Benchmarking circRNA Detection Tools from Long-Read Sequencing Using Data-Driven and Flexible Simulation Framework (Data for extension)

# nano-circ v3 benchmark extension: simulated datasets These are the simulated long-read circRNA datasets for the benchmark extension (Rusakovich et al., 2026): - Code: https://github.com/aerusakovich/nano-circ, folder `nanocirc_pipeline_v3/` (`simulator/` and `benchmark/`) - Results: `nanocirc_pipeline_v3/benchmark_results.md` This extends the published study (Rusakovich et al., 2026) by adding a new simulator, all four callers in containers, circFL-seq scoring, BSJ-based matching and a fix for the `bedtools intersect --split -f` bug. The callers were run with the nanocirc pipeline (https://github.com/aerusakovich/nanocirc_pipeline): CIRI-long, isoCirc, circFL-seq and circNICK-lrs. ## Datasets Each benchmark has 3 replicate runs. | Folder | Benchmark | Genome / annotation | circRNA set | Read error model | |---|---|---|---|---| | `human/` | Species benchmark, human | hg38 / GENCODE v47 | Generated in silico, fitted to human circBase+circAtlas | CIRI-long protocol | | `mouse/` | Species benchmark, mouse | GRCm38.p4 / GENCODE vM10 | Generated in silico, fitted to the mouse circRNA database | CIRI-long protocol | | `human_db/` | Database benchmark (annotated circRNA recovery) | hg38 / GENCODE v47 | 2,000 circRNAs per type, sampled from the human circBase/circAtlas union | CIRI-long protocol | | `protocols/` | Same circRNAs, three wet-lab protocols | hg38 / GENCODE v47 | One generated circRNA set shared by all three protocols | CIRI-long (`ciri_long/`), circFL-seq (`circfl/`), isoCirc (`isocirc/`) | Each read error model was trained with NanoSim on real ONT control reads from that wet-lab protocol. In `protocols/`, only the protocol changes between subfolders. The circRNAs, their abundances and the genome are the same, so any difference in results comes from the protocol. ## Files ### Per dataset (top level of each folder) Each top-level file is prefixed with its dataset name (`human_`, `mouse_`, `human_db_` or `protocols_`). | File | Content | |---|---| | `*_circRNAs.fa` | Circular (rolling-circle) sequences, the input to NanoSim `transcriptome` mode | | `*_circRNAs.mature.fa` | One linear mature copy per circRNA | | `*_circRNAs.bed` | BED12 of every simulated circRNA, with the exon/intron block structure | | `*_circRNA_metadata.tsv` | Per-isoform table: type, gene/transcript/locus, exon and intron counts, splice type, copy number, TPM and target read count | | `*_abundances.tsv` | TPM per circRNA (the NanoSim expression profile) | | `*_ground_truth.tsv` | BED12 fields plus type, locus/isoform, TPM and read counts for the full simulated set (10,000 circRNAs: 2,000 each of ecircRNA, EIciRNA, ciRNA, antisense and intergenic) | ### Per replicate (`runN/`, or `/runN/` in `protocols/`) | File | Content | |---|---| | `*_runN.fastq.gz` | Simulated ONT reads: BSJ-confirmed circular reads plus linear background reads from the transcriptome, in a 1:1 ratio | | `*_runN_ground_truth.tsv` | **The ground truth used for scoring.** It lists only the circRNAs that had at least one BSJ-confirmed read in this run, with their read counts (`n_reads_confirmed`) | | `*_runN_circRNAs.bed` | BED12 of the same circRNAs as that run's ground truth (`human/`, `mouse/` and `human_db/` only) | Each replicate draws a different, mostly overlapping subset of circRNAs from the full set (a Poisson draw around each circRNA's expected share). A circRNA with no BSJ-confirmed read in a run is left out of that run's ground truth, because of lack of circularity evidence. ## circRNA types - `ecirna`: exons only; may skip an internal exon - `eicirna`: exons plus one or more retained introns - `cirna`: a single intron - `antisense`: on the opposite strand of the gene - `intergenic`: no gene overlap on either strand (strand reported as `.`) The type is also the prefix of each circRNA ID (for example `ecirna0000000`). IDs are only unique within a dataset. ## Coordinates BED files are 0-based and half-open. `*_ground_truth.tsv` files use the same coordinates, with extra columns added after the 12 BED12 fields.