Sixth researcher

Another Bioinformatics website

Category: Scientific software

Scientific software

Discovering tumor neoantigens is easier than getting a job in Spain

The story of this post started few months ago at a immunotherapies conference in Barcelona, at this time I was looking for a job in cancer research. I had a great time learning about new cancer therapies and meeting nice people, but few doctors were interested, or had money, to contract a bioinformatician. Nevertheless, one of them told me: “I don’t know you, but if you show me that you are able to discover tumor neoantigens, then I’ll consider to contract you in the future”. I hope this person will read the post ūüėČ


But before starting, I want to dedicate this post to all the doctors, nurses, scientists and patients that fight everyday against cancer, specially those from Hospital Miguel Servet de Zaragoza and Cl√≠nica Universidad de Navarra that treat my father, also to many other brilliant and devoted people that I have met from places like Vall d’Hebron Hospital, Quiron Dexeus or 12 de Octubre, and finally, to my new team at Agenus Pharmaceuticals.


State of the art of tumor neoantigen discovery

With next generation sequencing (NGS) we are able to sequence the full genome of a tumor or at least of the coding regions or exome. If we identify with the help of NGS data which tumor-specific genes express mutant peptides that are able to activate T cells, then we can design vaccines to boost the natural immune response to the tumor, this mutant peptides are also called neoantigens. There is a nice and recent Nature Review about neoantigens discovery, and also there are interesting articles with practical cases of neoantigens discovery and personalize treatments for cancer patients (e.g. Robbins et al., Pritchard et al.).

But we are interested in the bioinformatics point of view, so let’s make a summary of some neoantigens discovery protocols and software in the literature:

  • pVAC-Seq v4.0: a pipeline for the identification of personalized Variant Antigens by Cancer Sequencing (pVAC-Seq) that integrates tumor mutation and expression data (DNA- and RNA-Seq) (original article).
  • TSNAD: an integrated software¬† for cancer somatic mutation and tumour-specific neoantigen detection (original article).
  • neoantigenR: an R package that identifies peptide epitope candidates (biorxir article).
  • MuPeXI: mutant peptide extractor and informer, a tool for predicting neo-epitopes from tumor sequencing data (original article).
  • INTEGRATE-neo: a pipeline for personalized gene fusion neoantigen discovery (original article).

We can conclude that the topic is quite new, most of the previous pipelines have been published during the last year, and that the neoantigen identification is a very complex task as we can see for example in the pVAC-Seq workflow schema:

Overview of the pipeline pVAC-Seq

To show an example of neoantigens discovery pipeline, I’ll try to simplify the process and follow the Best Practices guide from The Broad Institute for somatic variants discovery:

Best Practices for Somatic SNV and Indel Discovery


Tumor neoantigen discovery pipeline description

This protocol has been developed for didactic purposes and it is not formally validated neither intended for human diagnosis.

Basically I propose here a somatic variant calling pipeline with some additional filtering of the variants at the end. I tested the pipeline with whole exome sequencing (WES) Illumina paired-end reads from normal and tumor tissues.

The full pipeline can be summarized in 10 steps, here is briefly explanation of every step and some additional considerations about the chosen tools:

