predict.sh

#!/usr/bin/bash
#
# This script contains an example run of ClaMSA to classify MSAs as coding or not.
# It computes for each alignment given in the list of fasta files vertebrate.test.msa.lst
# and for each ClaMSA model the probability that the MSA is coding under the model.
# Models are specified with --model_ids { "label1" : "ID1", "label2" : "ID2, ...},
# where
# - The label is arbitrary, it will merely become the tile of the column in the output table
# - The ID is from training
#         - if you use pre-trained parameters, it could be "default", for example
#         - if you trained yourself, it is the time stamp used for storing weights (see train.sh)
# - A common choice may be --model_ids '{ "clamsa" : "default" }' for a single prediction each
#   with the default parameters.

TESTDATADIR=../data/test/
if [ ! -d $TESTDATADIR ]; then
    echo "Test data not found in $TESTDATADIR. Go to ../data and run download_vert_test.sh"
    exit 1
fi


../clamsa.py predict fasta \
   vertebrate.test.msa.lst \
   --clades ../examples/vertebrate.nwk \
   --use_codons \
   --model_ids '{ "clamsa" : "default" }' \
   --out clamsa_vert.csv

# takes about 30 seconds on 3000 MSAs

# Other examples.
# When called with
#   --model_ids '{  "tcmc_rnn" : "default", "tcmc_mean_log" : "logreg" }'
# predictions are done with two models and the output table has two prediction columns.
#
# After you have trained ClaMSA yourself as in train.sh
# look up the timestamp from the log file or saved_weights directory and predict with the
# self trained run like this
# --model_ids '{ "my_train_run" : "2021.03.18--15.17.25" }'


# A warning like
# WARNING:tensorflow:Skipping loading of weights for layer P_sequence_columns due to mismatch in shape ((21,) vs (53,)).
# can be ignored. This is about a mismatch between the number of leaves in the prediction clade and the training forest ( = set of training trees)