UMI Barcoded Illumina MiSeq 2x250 BCR mRNA

Overview of Experimental Data

This example uses publicly available data from:

B cells populating the multiple sclerosis brain mature in the draining cervical lymph nodes.
Stern JNH, Yaari G, and Vander Heiden JA, et al.
Sci Transl Med. 2014. 6(248):248ra107. doi:10.1126/scitranslmed.3008879.

Which may be downloaded from the NCBI Sequence Read Archive under BioProject accession ID: PRJNA248475. For this example, we will use the first 25,000 sequences of sample M12 (accession: SRR1383456), which may downloaded downloaded using fastq-dump from the SRA Toolkit:

fastq-dump --split-files -X 25000 SRR1383456

Primers sequences are available online at the supplemental website for the publication.

Read Configuration

../_images/Stern2014_ReadConfiguration.svg

Schematic of the Illumina MiSeq 2x250 paired-end reads with UMI barcodes. Each 250 base-pair read was sequenced from one end of the target cDNA, so that the two reads together cover the entire variable region of the Ig heavy chain. The V(D)J reading frame proceeds from the start of read 2 to the start of read 1. Read 1 is in the opposite orientation (reverse complement), and contains a 15 nucleotide UMI barcode preceding the C-region primer sequence.

Example Data

We have hosted a small subset of the data (Accession: SRR1383456) on the pRESTO website in FASTQ format with accompanying primer files. The sample data set and workflow script may be downloaded from here:

Stern, Yaari and Vander Heiden et al, 2014 Example Files

Overview of the Workflow

In the following sections, we demonstrate each step of the workflow to move from raw sequence reads to a fully annotated repertoire of complete V(D)J sequences. The workflow is divided into four high-level tasks:

A graphical representation of the workflow along with the corresponding sequence of pRESTO commands is shown below.

Flowchart

../_images/Stern2014_Flowchart.svg

Flowchart of processing steps. Each pRESTO tool is shown as a colored box. The workflow is divided into four primary tasks: (red) quality control, UMI annotation and primer masking; (orange) generation of UMI consensus sequences; (green) paired-end assembly of UMI consensus sequences; and (blue) deduplication and filtering to obtain the high-fidelity repertoire. Grey boxes indicate the initial and final data files. The intermediate files output by each tool are not shown for the sake of brevity.

Commands

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
#!/usr/bin/env bash
FilterSeq.py quality -s SRR1383456_1.fastq -q 20 --outname MS12_R1 --log FS1.log
FilterSeq.py quality -s SRR1383456_2.fastq -q 20 --outname MS12_R2 --log FS2.log
MaskPrimers.py score -s MS12_R1_quality-pass.fastq -p Stern2014_CPrimers.fasta \
    --start 15 --mode cut --barcode --outname MS12_R1 --log MP1.log
MaskPrimers.py score -s MS12_R2_quality-pass.fastq -p Stern2014_VPrimers.fasta \
    --start 0 --mode mask --outname MS12_R2 --log MP2.log
PairSeq.py -1 MS12_R1_primers-pass.fastq -2 MS12_R2_primers-pass.fastq \
    --1f BARCODE --coord sra
BuildConsensus.py -s MS12_R1_primers-pass_pair-pass.fastq --bf BARCODE --pf PRIMER \
    --prcons 0.6 --maxerror 0.1 --maxgap 0.5 --outname MS12_R1 --log BC1.log
BuildConsensus.py -s MS12_R2_primers-pass_pair-pass.fastq --bf BARCODE --pf PRIMER \
    --maxerror 0.1 --maxgap 0.5 --outname MS12_R2 --log BC2.log
PairSeq.py -1 MS12_R1_consensus-pass.fastq -2 MS12_R2_consensus-pass.fastq \
    --coord presto
AssemblePairs.py align -1 MS12_R2_consensus-pass_pair-pass.fastq \
    -2 MS12_R1_consensus-pass_pair-pass.fastq --coord presto --rc tail \
    --1f CONSCOUNT --2f CONSCOUNT PRCONS --outname MS12 --log AP.log
ParseHeaders.py collapse -s MS12_assemble-pass.fastq -f CONSCOUNT --act min
CollapseSeq.py -s MS12*reheader.fastq -n 20 --inner --uf PRCONS \
    --cf CONSCOUNT --act sum --outname MS12
SplitSeq.py group -s MS12_collapse-unique.fastq -f CONSCOUNT --num 2 --outname MS12
ParseHeaders.py table -s MS12_atleast-2.fastq -f ID PRCONS CONSCOUNT DUPCOUNT
ParseLog.py -l FS1.log FS2.log -f ID QUALITY
ParseLog.py -l MP1.log MP2.log -f ID PRIMER BARCODE ERROR
ParseLog.py -l BC1.log BC2.log -f BARCODE SEQCOUNT CONSCOUNT PRIMER PRCONS PRCOUNT \
    PRFREQ ERROR
ParseLog.py -l AP.log -f ID LENGTH OVERLAP ERROR PVALUE FIELDS1 FIELDS2