  1. Normal and tumor reads mapping, sorting and indexing:
    • In the protocol I have chosen BWA-MEM as recommended in Cornish and Guda paper, original authors used Novoalign, but other aligners could be considered. In this aspect, aligners as Bowtie2 may be too sensitive and could report some extra false positive variants, but it’s just a personal opinion.
  2. Removal of duplicated reads:
    • Maybe this step could be skipped as there are not many redundant reads and it’s unlikely that they map in variant positions. MarkDuplicates from Picard tools was used, I would suggest in the future to try ‘MarkDuplicatesWithMateCigar’ from the same suite, but I have not enough experience to recommend it.
  3. Adding sample/group information to the mapped reads:
    • This is just a formal requirement for the variant callers to work properly and distinguish between normal and tumor groups of reads.
  4. Realignment around indels:
    • This step is already integrated in the Mutect2 somatic variant caller and could be redundant, but other variant callers do not integrate it so it is strictly recommended to avoid false calling of indels and surrounding SNPs.
  5. Somatic variant calling with MuTect2:
    • MuTect2 is integrated in the GATK suite, the basic operation of MuTect2 proceeds similarly to that of the popular HaplotypeCaller, with a few key differences, taking as input tumor and normal alignments and producing somatic variant calls only.
  6. Filtering of somatic variants:
    • Some strict filters are imposed to discover only true somatic variants: a minimum depth of 10 reads and a variant frequency of at least 10%. Additionally, variants must be in coding regions of genes (it is filtered in the last step).
  7. Annotation of filtered somatic variants
    • Filtered variants are annotated: gene and transcript names and IDs, chromosome position, nucleotide and protein mutations and other relevant information about the antigens is annotated in a tabulated, CSV or HTML human readable format
  8. Neoantigen peptides design
    • Neoantigen peptides are designed extracting the amino acid sequence containing the variant (10-15 residues)
  9. HLA-typing with exome or RNA-seq data
    • Normal and tumor or RNA-seq (more accurate) data will be used as input of AmpliHLA to retrieve the HLA alleles of the patient.
  10. HLA-binding affinity calculation and peptide ranking
    • NetMHCpan software will calculate the binding affinities of the designed peptides to HLA alleles, these predictions together with expression data will be used to estimate the potential immunogenicity of the peptides.

Required tools and files

Before starting, you should install in your server or computer the following tools:

And download human genome files and SNP/indels annotations (be always consistent with the genome version and have in mind that new versions could be available):

Previous files (or equivalent ones) can be also downloaded from Google Cloud repository (not tested).

After this long introduction, let’s start the analysis…


Somatic variant calling and tumor neoantigen discovery pipeline

To simplify I’ll consider that all the input files, including the genome and reads ones, are in the working directory and all the output files will be stored into the ‘output’ folder. All the comments are marked with ‘#’ to differentiate them from the commands. Programs will be installed in the standard ‘/opt’ folder in the Linux system and run with 30 threads when it is possible (number of parallel threads/processes should be adjusted depending of your server or computer). Analysis time will depend of the system and the number of parallel processes, in a small server with 30 threads it may take around 15-20 hours to be completed. Steps 8 and 10 are sill not fully automatized.

1. Normal and tumor reads mapping, sorting and indexing

# Index genome FASTA file for BWA mapping:
/opt/bwa-0.7.15/bwa index Homo_sapiens.GRCh38.dna.primary_assembly.fa

# Normal reads mapping, sorting and indexing with BWA-MEM:
/opt/bwa-0.7.15/bwa mem -M -v 1 -t 30 Homo_sapiens.GRCh38.dna.primary_assembly.fa NORMAL_R1.fastq.gz NORMAL_R2.fastq.gz | /opt/samtools-1.6/bin/samtools view -Sb -F 4 -@ 30 | /opt/samtools-1.6/bin/samtools sort -@ 30 > output/NORMAL.bam ; /opt/samtools-1.6/bin/samtools index -@ 30 output/NORMAL.bam output/NORMAL.bai
# Tumor reads mapping, sorting and indexing:
/opt/bwa-0.7.15/bwa mem -M -v 1 -t 30 Homo_sapiens.GRCh38.dna.primary_assembly.fa TUMOR_R1.fastq.gz TUMOR_R2.fastq.gz | /opt/samtools-1.6/bin/samtools view -Sb -F 4 -@ 30 | /opt/samtools-1.6/bin/samtools sort -@ 30 > output/TUMOR.bam ; /opt/samtools-1.6/bin/samtools index -@ 30 output/TUMOR.bam output/TUMOR.bai

2. Removal of duplicated reads

# Removes normal duplicated reads:
java -jar /opt/picard-2.14.0/build/libs/picard.jar MarkDuplicates VERBOSITY=ERROR VALIDATION_STRINGENCY=SILENT REMOVE_DUPLICATES=true I=output/NORMAL.bam O=output/NORMAL.dups_removed.bam M=output/NORMAL.metrics.txt
# Removes tumor duplicated reads:
java -jar /opt/picard-2.14.0/build/libs/picard.jar MarkDuplicates VERBOSITY=ERROR VALIDATION_STRINGENCY=SILENT REMOVE_DUPLICATES=true I=output/TUMOR.bam O=output/TUMOR.dups_removed.bam M=output/TUMOR.metrics.txt

