Each tool has a set of Common options for input/output, profiling and debugging.

dedup - Deduplicate reads using UMI and mapping coordinates

Deduplicate reads based on the mapping co-ordinate and the UMI attached to the read

The identification of duplicate reads is performed in an error-aware manner by building networks of related UMIs (see --method). dedup can also handle cell barcoded input (see --per-cell).


umi_tools dedup --stdin=INFILE --log=LOGFILE [OPTIONS] > OUTFILE

Selecting the representative read

For every group of duplicate reads, a single representative read is retained.The following criteria are applied to select the read that will be retained from a group of duplicated reads:

1. The read with the lowest number of mapping coordinates (see --multimapping-detection-method option)

2. The read with the highest mapping quality. Note that this is not the read sequencing quality and that if two reads have the same mapping quality then one will be picked at random regardless of the read quality.

Otherwise a read is chosen at random.

Dedup-specific options


One can use the edit distance between UMIs at the same position as an quality control for the deduplication process by comparing with a null expectation of random sampling. For the random sampling, the observed frequency of UMIs is used to more reasonably model the null expectation.

Use this option to generate a stats outfile called:

Reports the (binned) average edit distance between the UMIs at each position. Positions with a single UMI are reported seperately. The edit distances are reported pre- and post-deduplication alongside the null expectation from random sampling of UMIs from the UMIs observed across all positions. Note that separate null distributions are reported since the null depends on the observed frequency of each UMI which is different pre- and post-deduplication. The post-duplication values should be closer to their respective null than the pre-deduplication vs null comparison

In addition, this option will trigger reporting of further summary statistics for the UMIs which may be informative for selecting the optimal deduplication method or debugging.

Each unique UMI sequence may be observed [0-many] times at multiple positions in the BAM. The following files report the distribution for the frequencies of each UMI.


The _stats_per_umi_per_position.tsv file simply tabulates the counts for unique combinations of UMI and position. E.g if prior to deduplication, we have two positions in the BAM (POSa, POSb), at POSa we have observed 2*UMIa, 1*UMIb and at POSb: 1*UMIc, 3*UMId, then the stats file is populated thus:

counts instances_pre
1 2
2 1
3 1

If post deduplication, UMIb is grouped with UMIa such that POSa: 3*UMIa, then the instances_post column is populated thus:

counts instances_pre instances_post
1 2 1
2 1 0
3 1 2

The _stats_per_umi_per.tsv table provides UMI-level summary statistics. Keeping in mind that each unique UMI sequence can be observed at [0-many] times across multiple positions in the BAM,

times_observed:How many positions the UMI was observed at
total_counts:The total number of times the UMI was observed across all positions
median_counts:The median for the distribution of how often the UMI was observed at each position (excluding zeros)

Hence, whenever times_observed=1, total_counts==median_counts.

Extracting barcodes

It is assumed that the FASTQ files were processed with umi_tools extract before mapping and thus the UMI is the last word of the read name. e.g:


where AATT is the UMI sequeuence.

If you have used an alternative method which does not separate the read id and UMI with a “_”, such as bcl2fastq which uses “:”, you can specify the separator with the option --umi-separator=<sep>, replacing <sep> with e.g “:”.

Alternatively, if your UMIs are encoded in a tag, you can specify this by setting the option –extract-umi-method=tag and set the tag name with the –umi-tag option. For example, if your UMIs are encoded in the ‘UM’ tag, provide the following options: --extract-umi-method=tag --umi-tag=UM

Finally, if you have used umis to extract the UMI +/- cell barcode, you can specify --extract-umi-method=umis

The start position of a read is considered to be the start of its alignment minus any soft clipped bases. A read aligned at position 500 with cigar 2S98M will be assumed to start at position 498.


How are the barcodes encoded in the read?

Options are:

  • read_id (default)
    Barcodes are contained at the end of the read separated as specified with --umi-separator option
  • tag
    Barcodes contained in a tag(s), see --umi-tag/--cell-tag options
  • umis
    Barcodes were extracted using umis (https://github.com/vals/umis)


Separator between read id and UMI. See --extract-umi-method above. Default=``_``


Tag which contains UMI. See --extract-umi-method above


Separate the UMI in tag by SPLIT and take the first element


Separate the UMI in by DELIMITER and concatenate the elements


Tag which contains cell barcode. See –extract-umi-method above


Separate the cell barcode in tag by SPLIT and take the first element


Separate the cell barcode in by DELIMITER and concatenate the elements

UMI grouping options


What method to use to identify group of reads with the same (or similar) UMI(s)?

All methods start by identifying the reads with the same mapping position.

The simplest methods, unique and percentile, group reads with the exact same UMI. The network-based methods, cluster, adjacency and directional, build networks where nodes are UMIs and edges connect UMIs with an edit distance <= threshold (usually 1). The groups of reads are then defined from the network in a method-specific manner. For all the network-based methods, each read group is equivalent to one read count for the gene.