Download Commands

Quality control, UMI annotation and primer masking

Removal of low quality reads

Quality control begins with the identification and removal of low-quality reads using the quality subcommand of the FilterSeq tool. In this example, reads with mean Phred quality scores less than 20 (-q 20) are removed:

2
3
FilterSeq.py quality -s SRR1383456_1.fastq -q 20 --outname MS12_R1 --log FS1.log
FilterSeq.py quality -s SRR1383456_2.fastq -q 20 --outname MS12_R2 --log FS2.log

The ParseLog tool is then used to extract results from the FilterSeq logs into tab-delimited files:

24
ParseLog.py -l FS1.log FS2.log -f ID QUALITY

Extracting the following information from the log:

Field Description
ID Sequence name
QUALITY Quality score

UMI annotation and masking of primer regions

Next, the score subcommand of MaskPrimers is used to identify and remove the PCR primers for both reads. When dealing with Ig sequences, it is important to cut or mask the primers, as B cell receptors are subject to somatic hypermutation (the accumulation of point mutations in the DNA) and degenerate primer matches can look like mutations in downstream applications. The MaskPrimers tool is also used to annotate each read 1 sequence with the 15 nucleotide UMI that precedes the C-region primer (MaskPrimers score --barcode):

4
5
6
7
MaskPrimers.py score -s MS12_R1_quality-pass.fastq -p Stern2014_CPrimers.fasta \
    --start 15 --mode cut --barcode --outname MS12_R1 --log MP1.log
MaskPrimers.py score -s MS12_R2_quality-pass.fastq -p Stern2014_VPrimers.fasta \
    --start 0 --mode mask --outname MS12_R2 --log MP2.log

To summarize these steps, the ParseLog tool is used to build a tab-delimited file from the MaskPrimers log:

25
ParseLog.py -l MP1.log MP2.log -f ID PRIMER BARCODE ERROR

Containing the following information:

Field Description
ID Sequence name
PRIMER Primer name
BARCODE UMI sequence
ERROR Primer match error rate

Note

For this data set the UMI is immediately upstream of the C-region primer. Another common approach for UMI barcoding involves placing the UMI immediately upstream of a 5’RACE template switch site. Modifying the workflow is simple for this case. You just need to replace the V-region primers with a fasta file containing the TS sequences and move the --barcode argument to the appropriate read:

MaskPrimers.py score -s R1_quality-pass.fastq -p CPrimers.fasta \
    --start 0 --mode cut --outname R1 --log MP1.log
MaskPrimers.py score -s R2_quality-pass.fastq -p TSSites.fasta \
    --start 17 --barcode --mode cut --maxerror 0.5 \
    --outname R2 --log MP2.log

In the above we have moved the UMI annotation to read 2, increased the allowable error rate for matching the TS site (--maxerror 0.5), cut the TS site (--mode cut), and increased the size of the UMI from 15 to 17 nucleotides (--start 17).

Generation of UMI consensus sequences

Copying the UMI annotation across paired-end files

In this task, a single consensus sequence is constructed for each set of reads annotated with the same UMI barcode. As the UMI barcode is part of read 1, the BARCODE annotation identified by MaskPrimers must first be copied to the read 2 mate-pair of each read 1 sequence. Propogation of annotations between mate pairs is performed using PairSeq which also removes unpaired reads and ensures that paired reads are sorted in the same order across files:

8
9
PairSeq.py -1 MS12_R1_primers-pass.fastq -2 MS12_R2_primers-pass.fastq \
    --1f BARCODE --coord sra

Note

For both the PairSeq and AssemblePairs commands using the correct --coord argument is critical for matching mate-pairs. If this was raw data from Illumina, rather than data downloaded from SRA/ENA, then the appropriate argument would be --coord illumina.

