Navigation Menu

Skip to content


Folders and files

Last commit message
Last commit date

Latest commit


Repository files navigation


PoreOver: Nanopore basecalling and consensus decoding


PoreOver is a basecalling tool for the Oxford Nanopore sequencing platform and is primarily intended for the task of consensus decoding raw basecaller probabilities for higher accuracy 1D2 sequencing. PoreOver includes a standalone RNN basecaller (PoreOverNet) that can be used to generate these probabilities, though the highest consensus accuracy is achieved in combination with Bonito, one of ONT's research basecallers.

More generally, PoreOver can serve as platform on which to explore new decoding algorithms and basecalling architectures.

If you find it useful, please cite: Silvestre-Ryan J, Holmes I. Pair consensus decoding improves accuracy of neural network basecallers for nanopore sequencing. Genome biology. 2021 Dec;22(1):1-6.


  • Python 3
  • TensorFlow 2


git clone
cd poreover
pip install -r requirements.txt
pip install -e .

Now the software can be run with:

poreover --help

Usage examples

PoreOver has four main modes:

  • call Run a forward pass of PoreOver's neural network and save the probabilities
  • decode Decode the output of PoreOver or another CTC basecaller using Viterbi or beam search
  • pair-decode Generate consensus sequences given a list of paired read probabilities
  • train: Train a new neural network using PoreOver

Basecalling with PoreOver's network (PoreOverNet)

PoreOver includes a simple basecalling network with an architecture inspired by other community basecallers such as DeepNano (Boža et al. 2017) and Chiron (Teng et al. 2018). It uses a single convolutional layer followed by three bidirectional GRU layers, and is trained with CTC loss. It is not as accurate as Bonito and is intended mostly for testing.

poreover call data/read.fast5

This will run the forward pass of the network and save the logits output to read.npy. This can then be passed to the decode

It will also take a directory as an input, e.g. poreover call data/reads

Decoding basecaller probabilities

A nucleotide sequence can then be decoded to FASTA format using decode. The basecaller must be specified to correctly parse the file.

poreover decode read.npy --basecaller poreover

By default, this does Viterbi (i.e. best path) decoding, though alternatively --algorithm beam will perform a beam search, with the beam width configurable with --beam_width. While beam search may outperform Viterbi decoding, in our experience any improvement is not usually worth the increased computational cost.

Pair decoding of 1D2 reads

Pair decoding can run either on a single pair as in

poreover pair-decode data/reads/read1.npy data/reads/read2.npy --reverse_complement --basecaller poreover

Or using a list of read pairs (provided probabilities have already been generated with a basecaller).

poreover pair-decode data/pairs.txt --reverse_complement --basecaller poreover

Each line in the pairs file must specify the paths to two reads' neural network output. As a convenience, if the listed file ends with .fast5 (as in the example) poreover will look for the appropriate .npy file.

Modifying Bonito for use with PoreOver

As Bonito does not currently support saving the basecaller probabilities, a slight modification must be made to allow this. This can be done using the bonito022.patch file (needs Bonito version 0.2.2).

git clone --branch v0.2.2
cd bonito
git apply ../poreover/data/bonito022.patch # substitute path to poreover repo
pip install -r requirements.txt
pip install -e .

Now, running Bonito will generate a .npy file for each read, named by the read ID.

Decoding flip-flop basecallers

Flip-flop is a CTC variation developed by ONT and implemented in their production basecaller Guppy, as well as the research bascaller Flappie. Both of these basecallers can optionally output a trace of probabilities using the --fast5-out (in Guppy) or the --trace (Flappie) option.

PoreOver can read and decode these probabilities, yielding a basecalled sequence.

poreover decode data/flappie_trace.hdf5

poreover decode data/guppy_flipflop.fast5

Note: Beam search decoding (and by extension pair decoding) does not seem to perform well on the flip-flop model and so is not recommended.

Training a new basecalling model for PoreOver

It is possible to use one of the architectures available in PoreOver to train a new basecalling model.

poreover train --data data/training.npz --loss_every 5 --epochs 10

This will use the default architecture of 1 convolutional + 3 bidirectional GRU layers, though there are a few other named architectures listed in poreover/network/ Checkpoints as well as the final parameters will be written to a directory named with [model architecture]_[run name]_[start date]_[start time]. This directory can be passed to the call command (as shown below).

Once there is a model to load, we can make a basecall on a sample read (of course, after only a little training on a toy dataset we would not expect it to be very accurate).

poreover call data/read.fast5 --weights $RUN_DIRECTORY --model $RUN_DIRECTORY/model.json