User guide


IgDiscover works on a single library at a time. It works within an “analysis directory” for the library, which contains all intermediate and result files.

To start an analysis, you need:

  1. A FASTA or FASTQ file with single-end reads or two FASTQ files with paired-end reads (also, the files must be gzip-compressed)
  2. A database of V/D/J genes (three FASTA files named V.fasta, D.fasta, J.fasta)
  3. A configuration file that describes the library

If you do not have a V/D/J database, yet, you may want to read the section about how to obtain V/D/J sequences.

To run an analysis, proceed as follows.


If you are on OS X, it may be necessary to run export SHELL=/bin/bash before continuing.

  1. Create and initialize the analysis directory.

    First, pick a name for your analysis. We will use myexperiment in the following. Then run igdiscover init:

    igdiscover init myexperiment

    A dialog will appear and ask for the file with the first (forward) reads. Find your compressed FASTQ file that contains them and select it. Typical file names may be Library1_S1_L001_R1_001.fastq.gz or mylibrary.1.fastq.gz. You do not need to choose the second read file! It is found automatically.

    Next, choose the directory with your database. The directory must contain the three files V.fasta, D.fasta, J.fasta. These files contain the V, D, J gene sequences, respectively. Even if have have only light chains in your data, a D.fasta file needs to be provided, just use one with the heavy chain D gene sequences.

    If you do not want a graphical user interface, use the two command-line parameters --db and --reads1 to provide this information instead:

    igdiscover init --db path/to/my/database/ --reads1 mylibrary.1.fastq.gz myexperiment

    Again, the second reads file will be found automatically. Use --single-reads instead of --reads1 if you have single-end reads or a dataset with already merged reads. For --single-reads, a FASTA file (not only FASTQ) is also allowed. In any case, an analysis directory named myexperiment will have been created.

  2. Adjust the configuration file

    The previous step created a configuration file named myexperiment/igdiscover.yaml, which you may need to adjust. In particular, the number of discovery rounds is set to 3 by default, which takes a long time. Reducing this to 2 or even 1 often works just as well.

  3. Run the analysis

    Change into the newly created analysis directory and run the analysis:

    igdiscover run

    Depending on the size of your library, your computer, and the number of iterations, this will now take from a few hours to a day. See the running IgDiscover section for more fine-grained control over what to run and how to resume the process if something failed.

Obtaining a V/D/J database

We use the term “database” to refer to three FASTA files that contain the sequences for the V, D and J genes. IMGT provides sequences for download. For discovering new VH genes, for example, you need to get the IGHV, IGHD and IGHJ files of your species. As IgDiscover uses this only as a starting point, using a similar species will also work.

When using an IMGT database, it is very important to change the long IMGT sequence headers to short headers as IgBLAST does not accept the long headers. We recommend using the program If you installed IgDiscover from Conda, the script is already installed and you can run it by typing the name. It is also available on the IgBlast FTP site.

Run it for all three downloaded files, and then rename files appropritely to make sure that they named V.fasta, D.fasta and J.fasta.

You always need a file with D genes even if you analyze light chains.

In case you have used IgBLAST previously, note that there is no need to run the makeblastdb tool as IgDiscover will do that for you.

Input data requirements

Paired-end or single-end data

IgDiscover can process input data of three different types:

  • Paired-end reads in gzipped FASTQ format,
  • Single-end reads in gzipped FASTQ format,
  • Single-end reads in gzipped FASTA format.

IgDiscover was tested mainly on paired-end Illumina MiSeq reads (2x300bp), but it can also handle 454 and Ion Torrent data.

Depending on the input file type, use a variant of one of the following commands to initialize the analysis directory:

igdiscover init --single-reads=file.fasta.gz  --database=my-database-dir/ myexperiment
igdiscover init --reads1=file.1.fasta.gz  --database=my-database-dir/ myexperiment
igdiscover init --reads1=file.1.fastq.gz  --database=my-database-dir/ myexperiment

Read layout

Paired-end reads are first merged and then processed in the same way as single-end reads. Reads that could not be merged are discarded. Single-end reads and merged paired-end reads are expected to follow this structure (from 5’ to 3’):

  • The forward primer sequence. This is optional.
  • A random barcode (molecular identifier). This is optional. Set the configuration option barcode_length_5p to 0 if you don’t have random barcodes or if you don’t want the program to use them.
  • Optionally, a run of G nucleotides. This is an artifact of the RACE protocol (Rapid amplification of cDNA ends). If you have this, set race_g to true in the configuration file.
  • 5’ UTR
  • Leader
  • Re-arranged V, D and J gene sequences for heavy chains; only V and J for light chains
  • An optional random barcode. Set the configuration option barcode_length_3p to the length of this barcode. You can currently not have both a 5’ and a 3’ barcode.
  • The reverse primer. This is optional.

