Added

• Python 3.12 support

Fixed

• Issue with checking for the presence of dwell information in fastq files.

Changed

• Behaviour of medaka_consensus with --bacteria option: if the basecaller model cannot
  be parsed or is not compatible with the bacterial polishing model, exit with an error instead of
  falling back to default model.
• Replaced pkg_resources with importlib

Removed

• Python 3.8 support (EOL).

Fixed

• Updated documentation with inference and sequence command renaming.
• Changed default model resolved from bam file from variant to consensus.
• Fixed issue with initializing inference in Medaka tandem model.
• Fixed a memory leak in the Medaka C library and removed redundant memory objects to reduce the footprint.

Changed

• Fully refactored and redesigned medaka tandem code and optimised CPU-based execution.
• Read-level models cannot be used with medaka tandem.
• get_trimmed_reads now also returns the phase-set, hap and read ids.

Added

• Consensus models for v5.2.0 basecaller models.
• Added support for read-level consensus models for v5.0.0 and v5.2.0 basecaller models.
• Models dna_r10.4.1_e8.2_5khz_400bps_sup and dna_r10.4.1_e8.2_5khz_400bps_hac added
  as aliases to those without _5kz_ in their names.
• Added -B option to medaka_consensus to allow passing a bed file or region to polish
  via medaka inference --regions.
• Added --cpu option to medaka inference to force CPU and avoid searching for GPUs.
• New output format for medaka tandem tailored for population studies.
• New fields to medaka tandem output: depth, read lengths, read names, phase sets, and MAD of read lengths.
• Read length–based outlier detection in medaka tandem.

Fixed

• medaka smolecule was broken by change from medaka consensus to medaka inference.

Changed

• Improved error message when model is not found.

Switched from tensorflow to pytorch.

Existing models for recent basecallers have been converted to the new format.
Pytorch format models contain a _pt suffix in the filename.

Changed

• Inference is now performed using PyTorch instead of TensorFlow.
• The medaka consensus command has been renamed to medaka inference to reflect
  its function in running an arbitrary model and avoid confusion with medaka_consensus.
• The medaka stitch command has been renamed to medaka sequence to reflect its
  function in creating a consensus sequence.
• The medaka variant command has been renamed to medaka vcf to reflect its function
  in consolidating variants and avoid confusion with medaka_variant.
• Order of arguments to medaka vcf has been changed to be more consistent
  with medaka sequence.
• The helper script medaka_haploid_variant has been renamed medaka_variant to
  save typing.
• Make --ignore_read_groups option available to more medaka subcommands including inference.

Removed

• The medaka snp command has been removed. This was long defunct as diploid SNP calling
  had been deprecated, and medaka variant is used to create VCFs for current models.
• Loading models in hdf format has been deprecated.
• Deleted minimap2 and racon wrappers in medaka/wrapper.py.

Added

• Release conda packages for Linux (x86 and aarch64) and macOS (arm64).
• Option --lr_schedule allows using cosine learning rate schedule in training.
• Option --max_valid_samples to set number of samples in a training validation batch.

Fixed

• Training models with DiploidLabelScheme uses categorical cross-entropy loss
  instead of binary cross-entropy.

(Probably) final version of medaka using tensorflow. Future versions will use
pytorch instead.

Fixed

• medaka_consensus: only keep bam tags if input file matches joint polishing pipeline.
• Pin numpy to <2.0.0.

Added

• Consensus and variant models lookup for v3.5.1 Dorado models.

Fixed

• tandem: Use haplotag 0 in unphased mode.
• tandem: Don't run consensus if regions set is empty.

Added

• Models for version 5 basecaller models.
• Expose sym_indels option for training.
• Expose --min_mapq minimum mapping quality alignment fitering option for medaka consensus.
• tandem: Option --ignore_read_groups to ignore read groups present in input file.
• Wrapper script medaka_consensus_joint and convenience tools (prepare_tagged_bam,
  get_model_dtypes) to facilitate joint polishing with multiple datatypes.

Added

• Consensus and variant models for v4.3.0 dorado models.

Added

• Parsing model information from fastq headers output by Guppy and MinKNOW.

Changed

• Additional explanatory information in VCF INFO fields concerning depth calculations.

Fixed

• Do not exit if model cannot be interpreted, use the default instead.
• An issue with co-ordinate handling in computing variants from alignments.

Added

• Ability to use basecaller model name as --model argument.
• Better handling or errors when running abpoa.

Fixed

• Correct suffix of consensus file when medaka_consensus outputs a fastq.

Added

• Choice of model file can be introspected from input files. For BAM files the
  read group (RG) headers are searched according to the dorado
  specification,
  whilst for .fastq files the comment section of a number of reads are checked
  for corresponding read group information. In the latter case see README for
  information on correctly converting basecaller output to .fastq whilst
  maintaining the relevant meta information.
• medaka tools resolve_model can display the model that would automatically
  be used for a given input file.

Changed

• If no model is provided on command-line interface (medaka consensus,
  medaka_consensus, and medaka_haploid_variant) automatic attempts will be made
  to choose the appropriate model.