted variation). This creates issues with benchmarking — when researchers evaluate a method, they simulate data only under the conditions that they assume (or minor deviations from these assumptions), and so their performance on real data is never accurately gauged. For my first solo-author publication (Gerard, 2020), I sought to create realistic simulation methods that accurately capture the unknown structure of unwanted variation in RNA-seq datasets. I did this by taking real RNA-seq data and adding known signal to it. I proved that my method of adding signal is guaranteed to work under a very flexible class of distributions on the RNA-seq read-counts. These simulation techniques are all implemented in my published software package, seqgendiff.
Gerard, David. (2020). Data-based RNA-seq simulations by binomial thinning. BMC Bioinformatics 21(1), 206. doi:10.1186/s12859-020-3450-9 | bioRxiv:758524.