armadillin 0.25

Lineage prediction from SARS-CoV-2 sequences

Armadillin

This is an experimental tool under development. The recommended method for calling lineages remains normal Pangolin: https://github.com/cov-lineages/pangolin

A Re-engineered Method Allowing DetermInation of viraL LINeages

Armadillin is an experimental alternative approach to training models on lineages designated by the PANGO team.

Armadillin uses dense neural networks for assignment, which means it doesn't have to assume that positions with an N are the reference sequence. Armadillin is still very fast, in part because it sparsifies the feature input to this neural net during training.

Installation (for inference)

conda create --name armadillin python=3.9
conda activate armadillin
pip3 install armadillin

Usage

You must already have aligned your files to the reference (doing this automatically is on the backlist).

We'll use the COG-UK aligned file for a demo:

wget https://cog-uk.s3.climb.ac.uk/phylogenetics/latest/cog_alignment.fasta.gz

armadillin https://cog-uk.s3.climb.ac.uk/phylogenetics/latest/cog_alignment.fasta.gz

or

armadillin https://cog-uk.s3.climb.ac.uk/phylogenetics/latest/cog_alignment.fasta.gz > output.tsv

Training your own models

Dataset generation

python -m armadillin.training_make_input --designations ~/gisaid/pango-designation-1.2.88/ --gisaid_meta_file ~/gisaid/metadata.tsv --gisaid_mmsa ~/gisaid/msa_2021-10-20.tar.xz --output ~/training_set_nov_02
 python -m armadillin.train --shard_dir /home/theo/training_set_nov_02 --use_wandb --checkpoint_path ~/nov2check1

 python -m armadillin.train --starting_model ~/nov2check1/checkpoint.h5 --use_wandb --checkpoint_path ~/nov2check1_sparse/ --do_pruning --shard_dir /home/theo/training_set_nov_02

 python -m armadillin.training_create_small_model -i /tmp/model_zeros.h5 -d  /home/theo/training_set_nov_02

Related tools

Pangolin is the OG for assigning lineages