Overview

As mentioned in the introduction tutorial as well as the read processing tutorial, read processing can make a huge impact on downstream work. While cutadapt which was introduced in the read processing tutorial is great for quick evaluation or dealing with a single bad sample, it is not as robust as some other trimmers in particular when it comes to removing sequence that you know shouldn't be present but may exist in odd orientations (such as adapter sequences from the library preparation). This tutorial is adapted from the 2021 trimmomatic tutorial which sought to do the same basic things as fastp: get rid of adapter sequences first and foremost, ideally even before fastQC so you can make any quality or length based improvements on actual data not artifacts. The #1 biggest reason why fastp is now the instructor's preferred trimming program is this box taken from the trimmomatic tutorial:

The more collaborative your work is, the less confidence you will have in picking the correct adapter file with trimmoatic, and while thanks to conda installations it can be pretty easy to test multiple different adapter filees, fastp does all the guess work for you, and can generate some interesting graphs itself.

Learning objectives:

1. Install fastp

2. Remove adapter sequences from some plasmids and evaluate effect on read quality, or assembly.

Installing fastp

fastp's home page can be found on github and has links to the paper discussing the program, installation instructions for conda, and information on each of the different options available to the program. This is far above the quality the average programs will have as most will not have a user manual (or not nearly so detailed), may not have been updated since originally published (or may not have been published), etc. It having been updated since the publication is one thing that makes it such a good tool as the more who use it the more likely problems are found, and having a group who is going to actively improve the program will significantly increase its longevity.

conda create -n GVA-ReadPreProcessing -c conda-forge -c bioconda fastp fastqc multiqc

Trimming adapter sequences

Example generic command

fastp -i <READ1> -I <READ2> -o <TRIM1> -O <TRIM2> --threads # --detect_adapter_for_pe -j <LOG.json> -h <LOG.html>

Breaking down the parts of the above command:

-j <LOG.json>

-h <LOG.html>

All of the above has been put together from the help fastp --help command.

Trimming a single sample

Get some data

mkdir -p $SCRATCH/GVA_fastp_1sample/Trim_Reads $SCRATCH/GVA_fastp_1sample/Raw_Reads
cd $SCRATCH/GVA_fastp_1sample
cp $BI/gva_course/plasmid_qc/E1-7* Raw_Reads

The ls command should show you 2 gzipped fastq files. You may notice that here that we used a wildcard in the middle of our copy path for the first time. This is done so that you can grab both R1 and R2 easily without having to type out the full command. Double tab will help tell you when you have a sufficiently specific base name to only get the files you are after.

Trim the fastq files

The following command can be run on the head node. Like with FastQC if we are dealing with less than say 1-2Million reads, it is reasonable to run the command on the head node unless we have 100s of samples in which case submitting to the queue will be faster as the files can be trimmed all at once rather than 1 at a time. Use what you have learned in the class to determine if you think this command should be run on the head node. (this was covered in more detail in the first part of the evaluating and processing read quality tutorial.)

zgrep -c "^+$" Raw_Reads/*.fastq.gz

fastp -i Raw_Reads/E1-7_S187_L001_R1_001.fastq.gz -I Raw_Reads/E1-7_S187_L001_R2_001.fastq.gz -o Trim_Reads/E1-7_S187_L001_R1_001.trim.fastq.gz -O Trim_Reads/E1-7_S187_L001_R2_001.trim.fastq.gz -w 4 --detect_adapter_for_pe 

Evaluating the output

Using everything you have learned so far in the class, can you answer the following questions?

1.  What are the 4 new files that were generated, and where did they come from?

   • E1-7_S187_L001_R2_001.trim.fastq.gz and E1-7_S187_L001_R1_001.trim.fastq.gz

     • These were created with the -o and -O options, they are in the Trim_Reads folder, and you likely found them using the ls command
       
   • fastp.html and fastp.json
     • These are log files created by default since we didn't specify their names. This is part of why -j and -h were discussed above with the general command. 
     • While the json file can be evaluated in the terminal (cat less more head tail), the html file has to be transferred back to your computer to view.
2.  How many reads were left after trimming?