3. Adding sample/group information to the mapped reads

# Adds normal read group, the name of the individual and the phenotype (3998_normal):
java -jar /opt/picard-2.14.0/build/libs/picard.jar AddOrReplaceReadGroups VERBOSITY=ERROR VALIDATION_STRINGENCY=SILENT CREATE_INDEX=true SO=coordinate RGID=3998_normal RGLB=3998_normal RGPL=illumina RGPU=3998_normal RGSM=3998_normal I=output/NORMAL.dups_removed.bam O=output/NORMAL.grouped.bam
# Adds tumor read group, the name of the individual and the phenotype (3998_tumor):
java -jar /opt/picard-2.14.0/build/libs/picard.jar AddOrReplaceReadGroups VERBOSITY=ERROR VALIDATION_STRINGENCY=SILENT CREATE_INDEX=true SO=coordinate RGID=3998_tumor RGLB=3998_tumor RGPL=illumina RGPU=3998_tumor RGSM=3998_tumor I=output/TUMOR.dups_removed.bam O=output/TUMOR.grouped.bam

4. Realignment around indels

# Normal reads realignment around indels using the GATK modules RealignerTargetCreator and IndelRealigner:
java -jar /opt/gatk-3.8/GenomeAnalysisTK.jar -T RealignerTargetCreator -nt 30 -R Homo_sapiens.GRCh38.dna.primary_assembly.fa -I output/NORMAL.grouped.bam -known /Mills_and_1000G_gold_standard.indels.b38.primary_assembly.vcf.gz -o output/NORMAL.intervals ; java -jar /opt/gatk-3.8/GenomeAnalysisTK.jar -T IndelRealigner -maxReads 100000 -maxInMemory 1000000 -R Homo_sapiens.GRCh38.dna.primary_assembly.fa -targetIntervals output/NORMAL.intervals -I output/NORMAL.grouped.bam -known /Mills_and_1000G_gold_standard.indels.b38.primary_assembly.vcf.gz -o output/NORMAL.realign.bam
# Tumor reads realignment around indels using the GATK modules RealignerTargetCreator and IndelRealigner:
java -jar /opt/gatk-3.8/GenomeAnalysisTK.jar -T RealignerTargetCreator -nt 30 -R Homo_sapiens.GRCh38.dna.primary_assembly.fa -I output/TUMOR.grouped.bam -known /Mills_and_1000G_gold_standard.indels.b38.primary_assembly.vcf.gz -o output/TUMOR.intervals ; java -jar /opt/gatk-3.8/GenomeAnalysisTK.jar -T IndelRealigner -maxReads 100000 -maxInMemory 1000000 -R Homo_sapiens.GRCh38.dna.primary_assembly.fa -targetIntervals output/TUMOR.intervals -I output/TUMOR.grouped.bam -known /Mills_and_1000G_gold_standard.indels.b38.primary_assembly.vcf.gz -o output/TUMOR.realign.bam

5. Somatic variant calling with Mutect2 (GATK tool)

# Creates dbSNP VCF index file required by GATK:
gunzip /All_20170710.vcf.gz ; bgzip /All_20170710.vcf ; tabix -p vcf /All_20170710.vcf.gz
# Base quality score recalibration for normal reads alignments (BQSR):
java -jar /opt/gatk-3.8/GenomeAnalysisTK.jar -T BaseRecalibrator -nct 30 -knownSites /All_20170710.vcf.gz -R Homo_sapiens.GRCh38.dna.primary_assembly.fa -I output/NORMAL.realign.bam -o output/NORMAL.bqsr ; java -jar /opt/gatk-3.8/GenomeAnalysisTK.jar -T PrintReads -nct 30 -BQSR output/NORMAL.bqsr -R Homo_sapiens.GRCh38.dna.primary_assembly.fa -I output/NORMAL.realign.bam -o output/NORMAL.recalib.bam
# Base quality score recalibration for tumor reads alignments (BQSR):
java -jar /opt/gatk-3.8/GenomeAnalysisTK.jar -T BaseRecalibrator -nct 30 -knownSites /All_20170710.vcf.gz -R Homo_sapiens.GRCh38.dna.primary_assembly.fa -I output/TUMOR.realign.bam -o output/TUMOR.bqsr ; java -jar /opt/gatk-3.8/GenomeAnalysisTK.jar -T PrintReads -nct 30 -BQSR output/TUMOR.bqsr -R Homo_sapiens.GRCh38.dna.primary_assembly.fa -I output/TUMOR.realign.bam -o output/TUMOR.recalib.bam

