Metagenomic Sequencing Pipeline

Runs all incoming metagenomic sequencing data through a consistent process, including cleaning, species assignment, and counting human-infecting viruses.

Table of contents

• Working with the output
• Adding new data
• Design
• Dependencies

Status

Deprecated

Data

[deprecated]

Output and intermediate steps for each delivery are in S3, under s3://nao-mgs/[deliveryid] (or s3://nao-restricted/... for private data). Available files, and their formats:

(Note that directories downstream from kraken have a -YYYY-MM suffix, representing the version of the kraken DB they were generated with.)

• raw/: Input, as downloaded from SRA

  • ex: SRR14530724_1.fastq.gz and SRR14530724_2.fastq.gz
  • Gzipped FastQ format
  • Paired-end reads split across two parallel files, one for the forward reads and another of exactly the same length for the reverse reads.

• cleaned/: Intermediate, output of AdapterRemoval2

  • ex: SRR14530724.collapsed.gz
  • Gzipped FastQ format
  • Multiple files, depending on what AdapterRemoval2 did
    • collapsed if it was able to combine the pair into one, pair1 and pair2 otherwise.
    • truncated if it did any quality trimming.
    • singleton if it dropped either the forward or reverse read.
    • discarded if it dropped the whole pair
    • Plus a settings file with info on how cleaning went

• ribofrac/: Intermediate, fraction of rRNA from fast RiboDetector

  • ex: COMO1.ribofrac.txt
  • txt
  • Fraction of rRNA in sample using subset of input reads

• riboreads/: Intermediate, read IDs of RiboDetector output

  • ex: SRR13167436.riboreads.txt.gz
  • txt
  • Stores rRNA read IDs

• processed/: Intermediate, output of Kraken2

  • ex: SRR14530724.collapsed.kraken2.tsv.gz
  • Gzipped TSV
  • Kraken output format
    • One line per input read with taxonomic classification and detailed hit information.

• samplereads/: Itermediate, example reads by taxonomic category

  • ex: SRR14530724.sr.tsv.gz
  • TSV
  • 100,000 examples (or as many as available) from each sample of:
    • a: all reads
    • b: bacterial reads
    • v: viral reads
    • h: human viral reads
  • These are sampled by assuming that order within a sequencing run doesn't matter: it takes the first 100,000 of each category.
  • Columns are:
    • Category (a/b/v/h)
    • Read ID

• readlengths: Output, distribution of read lengths by taxonomic category

  • ex: SRR14530724.rl.json.gz
  • JSON
  • category -> length -> count
  • Categories:
    • a: all reads
    • b: bacterial reads
    • v: viral reads
    • h: human viral reads
  • Length is a number, and then "NC" is cases where cleaning was unable to collapse the read. In these cases all we know is that the fragment was long enough to not get overlap.
  • Because we currently don't have a stage that does adapter removal without quality trimming, these will be an underestimate of fragment lengths, especially for lower quality reads.

• allmatches/: Intermediate, subset of Kraken2 output matching human viruses

  • ex: SRR14530724.allmatches.tsv
  • TSV
  • One row for each read with any hits against a human-infecting virus.
  • Kraken output format
  • Match can be in the assigned taxid, or any of the hits.

• hvreads/: Output, data that backs the read viewer

  • Ex: SRR14530724.hvreads.json
  • One record for each record in allmatches, joined back to the cleaned sequence and quality data.
  • JSON
  • Read ID to Kraken output and cleaned read

• alignments/: Output, alignment data that will later back the dashboard.

  • Ex: SRR21452137.hv.alignments.tsv.gz, SRR21452137.human.alignments.tsv.gz
  • Compressed TSV
  • One record for each read that Bowtie2 was able to map back to a genome in its DB.
  • We run this twice: once with a standard human genome DB ("human"), and again with a custom DB of human-infecting viruses ("hv").
    • For Human reads it runs with default setting, and the score is zero for perfect matches and increasingly negative for worse matches.
    • For HV reads we're running with custom settings, and the score is positive, higher for better matches. Some reads with positive scores are quite low quality matches: you likely want to filter with a combination of alignment score and trimmed read length (ex: Simon's been doing score/ln(length) > 22).
  • Non-collapsed reads will appear twice, one for the forward read and then one for the reverse.
  • Columns:
    • Read ID
    • Best-match genome
    • Best-match taxid
    • CIGAR string
    • Genome start position
    • Alignment score
    • Trimmed read length

• humanviruses/: Output, data that backs the dashboard.

