Simulation data for benchmarking de novo long read transcriptome assembly software

Method of simulation of differentially expressed biological replicates

We first obtained a subset of transcripts that are widely expressed in the GTEx v9 dataset (92 samples) using Gencode comprehensive annotation (v44). We kept transcripts with more than 5 reads in at least 15 samples after Salmon quantification (18145 genes, 40509 transcripts), and stored their mean count per million (CPM) values as the control group’s baseline expression. We then generated a perturbed set of CPM values where transcript expression was changed by: (1) randomly selecting 1000 genes and changing all transcripts belonging to that gene concordantly (500 genes 2 fold up and 500 genes 2 fold down), (2) selected another 1000 genes randomly, and then select 2 random transcripts from the gene and swap their expression, (3) selected another 1000 genes randomly, and then select 1 random transcript to change its expression (500 transcripts 2 fold up and 500 transcripts 2 fold down). The updated CPM were stored as the perturbed group baseline expression. We then generated a count matrix and CPM matrix for 3 control replicates and 3 perturbed replicates with gamma distribution, followed by a Poisson distribution (Baldoni et al., 2024). Both long-read and short-read FASTQ files were simulated using SQANTI-SIM with default settings and ONT R9.4 cDNA error profile (v 0.2.1) (Mestre-Tomás et al., 2023). The long read data contained 6 million reads in total, and an average read length of 1085 bp, and short read data was 100 bp paired-end. We then subsampled the short-read data to match the total number of base pairs in the long read data (6.5 billion bases). The simulated data was non-stranded, and contains 2000 DE genes, 2000 genes with DTU, 5927 transcripts with DTU and 6933 DE transcripts.