   • 5884 paired end reads

   • 11768 total reads

     • You likely found this out from using the zgrep command, or from the following blocks that printed as the command ran:
       

       Read1 after filtering:
       total reads: 5884
       total bases: 791763
       Q20 bases: 782948(98.8867%)
       Q30 bases: 765510(96.6842%)
       
       Read2 after filtering:
       total reads: 5884
       total bases: 791763
       Q20 bases: 711414(89.8519%)
       Q30 bases: 658164(83.1264%)
       
       Filtering result:
       reads passed filter: 11768
       reads failed due to low quality: 2014
       reads failed due to too many N: 0
       reads failed due to too short: 0
       reads with adapter trimmed: 3970
       bases trimmed due to adapters: 193972

       
       

3.  How big was our fragment size? How is this estimated, what might make it inaccurate?

   • From the information generated while the command ran we see:

     • Insert size peak (evaluated by paired-end reads): 171

     • This tells us that the average peak size was 171 bases, and that it was estimated by looking at the overlap between the read pairs. It is potentially inaccurate as reads which do not overlap each other can not estimate the size.
   • If you transferred the .html file back to your laptop, you would see this relevant histogram:
     
     
     • The general section of the summary at the top of the html tells us that the average insert size was 171, while the histogram tells us that 50% of our data is <18 or >272 bases
4.  Did our sample have any adapter present?

   • If you only look at the information that printed to the screen, you probably answer "No"
     • you likely see the following block and think this is the end of the answer:

     • Detecting adapter sequence for read1...
       No adapter detected for read1
       
       Detecting adapter sequence for read2...
       No adapter detected for read2

   • A more fuller answer might be "maybe" or "probably" or "I'm not sure" as:
     • 1. Not finding any adapter would be super rare
     • 2. If 45% of our reads have an insert size of 171 bases, and we did 151bp PE sequencing,  we should be able to find adapter sequences
     • 3. in the filtering results we see:

     • Filtering result:
       reads passed filter: 11768
       reads failed due to low quality: 2014
       reads failed due to too many N: 0
       reads failed due to too short: 0
       reads with adapter trimmed: 3970
       bases trimmed due to adapters: 193972

   • If you look at the html file you probably answered "yes"
     • There is a section for Read1 and Read2 adapters which show a growing stretch of DNA which recreates the illumina adapter sequences.
5.  Why was answering if there were adapters present not straight forward?

   Like we saw in our fastqc reports (over represented sequences having "no hit" and adapter content staying at bottom of graph), for something to be classified as an "adapter" in the first section of the printed information, it has to meet certain criteria that in this (and many other instances) is perhaps a bit too stringent. 

6.  What other interesting things can you find from this command?

   This is pretty open ended, take a look at the html file in particular, see what of it does or doesn't make sense and consider asking a question if you would like to know more.

   Of several things that you may stand out to you is large fraction of reads end with stretches of "G" on the end. There are 2 things to note with this: 1. 2 color sequencing on illumina (detailed information here) reads "no color" as "G", 2. This library is very fragmented and contains adapter dimers meaning that in some cases there are only ~40bp downstream of the sequencing primer location leaving 60 cycles that have no template available. If you look at the help for fastp the following options may stand out to you as a way to deal with this:

   • -g, --trim_poly_g                    force polyG tail trimming, by default trimming is automatically enabled for Illumina NextSeq/NovaSeq data
   • --poly_g_min_len                 the minimum length to detect polyG in the read tail. 10 by default. (int [=10])
   • -G, --disable_trim_poly_g            disable polyG tail trimming, by default trimming is automatically enabled for Illumina NextSeq/NovaSeq data
   • -x, --trim_poly_x                    enable polyX trimming in 3' ends.
   • --poly_x_min_len                 the minimum length to detect polyX in the read tail. 10 by default. (int [=10])

   Consider rerunning the fastp command while adding "-g" to the command line and see how the results differ.

   
   

Trim all the samples from the multiqc tutorial

As mentioned above, if you have already done the multiqc tutorial, you can use your new fastp command to remove the adapter sequences from all 544 samples or make other changes based on what you saw from the fastqc/multiqc reports. 

Get some data

mkdir -p $SCRATCH/GVA_fastp_multiqcSamples/Trim_Reads
cd $SCRATCH/GVA_fastp_multiqcSamples
cp -r $BI/gva_course/plasmid_qc/ Raw_Reads/

 The ls command will now show 2 directories, and like in the multiqc tutorial, you can use ls and wc commands to figure out how many files you will be working with.