  • Ex: SRR14530724.humanviruses.tsv
  • TSV
  • One row for each human-infecting virus observed in the sample
  • Columns are:
    • Taxid: NCBI Taxonomic ID, from the 2022-12-01 taxdmp release
    • Count: how many reads Kraken assigned to this taxid
    • Scientific name: the NCBI scientific name for this taxid

• cladecounts/: Output, taxonomic counts

  • ex: SRR21452136.tsv.gz
  • Gzipped TSV
  • Collated Kraken output
  • Columns:
    • Taxid: NCBI Taxonomic ID, from the 2022-12-01 taxdmp release
    • Direct assignments: how many reads in this sample Kraken assigned to this taxid.
    • Direct hits: how many reads in this sample had any 35-mer matches to the Kraken DB for this taxid
    • Clade assigments: how many reads in this sample Kraken assigned to this taxid or anywhere in its clade.
    • Clade hits: how many reads in this sample had any 35-mer matches to the Kraken DB for this taxid or anywhere in its clade.
  • For example, Tomaboviruses (12234) in SRR21452136.tsv.gz is: 12234 920 45914 954843 1027715 This is saying that:
    • 920 reads were assigned to "Tobamovirus".
    • 45,914 reads had some 35-mer that was common to Tobamoviruses and nothing more specific.
    • 954,843 reads were assigned to Tobamovirus or something within it's clade (ex: PMMoV).
    • 1,027,715 reads had at least one 35-mer hit within this clade.

Adding new data

We are no longer importing new public data.

Design

Each stage reads from an S3 directory under s3://nao-mgs/[deliveryId]/ and writes to a different one.

Species Classification

Kraken to assign taxonomic identifiers to reads.

Dependencies

E-utilities (for working with NCBI)

See the NCBI documentation.

sh -c "$(curl -fsSL https://ftp.ncbi.nlm.nih.gov/entrez/entrezdirect/install-edirect.sh)"
export PATH=${HOME}/edirect:${PATH}
sudo yum install perl-Time-HiRes

AdapterRemoval2

Visit https://github.com/MikkelSchubert/adapterremoval/ to get the latest version number, and then:

wget -O adapterremoval-2.3.3.tar.gz \
     https://github.com/MikkelSchubert/adapterremoval/archive/v2.3.3.tar.gz
tar xvzf adapterremoval-2.3.3.tar.gz
cd adapterremoval-2.3.3
make
sudo make install

The adapter removal step downloads the fastq files one sample at a time from S3 to the local machine. The step also generates local output files there before copying them to S3. On Linux these files are stored in /tmp. If you don’t have enough space available in /tmp, AdapterRemoval will crash. To check the available space in /tmp run df -H. You should have at least 2x the size of your largest pair of fastq files, multiplied by the maximum number of files you will need to process simultaneously.

To resize /tmp, edit /etc/fstab. If there is an entry for /tmp, add the option size=64G (or whatever size you need) to the 4th column. If not, add this line to the end of the file (tab-separated):

tmpfs  /tmp  tmpfs  size=64G  0  0

then run:

sudo mount -o remount /tmp/

RiboDetector

See documentation. To install:

pip install ribodetector

As of 2023-08-26, there is a bug in the latest version of one of the dependencies of RiboDetector, which causes the program to crash. To avoid this, install an earlier version of the dependency:

pip install onnxruntime==1.15.1

Kraken

Install

git clone [email protected]:DerrickWood/kraken2.git
cd kraken2/
./install_kraken2.sh ~/kraken2-install

Set up Kraken database

We're using the Standard pre-built database (see documentation). The prepare-shm-kraken.sh script automatically downloads and sets it up, but first you need to configure RAM. Edit /etc/fstab and add:

none     /dev/shm      tmpfs  defaults,size=85G      0 0

The 85G comes from starting with the 78GB listed for the Standard kraken DB at https://benlangmead.github.io/aws-indexes/k2 and then adding 6.5GB for the custom human virus and human bowtie DBs.

See "Updating the Taxnonomy and Kraken DB" below if it's out of date and you'd like it not to be.

BowTie2

Install

wget -O bowtie2.zip "https://sourceforge.net/deliverys/bowtie-bio/files/bowtie2/2.5.2/bowtie2-2.5.2-linux-x86_64.zip/download" 
unzip bowtie2.zip 
rm bowtie2.zip

Download pre-built human genome database

For detecting human reads we use the standard pre-built telomere-to-telomere "Human / CHM13plusY" database from https://benlangmead.github.io/aws-indexes/bowtie. See https://www.science.org/doi/10.1126/science.abj6987 for the construction of this genome.

cd mgs-pipeline/bowtie
aws s3 cp s3://genome-idx/bt/chm13.draft_v1.0_plusY.zip .
unzip chm13.draft_v1.0_plusY.zip
mv chm13.draft_v1.0_plusY/* .
rmdir chm13.draft_v1.0_plusY
rm chm13.draft_v1.0_plusY.zip

Build custom human viral database

To create a Bowtie2 database, we need to download genomes from NCBI, using ncbi-genome-download. To run the affiliated script gimme_taxa.py, you will also need to install the dependencies ete3 and six.

pip install ncbi-genome-download
python -m pip install ete3 six ncbi-genome-download

Now you can run build_bowtie_db.py, which will create the bowtie2 database. This will take quite some time so it's best to run this command within a screen session to not inadvertantly end the script when closing your terminal.