Note

If you have followed the 5’RACE modification above, then you must also modify the first PairSeq step to copy the UMI from read 2 to read 1, instead of vice versa (--2f BARCODE):

PairSeq.py -1 R1_primers-pass.fastq -2 R2_primers-pass.fastq \
    --2f BARCODE --coord sra

Multiple alignment of UMI read groups

Before generating a consensus for a set of reads sharing a UMI barcode, the sequences must be properly aligned. Sequences may not be aligned if more than one PCR primer is identified in a UMI read group - leading to variations in the the start positions of the reads. Ideally, each set of reads originating from a single mRNA molecule should be amplified with the same primer. However, different primers in the multiplex pool may be incorporated into the same UMI read group during amplification if the primers are sufficiently similar. This type of primer misalignment can be corrected using the AlignSets tool. In the example data used here, this step was not necessary due to the aligned primer design for the 45 V-region primers, though this does require that the V-region primers be masked, rather than cut, during the MaskPrimers step (--mode mask).

See also

If your data requires alignment, then you can create multiple aligned UMI read groups as follows:

AlignSets.py muscle -s R1_primers-pass_pair-pass.fastq --bf BARCODE \
    --exec ~/bin/muscle --outname R1 --log AS1.log
AlignSets.py muscle -s R2_primers-pass_pair-pass.fastq --bf BARCODE \
    --exec ~/bin/muscle --outname R2 --log AS2.log

Where the --bf BARCODE defines the field containing the UMI and --exec ~/bin/muscle is the location of the MUSCLE executable.

For additional details see the section on fixing UMI alignments.

Generating UMI consensus reads

After alignment, a single consensus sequence is generated for each UMI barcode using BuildConsensus:

10
11
12
13
BuildConsensus.py -s MS12_R1_primers-pass_pair-pass.fastq --bf BARCODE --pf PRIMER \
    --prcons 0.6 --maxerror 0.1 --maxgap 0.5 --outname MS12_R1 --log BC1.log
BuildConsensus.py -s MS12_R2_primers-pass_pair-pass.fastq --bf BARCODE --pf PRIMER \
    --maxerror 0.1 --maxgap 0.5 --outname MS12_R2 --log BC2.log

To correct for UMI chemistry and sequencing errors, UMI read groups having high error statistics (mismatch rate from consensus) are removed by specifiying the --maxerror 0.1 threshold. As the accuracy of the primer assignment in read 1 is critical for correct isotype identification, additional filtering of read 1 is carried out during this step. Specifying the --prcons 0.6 threshold: (a) removes individual sequences that do not share a common primer annotation with the majority of the set, (b) removes entire read groups which have ambiguous primer assignments, and (c) constructs a consensus primer assignment for each UMI.

Note

The --maxgap 0.5 argument tells BuildConsensus to use a majority rule to delete any gap positions which occur in more than 50% of the reads. The --maxgap argument is not really necessary for this example data set as we did not perform a multiple alignment of the UMI read groups. However, if you have performed an alignment, then use of --maxgap during consensus generation is highly recommended.

The ParseLog tool is then used to build a tab-delimited file contain the consensus results:

26
ParseLog.py -l BC1.log BC2.log -f BARCODE SEQCOUNT CONSCOUNT PRIMER PRCONS PRCOUNT \

With the following annotations:

Field Description
BARCODE UMI sequence
SEQCOUNT Number of total reads in the UMI group
CONSCOUNT Number of reads used for the UMI consensus
PRIMER Set of primer names in the UMI group
PRCONS Consensus primer name
PRCOUNT Count of primers in the UMI group
PRFREQ Frequency of primers in the UMI group
ERROR Average mismatch rate from consensus

Paired-end assembly of UMI consensus sequences

Syncronizing paired-end files

Following UMI consensus generation, the read 1 and read 2 files may again be out of sync due to differences in UMI read group filtering by BuildConsensus. To synchronize the reads another instance of PairSeq must be run, but without any annotation manipulation:

14
15
PairSeq.py -1 MS12_R1_consensus-pass.fastq -2 MS12_R2_consensus-pass.fastq \
    --coord presto

Assembling UMI consensus mate-pairs

Once the files have been synchronized, each paired-end UMI consensus sequence is assembled into a full length Ig sequence using the align subcommand of AssemblePairs:

16
17
18
AssemblePairs.py align -1 MS12_R2_consensus-pass_pair-pass.fastq \
    -2 MS12_R1_consensus-pass_pair-pass.fastq --coord presto --rc tail \
    --1f CONSCOUNT --2f CONSCOUNT PRCONS --outname MS12 --log AP.log

During assembly, the consensus isotype annotation (PRCONS) from read 1 and the number of reads used to define the consensus sequence (CONSCOUNT) for both reads are propagated into the annotations of the full length Ig sequence (--1f CONSCOUNT --2f CONSCOUNT PRCONS.