# Calls somatic variants (SNPs and indels) via local re-assembly of haplotypes with MuTect2:
java -jar /opt/gatk-3.8/GenomeAnalysisTK.jar -T MuTect2 -nct 30 -D /All_20170710.vcf.gz -cosmic /CosmicCodingMuts.vcf.gz -R Homo_sapiens.GRCh38.dna.primary_assembly.fa -I:normal output/NORMAL.recalib.bam -I:tumor output/TUMOR.recalib.bam -o output/normal_vs_tumor.mutect2.all.vcf

6. Filtering of somatic variants

MuTect2 may retrieve more than 10000 variants, but after filtering the number is reduced around 8-10 times.

# Filters somatic variants called by each method:
filter_somatic_variants.pl -pass -vardepth 10 -varfreq 0.1 -i output/normal_vs_tumor.mutect2.all.vcf -o output/normal_vs_tumor.mutect2.somatic.vcf

7. Annotation of filtered somatic variants

If we obtained around 1000 variants after previous filtering step, the annotation of only variants present in coding regions will reduce the number to few hundreds.

# Annotates final variants and prints the details in a tabulated format file:
annotate_variants.pl -i output/normal_vs_tumor.mutect2.somatic.vcf > output/normal_vs_tumor.mutect2.somatic.txt

8. Neoantigen peptides design

# Creates a FASTA file with the sequences of the peptides (13 aa) containing the somatic mutations previously detected
variants_to_peptides.pl -l 13 -i output/normal_vs_tumor.mutect2.somatic.vcf -o output/normal_vs_tumor.mutect2.peptides.fa

9. HLA-typing with exome or RNA-seq data

# Analyzes WES, or RNA-seq if available, data and retrieves the HLA alleles of the patient that will be used in the last step to calculate the affinities of the peptides.
ampliHLA.pl -thr 30 -ty mapping -r gen -i NORMAL_R1.fastq.gz NORMAL_R2.fastq.gz > amplihla_normal.txt
ampliHLA.pl -thr 30 -ty mapping -r gen -i TUMOR_R1.fastq.gz TUMOR_R2.fastq.gz > amplihla_tumor.txt

10. HLA-binding affinity calculation and peptide rankin

# Calculates and ranks the MHC-binding affinity
netMHCpan -BA -xls -f output/normal_vs_tumor.mutect2.peptides.fa -a HLA-A01:01,HLA-A02:01,HLA-B15:01...


Finally I didn’t get any job in Spain, but I learned a lot ūüôā

 

Counting blue and white bacteria colonies with Python and OpenCV

Last week I was teaching to my Polish students how to use the Python packages NumPy, SciPy and Matplotlib for scientific computing. Most of the examples were about numerical calculations, data visualization and generation of graphs and figures. If you are interested in the topic check the following links with nice tutorials and examples for NumPy, SciPy and Matplotlib.

At the end of the lesson we also explored the capabilities of the scipy.ndimage module for image processing as shown in this nice tutorial. After all, images are pixel matrices that may be represented as NumPy arrays.

After lesson my curiosity led me to OpenCV (Open Source Computer Vision Library), an open-source library for computer vision that includes several hundreds of algorithms.

It is highly recommended to install the last OpenCV version, but you should compile the code yourself as explained here. To use OpenCV in Python, just install its wrapper with PIP installer: pip install opencv-python and import it in any script as: import cv2. In this way you will be able to use any algorithm from OpenCV as Python native but in the background they will be executed as C/C++ code that will make image processing must faster.