Operations

We normally run this pipeline on c6a.16xlarge EC2 instances. We need 128GB+ of memory; the binding constraint is that the Standard Kraken DB is 78GB.

Rerunning for all Deliveries

While ./run.py runs serially for a single delivery, we can run in parallel at the sample level with ./reprocess.py. You specify a maximum number of parallel jobs. For example:

mgs-pipeline $ ./reprocess.py \
    --max-jobs 12 \
    --log-prefix prefix \
    --sample-level \
    --shuffle
    -- \
    --run-arguments-here \

This will run the pipeline once for each delivery, including restricted deliveries if available in ../mgs-restricted. If the stages you're running require a lot of disk, use a number lower than 12, potentially 1: the script doesn't know how "heavy" a job is, and will run out of disk space if you tell it to do do too much in parallel.

Job output is under log/ in files named by the date and the prefix you supply. So if I ran the above on 2023-01-01 I'd expect to see files like:

...
log/2023-01-01.prefix.PRJNA924011
log/2023-01-01.prefix.PRJNA943189
log/2023-01-01.prefix.PRJNA966185
...

If the job fails, the last line in the log file will start with "ERROR:" and then have the exit code.

Screen Oversight

You can check in on parallelized jobs under screen with:

mgs-pipeline $ pipeline-operation/screen-summary.py
8769..assembly:
4.collapsed.gz
hvreads: handling ERR7850094.collapsed.truncated.gz in /tmp/tmphu0ym26c
download: s3://nao-mgs/PRJEB49260/cleaned/ERR7850094.collapsed.truncated.gz to
.
/ERR7850094.collapsed.truncated.gz
hvreads: handling ERR7850094.discarded.gz in /tmp/tmpegmh_drc
download: s3://nao-mgs/PRJEB49260/cleaned/ERR7850094.discarded.gz to
./ERR785009
4.discarded.gz
hvreads: handling ERR7850094.pair1.truncated.gz in /tmp/tmpmel9okco
download: s3://nao-mgs/PRJEB49260/cleaned/ERR7850094.pair1.truncated.gz to
./ERR
7850094.pair1.truncated.gz

This prints the bottom ten lines of each active screen session on the computer.

You can also run:

mgs-pipeline $ pipeline-operation/print-running-jobs.sh
 --delivery PRJEB49260 --stages hvreads

For every currently executing instance of ./run.py this prints the arguments it was started with.

Regenerating Data

Normally the pipeline doesn't repeat work, but when the code changes some data typically also needs to change.

Normally the flow is:

1. Run the pipeline on a single sample (--sample RUN_ACCESSION) until you're happy with what it does.

2. Find where it checks whether your output already exists. For example, for hvreads there's a line like: existing_outputs = get_files(args, "hvreads").

3. Add a min_date argument to the get_files call, like min_date='2023-09-18'.

4. Follow the instructions above to rerun across all deliveries.

Updating the Taxnonomy and Kraken DB

Before starting, check in with all other users of the pipeline, and ensure no one wants you to hold off. Also check how recent the databases are on https://benlangmead.github.io/aws-indexes/k2. If they're not recent, it may be worth waiting for the next update.

Once you've decided to go ahead:

1. Move human-viruses-raw.tsv, human-viruses.tsv, plus, within dashboard, *.dmp, top_species_counts, and top_species_scratch into a scratch location. Delete hvreads, readlengths, ribofrac, cladecounts and allmatches.

2. Update download-taxonomy.sh to pull the latest taxonomy. Run the script.

3. Run download-human-viruses.sh

4. Put the year and month in reference-suffix.txt

5. Reprocess the deliveries we need to have up to date. The command will look something like:

./reprocess.py \
   --log-prefix kraken-update \
   --max-jobs 8 \
   --shuffle \
   --sample-level \
   --deliveries DELIVERY_1,DELIVERY1_,...DELIVERY_N \
   --

6. Manually compare the output to the previous generation.

7. Consider whether we need to keep the previous generation data (or previous previous etc) around.