Trim the fastq files

Just as we used a for loop to set up a set of FastQC commands in the multiqc tutorial, we can use a similar for loop to generate a single file with 272 trim commands for the 544 total files.

The following command will pair all R1 reads in the Raw_Reads folder with its R2 pair, determine the base name, and generate a command to trim the file 

for R1 in Raw_Reads/*_R1_001.fastq.gz; do R2=$(echo $R1| sed 's/_R1_/_R2_/'); name=$(echo $R1|sed 's/_R1_001.fastq.gz//'|sed 's/Raw_Reads\///'); echo "fastp -i $R1 -I $R2 -o Trim_Reads/$name.trim.R1.fastq.gz -O Trim_Reads/$name.trim.R2.fastq.gz -w 4 --detect_adapter_for_pe -j Trim_Logs/$name.json -h Trim_Logs/$name.html &> Trim_Logs/$name.log.txt";done > fastp.commands

Use the head and wc -l to see what the output is and how much there is of it respectively.

Again as we discussed in the multiqc tutorial, running this number of commands is kind of a boarder line case of being better to run in idev or as a submitted job, there are not a lot of total reads, but there are a large number of samples. Because our command does request additional processors we should not run on the head node, if we wanted to use only a single processor, the job would take so long to run on the head node that even waiting in the queue system would be faster.

Do only one of the following, but do read through both options as there are different discussions about the process in each.

Submitting as a job

cp /corral-repl/utexas/BioITeam/gva_course/GVA.launcher.slurm trim.slurm
nano trim.slurm

Again while in nano you will edit most of the same lines you edited in the in the breseq tutorial. Note that most of these lines have additional text to the right of the line. This commented text is present to help remind you what goes on each line, leaving it alone will not hurt anything, removing it may make it more difficult for you to remember what the purpose of the line is

Line 17 being set to -n 17 allows 17 jobs to run at the same time, since our command uses -w 4 (4 threads) this job will use all 48 threads available. The changes to lines 22 and 23 are optional but will give you an idea of what types of email you could expect from TACC if you choose to use these options. Just be sure to pay attention to these 2 lines starting with a single # symbol after editing them.

Again use ctl-o and ctl-x to save the file and exit.

mkdir Trim_Reads Trim_Logs
sbatch trim.slurm

The job should take less than 10 minutes once it starts if everything is working correctly, the showq -u command can be used to check for the job finishing.

Running on idev

The alternative to running all the commands as a submitted job is of course to run on an idev node. In the multiqc tutorial, we were able to tell fastqc to analyze all samples just by using the "*" wildcard as the only required input to fastqc is the filename. Here our command is much more intricate which may seem like it precludes us from being able to run interactively as we never would type 272 nearly identical commands. Obviously there is a trick for this.

chmod +x <filename>
./<filename>

Before we jump to making our commands file executable and executing it, we want to change it to be slightly different. Specifically, above we used -w 4 to specify we wanted to use 4 processor for every command. While this worked great when we also were launching 12 processes at the same time as it used all 48 processes, when executing a commands file from the command line without the help of the queue system, only 1 sample at a time will launch so you likely think we need to increase to 48 processors. 

#1 change the for loop:
for R1 in Raw_Reads/*_R1_001.fastq.gz; do R2=$(echo $R1| sed 's/_R1_/_R2_/'); name=$(echo $R1|sed 's/_R1_001.fastq.gz//'|sed 's/Raw_Reads\///'); echo "fastp -i $R1 -I $R2 -o Trim_Reads/$name.trim.R1.fastq.gz -O Trim_Reads/$name.trim.R2.fastq.gz -w 48 --detect_adapter_for_pe -j Trim_Logs/$name.json -h Trim_Logs/$name.html &> Trim_Logs/$name.log.txt";done > fastp.commands

#2 use sed to do an in file replacement (something new you haven't seen before in this class)
sed -i 's/ -w 4 / -w 48 /g' fastp.commands

Note that if you use the sed command above, you need to be very careful in what you choose to match to. If you just choose "4" and replace with "48" then the commands file will then change any file name that has a 4 into 48 and all those samples will fail. When using sed to do replacements, always make sure you have a unique handle, when you don't, and when you don't need one.

chomd +x fastp.commands
./fastp.commands

Once the command is started continue reading below.