  • unique
    Reads group share the exact same UMI
  • percentile
    Reads group share the exact same UMI. UMIs with counts < 1% of the median counts for UMIs at the same position are ignored.
  • cluster
    Identify clusters of connected UMIs (based on hamming distance threshold). Each network is a read group
  • adjacency
    Cluster UMIs as above. For each cluster, select the node (UMI) with the highest counts. Visit all nodes one edge away. If all nodes have been visited, stop. Otherwise, repeat with remaining nodes until all nodes have been visted. Each step defines a read group.
  • directional (default)
    Identify clusters of connected UMIs (based on hamming distance threshold) and umi A counts >= (2* umi B counts) - 1. Each network is a read group.


For the adjacency and cluster methods the threshold for the edit distance to connect two UMIs in the network can be increased. The default value of 1 works best unless the UMI is very long (>14bp).


Causes two reads that start in the same position on the same strand and having the same UMI to be considered unique if one is spliced and the other is not. (Uses the ‘N’ cigar operation to test for splicing).


Mappers that soft clip will sometimes do so rather than mapping a spliced read if there is only a small overhang over the exon junction. By setting this option, you can treat reads with at least this many bases soft-clipped at the 3’ end as spliced. Default=4.


If the sam/bam contains tags to identify multimapping reads, you can specify for use when selecting the best read at a given loci. Supported tags are “NH”, “X0” and “XT”. If not specified, the read with the highest mapping quality will be selected.


Use the read length as a criteria when deduping, for e.g sRNA-Seq.

Single-cell RNA-Seq options


Reads will be grouped together if they have the same gene. This is useful if your library prep generates PCR duplicates with non identical alignment positions such as CEL-Seq. Note this option is hardcoded to be on with the count command. I.e counting is always performed per-gene. Must be combined with either --gene-tag or --per-contig option.


Deduplicate per gene. The gene information is encoded in the bam read tag specified


BAM tag which describes whether a read is assigned to a gene. Defaults to the same value as given for --gene-tag


Use in conjunction with the --assigned-status-tag option to skip any reads where the tag matches this regex. Default ("^[__|Unassigned]") matches anything which starts with “__” or “Unassigned”:


Deduplicate per contig (field 3 in BAM; RNAME). All reads with the same contig will be considered to have the same alignment position. This is useful if you have aligned to a reference transcriptome with one transcript per gene. If you have aligned to a transcriptome with more than one transcript per gene, you can supply a map between transcripts and gene using the --gene-transcript-map option


File mapping genes to transcripts (tab separated), e.g:

gene1   transcript1
gene1   transcript2
gene2   transcript3


Reads will only be grouped together if they have the same cell barcode. Can be combined with --per-gene.

SAM/BAM Options


Minimium mapping quality (MAPQ) for a read to be retained. Default is 0.


How should unmapped reads be handled. Options are:
  • discard (default)
    Discard all unmapped reads
  • use
    If read2 is unmapped, deduplicate using read1 only. Requires --paired
  • output
    Output unmapped reads/read pairs without UMI grouping/deduplication. Only available in umi_tools group


How should chimeric read pairs be handled. Options are:
  • discard
    Discard all chimeric read pairs
  • use (default)
    Deduplicate using read1 only
  • output
    Output chimeric read pairs without UMI grouping/deduplication. Only available in umi_tools group


How should unpaired reads be handled. Options are:
  • discard
    Discard all unpaired reads
  • use (default)
    Deduplicate using read1 only
  • output
    Output unpaired reads without UMI grouping/deduplication. Only available in umi_tools group


Ignore the UMI and group reads using mapping coordinates only


Only consider a fraction of the reads, chosen at random. This is useful for doing saturation analyses.


Only consider a single chromosome. This is useful for debugging/testing purposes

Input/Output Options

--in-sam, --out-sam

By default, inputs are assumed to be in BAM format and outputs are written in BAM format. Use these options to specify the use of SAM format for input or output.


BAM is paired end - output both read pairs. This will also force the use of the template length to determine reads with the same mapping coordinates.

Group/Dedup options


By default, output is sorted. This involves the use of a temporary unsorted file since reads are considered in the order of their start position which may not be the same as their alignment coordinate due to soft-clipping and reverse alignments. The temp file will be saved (in --temp-dir) and deleted when it has been sorted to the outfile. Use this option to turn off sorting.


forces dedup to parse an entire contig before yielding any reads for deduplication. This is the only way to absolutely guarantee that all reads with the same start position are grouped together for deduplication since dedup uses the start position of the read, not the alignment coordinate on which the reads are sorted. However, by default, dedup reads for another 1000bp before outputting read groups which will avoid any reads being missed with short read sequencing (<1000bp).