We use IgBLAST to detect the location of the V, D, J genes through the igdiscover igblast subcommand. The G nucleotides after the barcode are split off if the configuration specifies race_g: true. The leader sequence is detected by looking for a start codon near 60 bp upstream of the start of the V gene match.


The igdiscover init command creates a configuration file igdiscover.yaml in the analysis directory. To configure your analysis, change that file with a text editor before running the analysis with igdiscover run.

The syntax should be mostly self-explanatory. The file is in YAML format, but you will not need to learn that. Just follow the examples given in the file. A few rules that may be good to know are the following ones:

  1. Lines starting with the # symbol are comments (they are ignored)

  2. A configuration option that is meant to be switched on or off will say something like stranded: false if it is off. Change this to stranded: true to switch the option on (and vice versa).

  3. The primer sequences are given as a list, and must be written in a certain way - one sequence per line, and a - (dash) in front, like so:


    Even if you have only one primer sequence, you still need to use this syntax.

To find out what the configuration options achieve, see the explanations in the configuration file itself.

The main parameters parameters that may require adjusting are the following.

The iterations option sets the number of rounds of V gene discovery that will be performed. By default, three iterations are run. Even with a very restricted starting V database (for example with only a single V gene sequence), this is usually sufficient to identify most novel germline sequences.

When the starting database is more complete, for example, when analyzing a human IgM library with the current IMGT heavy chain database, a single iteration may be sufficient to produce an individualized database.

If you do not want to discover any new genes and only want to produce an expression profile, for example, then use iterations: 0.

The ignore_j option should be set to true when producing a V gene database for a species where J sequences are unknown:

ignore_j: true

Setting the parameters stranded, forward_primers and reverse_primers to the correct values can be used to remove 5’ and 3’ primers from the sequences. Doing this is not strictly necessary for IgDiscover. It is simplest if you do not specify any primer sequences.

Pregermline and germline filter criteria

This provides IgDiscover with stringency requirements for V gene discovery that enable the program to filter out false positives. Usually the ”pregermline filter” can be used in the default mode since all these sequences will be subsequently passed to the higher stringency ”germline filter” where the criteria are set to maximize stringency. Here is how it looks in the configuration file:

  unique_cdr3s: 2      # Minimum number of unique CDR3s (within exact matches)
  unique_js: 2         # Minimum number of unique J genes (within exact matches)
  check_motifs: false  # Check whether 5' end starts with known motif
  whitelist: true      # Add database sequences to the whitelist
  cluster_size: 0      # Minimum number of sequences assigned to cluster
  differences: 0       # Merge sequences if they have at most this number of differences
  allow_stop: true     # Whether to allow non-productive sequences containing stop codons
  cross_mapping_ratio: 0.02  # Threshold for removal of cross-mapping artifacts (set to 0 to disable)
  allele_ratio: 0.1    # Required minimum ratio between alleles of a single gene

# Filtering criteria applied to candidate sequences in the last iteration.
# These should be more strict than the pre_germline_filter criteria.
  unique_cdr3s: 5      # Minimum number of unique CDR3s (within exact matches)
  unique_js: 3         # Minimum number of unique J genes (within exact matches)
  check_motifs: false  # Check whether 5' end starts with known motif
  whitelist: true      # Add database sequences to the whitelist
  cluster_size: 100    # Minimum number of sequences assigned to cluster
  differences: 0       # Merge sequences if they have at most this number of differences
  allow_stop: false    # Whether to allow non-productive sequences containing stop codons
  cross_mapping_ratio: 0.02  # Threshold for removal of cross-mapping artifacts (set to 0 to disable)
  allele_ratio: 0.1    # Required minimum ratio between alleles of a single gene

Factors that affect germline discovery include library source (IgM vs IgK, IgL or IgG) library size, sequence error rate and individual genomic factors (for example the number of J segments present in an individual).

