search

al-mcintyre / mCaller

A python program to call methylation (m6A in DNA) from nanopore signal data
MIT License
45 stars 16 forks source link

readme

    H
     |
H-C-aller
     |
    H

This program is designed to call m6A from nanopore data using the differences between measured and expected currents.

Dependencies/requirements

  • python3 (as of mCaller version 1.0)
  • nanopolish (https://github.com/jts/nanopolish)
  • an aligner to create a bam file (has been tested with graphmap and bwa mem) python packages
  • scikit-learn
  • h5py
  • biopython
  • matplotlib
  • seaborn
  • numpy
  • pysam
  • scipy
  • pandas

other

  • nanopore sequencing data (fastq format + fast5 to run nanopolish, basecalled using Albacore or another basecaller that saves event data)
  • a reference sequence file (fasta)

Installation

Add softlinks for make_bed.py and mCaller.py to path or run from mCaller directory

Options

usage: mCaller.py [-h] (-p POSITIONS | -m MOTIF) -r REFERENCE -e
                             TSV -f FASTQ [-t THREADS] [-b BASE]
                             [-n NUM_VARIABLES] [--train] [-d MODELFILE]
                             [-s SKIP_THRESH] [-q QUAL_THRESH] [-c CLASSIFIER]
                             [-v]

arguments:

  -h, --help            show this help message and exit
  -p, --positions
                        file with a list of positions at which to classify
                        bases (must be formatted as space- or tab-separated
                        file with chromosome, position, strand, and label if
                        training)
  -m, --motif 
                        classify every base of type --base in the motif
                        specified instead (can be single one-mer)
  -r, --reference 
                        fasta file with reference aligned to
  -e, --tsv     tsv file with nanopolish event alignment
  -f, --fastq 
                        fastq file with nanopore reads
  -t, --threads
                        specify number of processes (default = 1) 
            multiple threads now includes sorting but this can create trade-offs
            in terms of speed
  -b, --base  bases to classify as methylated or unmethylated (A or
                        C, default A)
  -n, --num_variables
                        change the length of the context used to classify
                        (default of 6 variables corresponds to 11-mer context
                        (6*2-1))
  --train               train a new model (requires labels in positions file)
  --training_tsv 
                        provide an mCaller output file with labels for training
  -d, --modelfile 
                        model file name
  -s, --skip_thresh 
                        number of skips to allow within an observation
                        (default 0)
  -q, --qual_thresh
                        quality threshold for reads (default none)
  -c, --classifier 
                        use alternative classifier: options = NN (default), RF,
                        LR, or NBC (others may severely increase runtime)
  --plot_training       plot probabilities distributions for training
                        positions (requires labels in positions file and
                        --train)
  -v, --version         print version

Pipeline for methylation detection from R9 data

  1. extract template strand reads from fast5 files. Follow the most up-to-date guidelines from nanopolish (https://github.com/jts/nanopolish). As of April 2019:

    nanopolish index -d <fast5 directory> -s sequencing_summary.txt <filename>.fastq

    or 

    poretools fastq --type fwd <fast5 directory> > <filename>.fastq

  2. align fastq reads to reference assembly (we have used both GraphMap and bwa mem, with comparable results):

    bwa index <reference>.fasta 
bwa mem -x ont2d -t <num_threads> <reference>.fasta <filename>.fastq | samtools view -Sb - | samtools sort -T /tmp/<filename>.sorted -o <filename>.sorted.bam 
samtools index <filename>.sorted.bam

    or 

    graphmap align -r <reference>.fasta -d <filename>.fastq -o <filename>.sam 
samtools view -bS <filename>.sam | samtools sort -T /tmp/<filename>.sorted -o <filename>.sorted.bam 
samtools index <filename>.sorted.bam

  3. run nanopolish with the following command to save a tsv file and the event values scaled towards the model:

    nanopolish eventalign -t <num_threads> --scale-events -n -r <filename>.fastq -b <filename>.sorted.bam -g <reference>.fasta > <filename>.eventalign.tsv

  4. run mCaller to detect m6A:

    mCaller.py <-m GATC or -p positions.txt> -r <reference>.fasta -d r95_twobase_model_NN_6_m6A.pkl -e <filename>.eventalign.tsv -f <filename>.fastq -b A

    This returns a tabbed file with per-read predictions, where columns indicate chromosome, read name, genomic position, position k-mer context, features, strand, label, and probability of methylation predicted by mCaller for that position and read

  5. (optionally) run summary script to generate a bed file of per-position methylation predictions:

    make_bed.py -f <filename>.eventalign.diffs.6 -d 15 -t 0.5

    -d indicates the minimum read depth for inclusion, while -t indicates the --mod_threshold, or the minimum fraction of observations of a position (ie. currents deviations in individual reads) labelled as methylated, based on a >= 50% predicted probability of methylation with the mCaller model. Check make_bed.py --help for all options.

Results and analysis scripts for the E. coli datasets are provided in the bioRxiv folder.

Test data

Reference fasta, PacBio calls for m6A and a subset of A positions, and eventalign tsv + fastq are provided for a single read for testing purposes in the "testdata" folder.

  1. To run mCaller on the testdata, use:

    ./mCaller.py -p testdata/test_positions_<m6A/A>.txt -r testdata/pb_ecoli_polished_assembly.fasta -e testdata/masonread1.eventalign.tsv -d r95_twobase_model_NN_6_m6A.pkl -f testdata/masonread1.fastq

    Can also try using -m GATC, although not all GATC positions within the read were identified as methylated on both strands using PacBio and the model is slightly weighted to accept more false negatives than false positives at the moment.

    This will generate the output file testdata/masonread1.eventalign.diffs.6

./make_bed.py -f testdata/masonread1.eventalign.diffs.6 -d 1 -t 0.5

Will then generate the output bed file testdata/masonread.methylation.summary.bed with columns chrom, chromStart, chromEnd, context, % methylated, strand, depth of coverage

  1. To train on the testdata (don't actually use this model trained on one read), try:

    ./mCaller.py -p testdata/test_positions.txt -r testdata/pb_ecoli_polished_assembly.fasta -e testdata/masonread1.eventalign.tsv -t 4 --train -f testdata/masonread1.fastq

    This will generate the output file model_NN_6_m6A.pkl.

    Results with the R9.5 model

    latest results Unsurprisingly, mCaller performs best at identifying motifs similar to those it's trained on (here, E. coli). This is a clear limitation to the method, but the results are still sufficient in many cases to verify a motif of interest.

Models provided