Preprocessing

csRNA-seq and small RNA-seq input sequencing reads were trimmed of their adapter sequences using HOMER (homerTools trim -3 AGATCGGAAGAGCACACGTCT -mis 2 -minMatchLength 4 -min 20 -max 58, see also http://homer.ucsd.edu/homer/ngs/csRNAseq/) and aligned to species specific genome using HISAT2 v.2.2.1 (hisat2 -p 30 --rna-strandness F --dta -x {hisat2_index} -U {file_prefix}-trimmed.fastq.gz -S {file_prefix}.sam) outputting a single sam file. Using samtools version 1.16.1, the SAM files were then converted to BAMs (samtools view -bS {file_prefix}.sam > {file_prefix}.bam), and reads were filtered for only the ones uniquely mapped (samtools view -h -q 1 -e ' [NH] == 1 ' -b -o {file_prefix}_uniqueAligned.bam {file_prefix}.bam).

Next, the 5' ends of the aligned reads were used to construct the genome-wide read count profiles of the transcription start sites. To prepare the challenge data, we selected contigs longer than 100,000 bp and masked the genomic regions that overlap with the repeats (using the following repeat masking tracks: rmsk, simpleRepeat, tandemDups, except Squalus Suckleyi, which doesn't have a tandemDups track, all available from GenArk for each of the EPIC genomes). Afterwards, the non-masked regions shorter than 100 bp were discarded. Next, we assigned the individual chromosomes or contigs to training or test data to maintain the balance in GC composition (considering only non-masked regions). Third, we subsetted nearly half of the test data for scoring during the online leaderboard stage, while the other half will be used for the final evaluation. By default, the data are distributed as 80%:20% between training and test subsets, except for larger genomes of Octopus bimaculoides and Squalus suckleyi for which we use 90%:10% and 95%:5% splits to keep the submission file size manageable.