In general, setting a higher cutoff of unique_cdr3s and unique_js will minimize the number of false positives in the output. Example:

unique_cdr3s: 10      # Minimum number of unique CDR3s (within exact matches)
unique_js: 4          # Minimum number of unique J genes (within exact matches)

If the differences parameter is set to a value higher than 0, the germline filter inspects clusters of sequences that are closely related (when the edit distance between them is at most differences) and retains only the most common sequence of each cluster. Previously, we believed this would removes some false positives due to accumulated random sequence errors of highly expressed alleles that otherwise would pass the cutoff criteria. However, we found out that we miss true positives, in particular if there are two alleles in the sample that differ in only a single nucleotide. We have now implemented other measures to avoid false positives and recommend against setting the differences to something other than 0.

Read also about the cross mapping, for which germline filtering corrects, and about the germline filters.

Changed in version The: default for the differences setting was changed from 1 to 0.

Running IgDiscover

Resuming failed runs

The command igdiscover run, which is used to start the pipeline, can also be used to resume execution if there was an interruption (a transient failure). Reasons for interruptions might be:

  • Ctrl+C was pressed on the keyboard
  • A full harddisk
  • If running on a cluster, the program may have been terminated because it exceeded its allocated time
  • Too little RAM
  • Power loss

To resume execution after you have fixed the problem, go to the analysis directory and run igdiscover run again. It will skip the steps that have already finished successfully. This capability comes from the workflow management system snakemake, on which igdiscover run is based. Snakemake will determine automatically which steps need to be re-run in order to get to a full result and then run only those.

Alterations to the configuration file after an interruption are possible, but affect only steps that have not already finished successfully. For example, assume you interrupted a run with Ctrl+C after it is already past the step in which barcodes are removed. Then, even if you change the barcode length in the configuration, the barcode removal step will not be re-run when you resume the pipeline and the previous barcode length is in effect. See also the next section.

Changing parameters and re-running parts of the pipeline

When you experiment with parameters in the igdiscover.yaml file, such as germline filtering criteria, you do not need to re-run the entire pipeline from the beginning, but can re-use the results that already exist. This can save a lot of processing time, in particular when you avoid re-running IgBLAST in this way.

As described in the previous section, igdiscover run automatically figures out which files need to be re-created if a run was interrupted. Unfortunately, this mechanism is currently not smart enough to also look for changes in the igdiscover.yaml file. Thus, if the full pipeline has finished successfully, then re-running igdiscover run will just print the message Nothing to be done. even after you have changed the configuration file.

You will therefore need to know yourself which file you want to regenerate. Then follow the following steps. Note that these will remove parts of the existing results, and if you need to keep them, make a copy of your analysis directory first.

  1. Change the configuration setting.
  2. Delete the file that needs to be re-generated. Assume it is filename
  3. Run igdiscover run filename to re-create the file. Only that file will be created, not the ones that usually would be created afterwards.
  4. Optionally, run igdiscover run (without a file name this time) to update the remaining files (those that depend on the file that was just updated).

For example, assume you want to modify some germline filtering setting and then re-run the pipeline. Change the setting in your igdiscover.yaml, then run these commands:

rm iteration-01/
igdiscover run iteration-01/

The above will have regenerated the iteration-01/ file and also the iteration-01/new_V_germline.fasta file since they are generated by the same script. If you want to update any other files, then also run

igdiscover run

The analysis directory

IgDiscover writes all intermediate files, the final V gene database, statistics and plots into the analysis directory that was created with igdiscover init. Inside that directory, there is a final/ subdirectory that contains the analysis results.

These are the files and subdirectories that can be found in the analysis directory. Subdirectories are described in detail below.

The configuration file. Make sure to adjust this to your needs as described above.
reads.1.fastq.gz, reads.2.fastq.gz
Symbolic links to the raw paired-end reads.
The input V/D/J database (as three FASTA files). The files are a copy of the ones you selected when running igdiscover init.
Processed reads (merged, de-duplicated etc.)
Iteration-specific analysis directory, where “xx” is a number starting from 01. Each iteration is run in one of these directories. The first iteration (in iteration-01) uses the original input database, which is also found in the database/ directory. The database is updated and then used as input for the next iteration.
After the last iteration, IgBLAST is run again on the input sequences, but using the final database (the one created in the very last iteration). This directory contains all the results, such as plots of the repertoire profiles. If you set the number of iterations to 0 in the configuration file, this directory is the only one that is created.