Comparing different run options 

In previous years it has been common to question what the fastest way of getting a large set of samples analyzed is with respect to threads and Nodes and tasks. Here we hav an opportunity to do just that, and have some surprising results. Since we have been working with idev sessions all along we'll start with the following:

This brings us to our first question: how can using fewer processors speed things up. I am not completely sure but I do have some hypotheses, and expect the answer to be somewhere in the middle:

1. We have so few reads (10,000s) that trying to spread them in 68 different groups, read them, modify them, write them back to the growing trimmed files, perform statistics on all the reads. That the computation required to spread them out (the overhead) actually exceeds whatever speed bonus is available with the extra threads.

2. If you used this command and look in the .log.txt file that was generated thanks to our pipe "&>", you will find a interesting note:

   WARNING: fastp uses up to 16 threads although you specified 68

   this means that despite requesting fastp use 68 threads, it is only capable of, and only used 16 so there is an even smaller difference in number of threads available. Unfortunately this is something that doesn't seem to be documented on their website or in the help information.

Given that threads only effect the speed of a single sample I also attempted to trim all the read files at once by telling the idev node to run them in the background (this was done by adding a single & as the last character on each line of the commands file)

While both of these errored after ~1/5 and ~1/2 of the total samples making them a non viable single pass method, they did finish in <10 seconds. Using grep we could generate a list of samples which did finish, store those in a file, then use another grep command with the -v and -f options to remove the finished samples from analysis. This is something that can be very helpful if you have a large number of samples and your run timed out as you can focus on just samples that have not yet finished instead of having to rerun all samples, or manually edit the commands file to remove samples which finished. Since this run failed due to lack of memory, not lack of time its less reasonable to try to tackle it in this manner.

That brings us to our last set of runs: those working with the commands as a submitted job.

Based on what we have already seen, it is probably not surprising that using 68 (really 16) threads and only evaluating 1 sample at a time took approximately the same amount of time as it did when running on an idev node as those conditions are functionally equivalent. What may be surprising is the lack of improvement despite running 4x more samples at the same time. Potential hypotheses:

1. The amount of overhead the job launcher has in checking when a command finishes and starting the next command is similar to the amount of overhead in splitting reads into 4 sets
2. The biggest contributor to time is not something that can be improved with additional threads

Evaluating the output

The first thing you always want to check when working with a lot of commands simultaneously is if they all finished correctly (note above how running all 272 jobs in the background at basically the same time did not finish correctly). Typically this is where the log files we generate with "&>" come in handy. if you look at the tail of any 1 of the files you are likely to see something like:

Duplication rate: 6.5207%

Insert size peak (evaluated by paired-end reads): 80

JSON report: Trim_Logs/E2-1_S189_L001.json
HTML report: rim_Logs/E2-1_S189_L001.html

fastp -i Raw_Reads/E2-1_S189_L001_R1_001.fastq.gz -I Raw_Reads/E2-1_S189_L001_R2_001.fastq.gz -o Trim_Reads/E2-1_S189_L001.trim.R1.fastq.gz -O Trim_Reads/E2-1_S189_L001.trim.R2.fastq.gz -w 68 --detect_adapter_for_pe -j Trim_Logs/E2-1_S189_L001.json -h Trim_Logs/E2-1_S189_L001.html 
fastp v0.23.4, time used: 3 seconds

Typically what you should look for is some kind of anchor that you can pass to grep that is as far down the file as possible. Sometimes you will be lucky and the program will actually print something like "successfully complete". In our case the last line looks promising, "fastp v0.23.4, time used:" seems likely to be printed as the last step in the program.

ls Trim_Logs/*.log.txt|wc -l
wc -l fastp.commands
tail -n 1 Trim_Logs/*.log.txt|grep -c "^fastp v0.23.4, time used:" 

The above 3 commands are all expected to return 272. If so remember I'm on zoom if you need help looking at whats going on. 

Beyond the job finishing successfully, the best way to evaluate this data would actually be to actually use it. That could me running it through multiqc to see how the trimming has effected overall quality, or assembling the reads, or mapping them to a reference (not applicable in this case since they are from a variety of sources and you have not been provided reference files.

Optional next steps:

1. Consider moving back over to the multiqc tutorial use multiqc to determine well the trimming worked. 
2. The reads could then further be assembled in the Spades tutorial as they are all plasmid samples.

Return to GVA2023 home page