Basecaller

Click here for an introduction of how nanopore sequencing and basecalling works.

Basecalling quick start instructions can be found here.

Basecalling pipeline

Dorado is the Oxford Nanopore basecaller. Basecalling is the process of calling nucleotide bases (ACGT) from a nanopore signal which is recorded from your sequencing device. This signal data is stored in POD5 or .fast5 files, although support for the .fast5 file format has been deprecated.

There are many parts of the basecalling process many of which we will explain in this document.

graph LR
  IN[Data Ingest
  ...
  POD5 / .fast5
  ] --> PREP[Read
  Pre-Processing
...
  Scaling
  Normalisation
  Filtering
  Signal Trimming
  ...];
  PREP --> ML[Machine Learning
  Algorithm &
  Decoder];
  ML --> EXTRA[Additional
  Processing
  ...
  Alignment
  Barcoding
  Mod Basecalling
  ...]
  EXTRA --> POST["Sequence
  Post-Processing
  ...
  Filtering
  Trimming
  Poly(A) Estimation
  ..."];

  POST --> WRITER[Writer
  ...
  Filtering
  ];

---

Data ingest

Support for .fast5 files is deprecated

Reads are sourced from POD5 and .fast5 files.

-n / --max-reads N - Basecall only N reads

-l / --read-ids FILE - Basecall only reads in FILE. This read ids file should contain newline delimited read ids only.

-r / --recursive - Consider all POD5 or FAST5 files in data and all subdirectories.

---

Signal pre-processing

graph LR
SRC[Signal In] --> SCALE[Scaling & Normalisation];
SCALE --> TRIM_ST[Signal Trimming];
TRIM_ST --> OUT[Signal Out];

The signal chunk pre-processing stage takes raw signal chunks from their source and performs some manipulations to improve basecalling accuracy.

Signal scaling aims to bring raw signal values into a consistent regime to improve the performance of the machine learning algorithm using during basecalling. The scaling strategies used are pico-Ampere (PA) scaling with and without standardisation, quantile scaling and median absolute deviation (medmad) scaling. The scaling strategy used depends on the how the basecalling model was trained and as such the scaling strategy selection is controlled by the model configuration file. Note that scaling is performed on the whole read before trimming.

DNA rapid adapter trimming is disabled

In the trimming stage, RNA adapters and DNA rapid adapters* are detected in the signal and their position is recorded.

The very beginning of a nanopore sequencing read often contains samples that are far out of the normal distribution and as such we try to remove these samples in DNA reads only by detecting the first samples of useful data.

Finally, if any signal trimming is required, it is actioned.

---

Machine learning algorithm

The machine learning algorithm is at the core of the basecalling process. These highly accurate and efficient models can decode the sequencing signal to produce nucleotide base calls. The machine learning model predicts the probabilities of each base being present throughout the signal. These probabilities are then decoded to determine the most likely sequence of bases which is then emitted as the output of the algorithm.

There have been many versions of basecalling models with improvements to model architecture and training data to enhance basecalling speed and accuracy over time. The basecalling model is set by the user and is normally one of fast, hac (high-accuracy), and sup (super-accurate).

More details can be found in the dedicated models documentation.

Advanced basecalling controls

Here are the advanced basecalling parameters available to the user:

Warning

We generally do not recommend users set the arguments listed here unless absolutely necessary.

The default values are often good for most users however if you're facing issues please read the troubleshooting guide.

--batchsize SIZE - The batch size controls the number of signal chunks passed the the model at once. This can be thought of as the number of rows in a table. Often, increasing the batch size results in better overall throughput especially on modern GPU hardware. However, this comes at the cost of increases memory usage which is limited. Dorado will use a default batchsize if one is has been computed for the system hardware.

---

Additional processing

Alignment

Please see alignment for details.

Barcoding

Please see barcoding for details.

Modified basecalling

Please see modified basecalling for details.

---

Sequence post-processing

The sequence post-processing stage includes a number of stages depending on the options selected.

Read trimming

Please see read trimming for details.

Read splitting

Please see read splitting for details.

Poly(A) estimation

Please see poly(A) estimation for details.

---

File writer

See the Dorado SAM file specification for details on the SAM / BAM output generated by Dorado.

BAM is the default file format written by Dorado basecaller, however SAM and FASTQ file formats can be selected. The content of the BAM and SAM files will be identical but the metadata content in the FASTQ file will differ and some of the metadata tags are not supported.

By default, the output file is written to stdout. This process is summarised in the getting started help for redirecting output but the output can be written directly to files using the -o / --output-dir argument.

--emit-sam - This flag sets the output file format to be SAM.

--emit-fastq - This flag sets the output file format to be FASTQ.

--emit-moves - This flag will write the move table into the SAM / BAM outputs.

-o / --output-dir DIR - This optional argument can be used to specify an output directory which follows the MinKnow output structure.

---