Final results

Final results are found in the final/ subdirectory of the analysis directory.

These three files represent the final, individualized V/D/J database found by IgDiscover. The D and J files are copies of the original starting database; they are not updated by IgDiscover.
A dendrogram of all V sequences in the individualized database.
IgBLAST result (compressed) of running IgBLAST with the discovered database.
V/D/J gene assignments and other information for each sequence. The file is created by parsing the IgBLAST output in the igblast.txt.gz file. This is a table that contains one row for each input sequence. See below for a detailed description of the columns.
Filtered V/D/J gene assignments. This is the same as the file mentioned above, but with low-quality assignments filtered out. Run igdiscover filter --help to see the filtering criteria.
final/, final/V_usage.pdf
The V gene expression counts, derived from the IgBLAST results. The .tab file contains the counts as a table, while the pdf file contains a plot of the same values.
A PDF with one page per V gene/allele. Each page shows a histogram of the percentage differences for that gene.
This is a directory that contains one PNG file for each discovered gene/allele. Each image shows a clusterplot of all the sequences assigned to that gene. Note that the shown clusterplots are by default restricted to showing only at most 300 sequences, while the actual clustering used by IgDiscover uses 1000 sequences.

If you are interested in the results of each iteration, you can inspect the iteration-xx/ directories. They are structured in the same way as the final/ subdirectory, except that the results are based on the intermediate databases of that iteration. They also contain the following additional files.

A table with candidate novel V alleles (or genes). This is a list of sequences found through the windowing strategy or linkage cluster analysis, as discussed in our paper.
iteration-xx/new_V_germline.fasta, iteration-xx/new_V_pregermline.fasta
The discovered list of V genes for this iteration. The file is created from the file by applying either the germline or pre-germline filter. The file resulting from application of the germline filter is used in the last iteration only. The file resulting from application of the pre-germline filter is used in earlier iterations.

Other files

For completeness, here is a description of the files in the reads/ and stats/ directories. They are created during pre-processing and are not iteration specific.

reads/1-limited.1.fastq.gz, reads/1-limited.1.fastq.gz
Input reads file limited to the first N entries. This is just a symbolic link to the input file if the limit configuration option is not set.
Reads merged with PEAR or FLASH
Merged reads with 5’ primer sequences removed. (This file is automatically removed when it is not needed anymore.)
Merged reads with 5’ and 3’ primer sequences removed.
Merged, primer-trimmed sequences converted to FASTA, and too short sequences removed. (This file is automatically removed when it is not needed anymore.)
Fully pre-processed sequences. That is, filtered sequences without duplicates (using VSEARCH)
Statistics of pre-processed sequences.
stats/readlengths.txt, stats/readlengths.pdf
Histogram of the lengths of pre-processed sequences (created from reads/sequences.fasta)

Format of output files

This file is a gzip-compressed table with tab-separated values. It is created by the igdiscover igblast subcommand and is the result of parsing raw output from IgBLAST. It contains a few additional columns that do not come directly from IgBLAST. In particular, the CDR3 sequence is detected, the sequence before the V gene match is split into UTR and leader, and the RACE-specific run of G nucleotides is also detected. The first row is a header row with column names. Each subsequent row describes the IgBLAST results for a single pre-processed input sequence.

Note: This file is typically quite large. LibreOffice can open the file directly (even though it is compressed), but make sure you have enough RAM.


How many copies of input sequence this query sequence represents. Copied from the ;size=3; entry in the FASTA header field that is added by VSEARCH -derep_fulllength.
V_gene, D_gene, J_gene
V/D/J gene match for the query sequence
whether the sequence contains a stop codon (either “yes” or “no”)


V_covered, D_covered, J_covered
percentage of bases of the reference gene that is covered by the bases of the query sequence
V_evalue, D_evalue, J_evalue
E-value of V/D/J hit
rate of somatic hypermutation (actually, an error rate)
V_errors, J_errors
Absolute number of errors (differences) in the V and J gene match
Sequence of the 5’ UTR (the part before the V gene match up to, but not including, the start codon)
Leader sequence (the part between UTR and the V gene match)
CDR1_nt, CDR1_aa, CDR2_nt, CDR2_aa, CDR3_nt, CDR3_aa
nucleotide and amino acid sequence of CDR1/2/3
V_nt, V_aa
Nucleotide and amino acid sequence of V gene match
Start coordinate of CDR3 within V_nt. Set to zero if no CDR3 was detected. Comparisons involving the V gene ignore those V bases that are part of the CDR3.
V_end, VD_junction, D_region, DJ_junction, J_start
nucleotide sequences for various match regions
name, barcode, race_G, genomic_sequence
see the following explanation

The UTR, leader, barcode, race_G and genomic_sequence columns are filled in the following way.