After the technical introduction, let’s go to the interesting stuff…

Figure 1. Original blue and white bacteria colonies in Petri dish.

Let’s imagine that we are working at the lab trying to optimize a new cloning protocol. We have dozens of Petri dish with transformed bacteria and we want to check and quantify the transformation efficiency. Each Petri dish will contain blue and white bacteria colonies, white ones will be the bacteria that have inserted our vector disrupting the lacZ gene that generates the blue color.

We want to take photos of the Petri dishes, transfer them to the computer and use a Python script to count automatically the number of blue and white colonies.

For example, we will analyze one image (Figure 1: ‘colonies01.png’) running the following command:

> python bacteria_counter.py -i colonies01.png -o colonies01.out.png
   30 blue colonies
   17 white colonies

Figure 2. Blue and white bacteria colonies are marked in red and green colors respectively as recognized by the Python script.

It will print the number of colonies of each type (blue and white) and it will create an output image (Figure 2: ‘colonies01.out.png’) where blue colonies are marked in red color and white ones in green.

Before showing the code I’ll do a few remarks:

  • Code works quite well but it is not perfect, it fails to recognize 2 small white colonies and also groups other 2 small colonies with 2 bigger ones of the same color. Finally, the script counts 30 blue and 17 white colonies.
  • One of the most tricky parts of the code are the color boundaries to recognize blue and white spots. These thresholds have been manually adjusted (with Photoshop help) before the analysis and they could change with different camera illumination conditions.
  • White colonies are more difficult to identify because their colors are¬†grayish and similar colors are in the blue colonies edges and background. For that reason, image colors are inverted previously to white colonies analysis for an easier recognition.
  • It’s not AI (Artificial Intelligence). I’d call it better ‘Human Intelligence’ because the recognition algorithm thresholds are manually trained. If you are interested in AI and image recognition I can suggest to read the recent article in Nature about skin cancer classification with deep neural networks.
  • I wanted to show a scientific application of image processing, but many other examples are available in internet: recognizing Messi face in a photo, classifying Game Boy cartridges by color

And here is the commented code that performs the magic:

# -*- coding: utf-8 -*-
"""
Bacteria counter

    Counts blue and white bacteria on a Petri dish

    python bacteria_counter.py -i [imagefile] -o [imagefile]

@author: Alvaro Sebastian (www.sixthresearcher.com)
"""

# import the necessary packages
import numpy as np
import argparse
import imutils
import cv2
 
# construct the argument parse and parse the arguments
ap = argparse.ArgumentParser()
ap.add_argument("-i", "--image", required=True,
    help="path to the input image")
ap.add_argument("-o", "--output", required=True,
    help="path to the output image")
args = vars(ap.parse_args())
 
# dict to count colonies
counter = {}

# load the image
image_orig = cv2.imread(args["image"])
height_orig, width_orig = image_orig.shape[:2]

# output image with contours
image_contours = image_orig.copy()

# DETECTING BLUE AND WHITE COLONIES
colors = ['blue', 'white']
for color in colors:

    # copy of original image
    image_to_process = image_orig.copy()

    # initializes counter
    counter[color] = 0

    # define NumPy arrays of color boundaries (GBR vectors)
    if color == 'blue':
        lower = np.array([ 60, 100,  20])
        upper = np.array([170, 180, 150])
    elif color == 'white':
        # invert image colors
        image_to_process = (255-image_to_process)
        lower = np.array([ 50,  50,  40])
        upper = np.array([100, 120,  80])

    # find the colors within the specified boundaries
    image_mask = cv2.inRange(image_to_process, lower, upper)
    # apply the mask
    image_res = cv2.bitwise_and(image_to_process, image_to_process, mask = image_mask)

    ## load the image, convert it to grayscale, and blur it slightly
    image_gray = cv2.cvtColor(image_res, cv2.COLOR_BGR2GRAY)
    image_gray = cv2.GaussianBlur(image_gray, (5, 5), 0)

