Dataset overview & statistics

Dataset overview

Task columns: BC base calling · PD polyA detection · SA segmentation & event alignment · MD modification detection. For MD the cell shows which modification the sample carries ground truth for.

Dataset statistics

Read counts and average lengths come from the unified pipeline. * Base-called sequence lengths are from Guppy 6.0.1.

Processed dataset download

All processed data can be downloaded from Zenodo record 10889896. The dataset is dynamic: the current version is 1.0.0, and processed results for additional software are added over time. A small demo dataset is the fastest way to see every intermediate format.

• Demo dataset, clean (raw fast5 + reference)319.1 MBDownload
• Demo dataset, fully processed (all six pipeline steps)1.1 GBDownload
• DNA base calling benchmark (NA12878)70.6 GBDownload
• RNA base calling benchmark (hek293t)81.9 GBDownload
• PolyA benchmark (ont_ployA_standard)Download
• All files and versionsZenodoBrowse record

Processed dataset structure

Every dataset follows the same directory layout, numbered by pipeline step. Not all datasets contain all modules: Tailfindr and SegPore cannot process DNA, so DNA datasets usually lack 3_tailfindr and 6_segpore. The example below has two samples.

<dataset>/
├── 0_reference/
│   └── ref.fa                      # reference genome / transcriptome
├── 1_raw_signal/
│   ├── single_fast5/               # as downloaded (if applicable)
│   ├── multi_fast5/                # standardized input for all tools
│   └── multi_pod5/                 # for Dorado
├── 2_base_called/
│   ├── guppy/{pass,fail,workspace}/
│   ├── dorado/dorado.bam
│   ├── <dataset>.fastq
│   └── reads-ref.sorted.filter.bam # minimap2 + samtools
├── 3_tailfindr/
│   └── tails.csv
├── 4_nanopolish/
│   ├── polya.tsv, polya-pass-only-with-head.tsv
│   ├── eventalign.txt, eventalign_combined.txt
│   └── summary.txt
├── 5_tombo/
│   └── single_reads/{tombo_resquiggle.txt, tombo_summary.txt}
└── 6_segpore/

Dataset documentation

Collection criteria. Raw fast5/pod5 must be freely downloadable from a public repository; datasets should span diverse species; common RNA modifications (m6A, m5C, hm5C, inosine, pseudouridine) should be covered; and ground truth should be available for several tasks. All collected datasets use R9.4 or R9.4.1 chemistry.

Intended use. Developing and benchmarking machine-learning methods for the four tasks. Uses that re-identify individuals or contradict the original data providers' intent are out of scope; see the paper's discussion of societal impact.

Licensing and maintenance. Processed data: CC BY 4.0. Software: Apache 2.0. Data are hosted on Zenodo for long-term preservation; this website and the GitHub repository are maintained by the authors. A machine-readable datasets.json and schema.org Dataset metadata (embedded in this page) describe the collection.