  1. Split the 5’ end barcode from the sequence (if barcode length is zero, this will be empty), put it in the barcode column.
  2. Remove the initial run of G bases from the remaining sequence, put that in the race_G column.
  3. The remainder is put into the genomic_sequence column.
  4. If there is a V gene match, take the sequence before it and split it up in the following way. Search for the start codon and write the part before it into the UTR column. Write the part starting with the start column into the leader column.

This table is the same as the table, except that rows containing low-quality matches have been filtered out. Rows fulfilling any of the following criteria are filtered:

  • The J gene was not assigned
  • A stop was codon found
  • The V gene coverage is less than 90%
  • The J gene coverage is less than 60%
  • The V gene E-value is greater than 10-3

This table contains the candidates for novel V genes found by the discover subcommand. As the other files, it is a text file in tab-separated values format, with the first row containing the column headings. It can be opened directly in LibreOffice, for example.

Candidates are found by inspecting all the sequences assigned to a database gene, and clustering them in multiple ways. The candidate sequences are found by computing a consensus from each found cluster.

Each row describes a single candidate, but possibly multiple clusters. If there are multiple clusters from a single gene that lead to the same consensus sequence, then they get only one row. The cluster column lists the source clusters for the given sequence. Duplicate sequences can still occur when two different genes lead to identical consensus sequences. (These duplicated sequences are merged by the germline filters.)

Below, we use the term cluster set to refer to all the sequences that are in any of the listed clusters.

Some clusters lead to ambiguous consensus sequences (those that include N bases). These have already been filtered out.

The name of the candidate gene. See novel gene names.
The original database gene to which the sequences from this row were originally assigned. All candidates coming from the same source gene are grouped together.
Chain type: VH for heavy, VK for light chain lambda, VL for light chain kappa

From which type of cluster or clusters the consensus was computed. If there are multiple clusters that give rise to the same consensus sequence, they are all listed here, separated by semicolon. A cluster name such as 2-4 is for a percentage difference window: Such a cluster consists of all sequences assigned to the source gene that have a percentage difference to it between 2 and 4 percent.

A cluster name such as cl3 describes a cluster generated through linkage cluster analysis. The clusters are simply named cl1, cl2, cl3 etc. If any cluster number seems to be missing (such as when cl1 and cl3 occur, but not cl2), then this means that the cluster led to an ambiguous consensus sequence that has been filtered out. Since the cl clusters are created from a random subsample of the data (in order to keep computation time down), they are never larger than the size of the subsample (currently 1000).

The name db represents a cluster that is identical to the database sequence. If no actual cluster corresponding to the database sequence is found, but the database sequence is expressed, a db cluster is inserted artificially in order to make sure that the sequence is not lost. The cluster name all represents the set of all sequences assigned to the source gene. This means that an unambiguous consensus could be computed from all the sequences. Typically, this happens during later iterations when there are no more novel sequences among the sequences assigned to the database gene.

The number of sequences from which the consensus was computed. Equivalently, the size of the cluster set (all clusters described in this row). Sequences that are in multiple clusters at the same time are counted only once.

The number of unique J genes associated with the sequences in the cluster set.

Consensus sequences are computed only from V gene sequences, but each V gene sequence is part of a full V/D/J sequence. We therefore know for each V sequence which J gene it was found with. This number says how many different J genes were found for all sequences that the consensus in this row was computed from.

The number of unique CDR3 sequences associated with the sequences in the cluster set. See also the description for the Js column. This number says how many different CDR3 sequences were found for all sequences that the consensus in this row was computed from.

The number of exact occurrences of the consensus sequence among all sequences assigned to the source gene.

To clarify, we describe how the set of exact sequences is found: First, all sequences assigned to a source gene are clustered. A consensus is then computed from each cluster. Then we look back at all sequences assigned to the source gene and find exact occurrences of that consensus sequence.