    # perform edge detection, then perform a dilation + erosion to close gaps in between object edges
    image_edged = cv2.Canny(image_gray, 50, 100)
    image_edged = cv2.dilate(image_edged, None, iterations=1)
    image_edged = cv2.erode(image_edged, None, iterations=1)

    # find contours in the edge map
    cnts = cv2.findContours(image_edged.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
    cnts = cnts[0] if imutils.is_cv2() else cnts[1]

    # loop over the contours individually
    for c in cnts:
        
        # if the contour is not sufficiently large, ignore it
        if cv2.contourArea(c) < 5:
            continue
        
        # compute the Convex Hull of the contour
        hull = cv2.convexHull(c)
        if color == 'blue':
            # prints contours in red color
            cv2.drawContours(image_contours,[hull],0,(0,0,255),1)
        elif color == 'white':
            # prints contours in green color
            cv2.drawContours(image_contours,[hull],0,(0,255,0),1)

        counter[color] += 1
        #cv2.putText(image_contours, "{:.0f}".format(cv2.contourArea(c)), (int(hull[0][0][0]), int(hull[0][0][1])), cv2.FONT_HERSHEY_SIMPLEX, 0.65, (255, 255, 255), 2)

    # Print the number of colonies of each color
    print("{} {} colonies".format(counter[color],color))

# Writes the output image
cv2.imwrite(args["output"],image_contours)

When Blast is not Blast

blastnBLAST changed some time ago for better and faster to BLAST+ version, but along the way small differences were introduced that may confound more than one.

Old BLAST was run with the command ‘blastall. In this way, a protein alignment should be called as ‘blastall -p blastp‘ and the same for nucleotides should be ‘blastall -p blastn‘, and if we want to use MEGABLAST then it exists the different command ‘megablast‘.

Nevertheless, in the ‘new’ BLAST+ version, the alignments of proteins and nucleotides were splitted in two different commands: ‘blastp‘ and ‘blastn‘ (see manual).

Till here, everything looks natural and logic, but not everyone knows that THE DEFAULT OPTION IN THE NEW BLASTN COMMAND IS MEGABLAST.

If we run ‘blastn -help‘ we will obtain the following explanation:

 *** General search options
 -task <String, Permissible values: 'blastn' 'blastn-short' 'dc-megablast'
                'megablast' 'rmblastn' >
   Task to execute
   Default = `megablast'

But not everyone is interested in increasing the alignment speed, because many of us that still use Blastn is because we appreciate its great sensitivity to detect short local alignments. Nowadays Blast can align thousands of sequences in a reasonable time of minutes or even seconds. To increase the alignment speed in alignments with millions of sequences involved there are other better alternatives as Bowtie2.

As CONCLUSION: if we use Blastn and we are interested in a high sensitivity, we should run it as:

blastn -task blastn

If not, we will execute MEGABLAST and we can lose a great sensitivity and risk to not to find the expected alignments.

As example, let’s search homology between the human 2’beta microglobuline (NM_004048.2) and murine one (NM_009735.3) using online version of Blastn with default options (‘Highly similar sequences – megablast’):

But if we change the search option to ‘Somewhat similar sequences (blastn)’:

It is a big difference, megablast does not find any significant similarity, whereas blastn outputs a high similarity alignment with an E-value of 1.5E-56!!!!

You can read my original post in Spanish at ‘Bioinfoperl’ blog.

A list of R commands helpful for science and research

Rlogo.jpg

Here is a comprehensive compilation of many R commands that most of the scientists can find useful for analyzing research data.

Basics

  • Check working directory:
getwd()
  • Change working directory:
setwd("path")
  • Show help about a command (quotes are important):
help("command")
?"command"
  • Show the objects and variables of the working space:
ls()
  • Remove an object from the working space:
rm(x)
rm(list=ls()) # Removes all the objects
  • Save the working space:
save.image() # in the file .RData at the working directory
save(object list,file="myfile.RData") # saves the object list into the chosen file
  • Reload a working space:
load("myfile.RData")
  • Quit R:
q()

Continue reading

© 2018 Sixth researcher

Theme by Anders NorenUp ↑