ParseLog is then uses to extract the results from the AssemblePairs log into a tab-delimited file:

27
    PRFREQ ERROR

Containing the following information:

Field Description
ID Sequence name (UMI)
LENGTH Length of the assembled sequence
OVERLAP Length of the overlap between mate-pairs
ERROR Mismatch rate of the overlapping region
PVALUE P-value for the assembly
FIELDS1 Annotations copied from read 2 into the assembled sequence
FIELDS2 Annotations copied from read 1 into the assembled sequence

See also

Depending on the amplicon length in your data, not all mate-pairs may overlap. For the sake of simplicity, we have excluded a demonstration of assembly in such cases. pRESTO provides a couple approaches to deal with such reads. The reference subcommand of AssemblePairs can use the ungapped V-region reference sequences to properly space non-overlapping reads. Or, if all else fails, the join subcommand can be used to simply stick mate-pairs together end-to-end with some intervening gap.

Deduplication and filtering

Combining UMI read group size annotations

In the final stage of the workflow, the high-fidelity Ig repertoire is obtained by a series of filtering steps. First, the annotation specifying the number of raw reads used to build each sequence (-f CONSCOUNT) is updated to be the minimum (--act min) of the forward and reverse reads using the collapse subcommand of ParseHeaders:

19
ParseHeaders.py collapse -s MS12_assemble-pass.fastq -f CONSCOUNT --act min

Removal of duplicate sequences

Second, duplicate nucleotide sequences are removed using the CollapseSeq tool with the requirement that duplicate sequences share the same isotype primer (--uf PRCONS). The duplicate removal step also removes sequences with a high number of interior N-valued nucleotides (-n 20 and --inner) and combines the read counts for each UMI read group (--cf CONSCOUNT and --act sum).

20
21
CollapseSeq.py -s MS12*reheader.fastq -n 20 --inner --uf PRCONS \
    --cf CONSCOUNT --act sum --outname MS12

Filtering to sequences with at least two representative reads

Finally, unique sequences are filtered to those with at least 2 contributing sequences using the group subcommand of SplitSeq, by splitting the file on the CONSCOUNT annotation with a numeric threshold (-f CONSCOUNT and --num 2):

22
SplitSeq.py group -s MS12_collapse-unique.fastq -f CONSCOUNT --num 2 --outname MS12

Creating an annotation table

For further analysis, the annotations of the final repertoire are then converted to into a table using the table subcommand of ParseHeaders:

23
ParseHeaders.py table -s MS12_atleast-2.fastq -f ID PRCONS CONSCOUNT DUPCOUNT

Output files

The final set of sequences, which serve as input to a V(D)J reference aligner (Eg, IMGT/HighV-QUEST or IgBLAST), and tables that can be plotted for quality control are:

File Description
M12_collapse-unique.fastq Total unique sequences
M12_atleast-2.fastq Unique sequences represented by at least 2 reads
M12_atleast-2_headers.tab Annotation table of the atleast-2 file
FS1_table.tab Table of the read 1 FilterSeq log
FS2_table.tab Table of the read 2 FilterSeq log
MP1_table.tab Table of the C-region MaskPrimers log
MP2_table.tab Table of the V-region MaskPrimers log
BC1_table.tab Table of the read 1 BuildConsensus log
BC2_table.tab Table of the read 2 BuildConsensus log
AP_table.tab Table of the AssemblePairs log

A number of other intermediate and log files are generated during the workflow, which allows easy tracking/reversion of processing steps. These files are not listed in the table above.

Performance

Example performance statistics for a comparable, but larger, MiSeq workflow are presented below. Performance was measured on a 64-core system with 2.3GHz AMD Opteron(TM) 6276 processors and 512GB of RAM, with memory usage measured at peak utilization. The data set contained 1,723,558 x 2 raw reads, and required matching of 1 constant region primer, 45 V-segment primers, and averaged 24.3 reads per UMI.

Line Tool Reads Cores MB Minutes
01 R1: FilterSeq.py quality 1,723,558 10 1,219 13.0
02 R2: FilterSeq.py quality 1,723,558 10 1,219 12.8
03 R1: MaskPrimers.py score 1,722,116 10 1,221 18.9
04 R2: MaskPrimers.py score 1,684,050 10 1,221 46.7
06 PairSeq.py 1,665,584 1 4,734 25.5
07 R1: BuildConsensus.py 1,565,017 10 1,228 50.1
08 R2: BuildConsensus.py 1,565,017 10 1,229 58.6
11 AssemblePairs.py align 66,285 10 358 3.0
13 ParseHeaders.py collapse 56,104 1 88 0.4
14 CollapseSeq.py 55,480 1 822 0.7
15 SplitSeq.py group 51,047 1 88 0.2