How many unique barcode sequences were used by the sequences in the set of exact sequences (described above).
How many unique D genes were used by the sequences in the set of exact sequences (described above). Only those D gene assignments are included in this count for which the number of errors is zero, the E-value is at most a given threshold, and for which the number of covered bases is at least a given percentage.
How many unique J genes were used by the sequences in the set of exact sequences (described above).
How many unique CDR3 sequences were used by the sequences in the set of exact sequences (described above).
The number of differences between the consensus sequence and the sequence of the source gene. (Given as edit distance, that is insertion, deletion, mismatch count as one difference each.)
Indicates whether the consensus sequence contains a stop codon.
Whether the consensus sequence “looks like” a true V gene (1 if yes, 0 if no). Currently, this checks whether the 5’ end of the sequence matches a known V gene motif.
Where the CDR3 starts within the discovered V gene sequence. This uses the most common CDR3 start location among the sequences from which this consensus is derived.
The consensus sequence itself.

The igdiscover discover command can also be run by hand with other parameters, in which case additional columns may appear.

Number of N bases in the consensus
Number of approximate occurrences of the consensus sequence among all sequences assigned to the source gene. See the description for the exact column. This approximate set is similar to the exact set, except that a difference up to a given percentage is allowed when comparing the consensus sequence to the other sequences.
Same as Js_exact, except that it refers to the approximate occurrences of the consensus sequence.
Same as CDR3s_exact, except that it refers to the approximate occurrences of the consensus sequence.

Novel V gene names

Each V gene discovered by IgDiscover gets a unique name such as “VH4.11_S1234”. The “VH4.11” is the name of the database gene to which the novel V gene was initially assigned. The number 1234 is derived from the nucleotide sequence of the novel gene. That is, if you discover the same sequence in two different runs of the IgDiscover, or just in different iterations, the number will be the same. This may help when manually inspecting results.

Be aware that you still need to check the sequence itself since even different sequences can sometimes lead to the same number (a “hash collision”).

The _S1234 suffixes do not accumulate. Before IgDiscover adds the suffix in an iteration, it removes the suffix if it already exists.


The igdiscover program has multiple subcommands. You should already be familiar with the two commands init and run. Each subcommand comes with its own help page that shows how to use that subcommand. Run the command with the --help option to see the help. For example,

igdiscover run --help

shows the help for the run subcommand.

The following additional subcommands may be useful for further analysis.

Find common V genes between two different antibody libraries
Cluster upstream sequences (UTR and leader) for each gene
Draw a dendrogram of sequences in a FASTA file.
Rename sequences in a target FASTA file using a template FASTA file
Compute union of sequences in multiple FASTA files

The following subcommands are used internally, and listed here for completeness.

Filter a table with IgBLAST results
Count and plot V, D, J gene usage
Group sequences by barcode and V/J assignment and print each group’s consensus (unused in IgDiscover)
Create new V gene database from V gene candidates using the germline and pre-germline filter criteria.
Discover candidate new V genes within a single antibody library
For each V gene, plot a clustermap of the sequences assigned to it
Plot histograms of differences to reference V gene

Germline and pre-germline filtering

V gene sequences found by the clustering step of the program (the discover subcommand) are stored in the file. The entries are “candidates” because many of these will be PCR or other artifacts and therefore do not represent true novel V genes. The germline and pre-germline filters take care of removing artifacts. They germline filter is the “real” filter and used only in the last iteration in order to obtain the final gene database. The pre-germline filter is less strict and used in all the earlier iterations.

The germline filters are implemented in the igdiscover germlinefilter subcommand. It performs the following filtering and processing steps:

  • Discard sequences with N bases
  • Discard sequences that come from a consensus over too few source sequences
  • Discard sequences with too few unique CDR3s (CDR3s_exact column)
  • Discard sequences with too few unique Js (Js_exact column)
  • Discard sequences identical to one of the database sequences (if DB given)
  • Discard sequences that do not match a set of known good motifs
  • Discard sequences that contain a stop codon (has_stop column)
  • Discard near-duplicate sequences
  • Discard cross-mapping artifacts
  • Discard sequences whose “allele ratio” is too low.

If a whitelist of sequences is provided (by default, this is the input V gene database), then the candidates that appear on it

  • are not checked for the cluster size criterion,
  • do not need to match a set of known good motifs,
  • are never considered near-duplicates (but they are checked for cross-mapping and for the allele ratio),
  • are allowed to contain a stop codon.

Whitelisting allows IgDiscover to identify known germline sequences that are expressed at low levels in a library. If enabled with whitelist: true (the default) in the pregermline and germline filter sections of the configuration file, the sequences present in the starting database are treated as validated germline sequences and will not be discarded if due to too small cluster size as long as they fulfill the remaining criteria (unique_cdr3s, unique_js etc.).

Cross-mapping artifacts

If two very similar sequences appear in the database used by IgBLAST, then sequencing errors may lead to one sequence incorrectly being assigned to the other. This is particularly problematic if one of the sequences is highly expressed while the other is not expressed at all. The not expressed sequence is even included in the list of V gene candidates because it is in the input database and therefore whitelisted. We call this a “cross-mapping artifact”.

The germline filtering step of IgDiscover therefore aims to eliminate cross-mapping artifacts by checking all pairs of sequences for the following:

  • The two sequences have a distance of 1,
  • they are both in the database for that particular iteration (only then can cross-mapping occur)
  • the ratio between the expression levels of the two sequences (using the cluster_size field in the file) is less than the value cross_mapping_ratio defined in the configuration file (0.02 by default).

If all that is the case, then the sequence with the lower expression is discarded.

Allele-ratio filtering

When multiple alleles of the same gene appear in the list of V gene candidates, such as IGHV1-2*02 and IGHV1-2*04, the germline filter computes the ratio of CDR3s_exact between them. If the ratio is under a threshold, the lower-expressed candidate is discarded. The default threshold is 0.1 and can be modified in the configuration file by adjusting the allele_ratio settings within the germline filter sections.

New in version 0.7.0.

Data from the Sequence Read Archive (SRA)

To work with datasets from the Sequence Read Archive, you may want to use the tool fastq-dump, which can download the reads in the format required by IgDiscover. You just need to know the accession number, such as “SRR2905710” and then run this command to download the files to the current directory:

fastq-dump --split-files --gzip SRR2905710

The --split-files option ensures that the paired-end reads are stored in two separate files, one for the forward and one for the reverse read, respectively. (If you do not provide it, you will get an interleaved FASTQ file that currently cannot be read by IgDiscover). The --gzip option creates compressed output. The command creates two files in the current directory. In the above example, they would be named SRR2905710_1.fastq.gz and SRR2905710_2.fastq.gz.

The program fastq-dump is part of the SRA toolkit. On Debian-derived Linux distributions, you can typically install it with sudo apt-get install sra-toolkit. On Conda, install it with conda install -c bioconda sra-tools.

Does random subsampling influence results?

Random subsampling indeed influences somewhat which sequences are found by the cluster analysis, particularly in the beginning. However, the probability is large that all highly expressed sequences are represented in the random sample. Also, due to the database growing with subsequent iterations, the set of sequences assigned to a single database gene becomes smaller and more homogeneous. This makes it increasingly likely that also sequences expressed at lower levels result in a cluster since they now make up a larger fraction of each subsample.

Also, many of the clusters which are captured in one subsample but not in the other are artifacts that are then filtered out anyway by the pre-germline or germline filter.

On human data with a nearly complete starting database, the subsampling seems to have no influence at all, as we determined experimentally. We repeated a run of the program four times on the same human dataset, using identical parameters each time except that the subsampling was done in a different way. Although intermediate results differed, all four personalized databases that the program produced were exactly identical.

Concordance is lower, though, when the input database is not as complete as the human one.

The way in which random subsampling is done is modified by the seed configuration setting, which is set to 1 by default. If its value is the same for two different runs of the program with otherwise identical settings, the numbers chosen by the random number generator will be the same and therefore also subsampling will be done in an identical way. This makes runs of the program reproducible. In order to test how results differ when subsampling is done in a different way, change the seed to a different value.

Logging the program’s output to a file

When you report a bug or unusual behavior to us, we might ask you to send us the output of igdiscover run. You can send its output to a file by running the program like this:

igdiscover run >& logfile.txt

And here is how to send the logging output to a file and also see the output in your terminal at the same time (but you lose the colors):

igdiscover run |& tee logfile.txt


Analysis directory
The directory that was created with igdiscover init. Separate ones are created for each experiment. When you used igdiscover init myexperiment, the analysis directory would be myexperiment/.
Starting database
The initial list of V/D/J genes. These are expected to be in FASTA format and are copied into the database/ directory within each analysis directory.