Duplex Consensus Calling

Overview

Duplex consensus calling takes reads generated from both strands of a double-stranded source molecule and produces consensus reads with extremely low error rates. This is the process used in duplex sequencing methods such as those described by Kennedy et al, where UMIs are attached to each end of the source molecule.

The mathematical model is similar to single-strand consensus calling, but the mechanics differ because reads from both strands must be combined.

Duplex consensus calling is run after grouping reads with fgumi group --strategy paired.

Process

Starting from a group of reads identified as originating from the same double-stranded molecule, the two strands are labeled A and B. The process proceeds through these steps:

1. Split reads into four sub-groups: A1 (strand A, read 1), A2, B1, B2
2. Unmap and revert to sequencing order
3. Quality trim (optional, recommended)
4. Mask remaining low-quality bases to N
5. Trim to insert length to avoid reading into adapters
6. Filter by CIGAR to ensure reads are in phase
7. Call four single-strand consensus reads (one each for A1, A2, B1, B2)
8. Call two duplex consensus reads by combining A1+B2 and A2+B1

Splitting Reads into Groups

Reads are split by strand of origin (A or B) and whether they are sequencing read 1 or 2. R1s from strand A correspond to R2s from strand B, and vice versa.

Quality Trimming

Reads can be end-trimmed to remove low-quality bases. This is highly recommended as it reduces disagreements in the consensus and fewer no-calls (Ns). Trimming uses the same running-sum algorithm as BWA.

Masking Low-Quality Bases

Bases below the minimum quality threshold are converted to Ns so they are not used in consensus calling. If quality trimming is disabled, reads are truncated to remove contiguous trailing Ns.

Trimming to Insert Length

Reads longer than the insert length read into adapter sequence. For duplex data, A1 and B2 reads may read into different adapter sequences. Calling consensus across different adapters produces many disagreements and no-calls, potentially causing consensus reads to be erroneously filtered. Reads are therefore trimmed to insert length before consensus calling.

CIGAR Filtering

Without multiple alignment, length errors (indels) in raw reads cause reads to be out of phase with each other. For example:

1: ACGTGACTGACTAGCTTTTTTT-AGACTAGCTACTACT
2: ACGTGACTGACTAGCTTTTTTT-AGACTAGCTACTACT
3: ACGTGACTGACTAGCTTTTTTTT-GACTAGCTACTACT

Read 3 has an extra T, causing many disagreements with reads 1 and 2.

To handle this, reads are grouped by compatible CIGAR alignments, and only the largest group is used for consensus. This is performed independently on A1+B2 and B1+A2 reads.

Calling Single-Strand Consensus Reads

Four single-strand consensus reads are generated (A1, A2, B1, B2) using the standard consensus calling model.

Calling Duplex Consensus Reads

The final duplex R1 and R2 are produced by merging the appropriate A and B reads base-by-base:

1. Bases agree: quality = Q(A) + Q(B)
2. Bases disagree, different qualities: base = higher quality base, quality = Q(higher) - Q(lower)
3. Bases disagree, same quality: base is arbitrarily from A, quality = 2 (minimum Phred score)

The min-reads Parameter

For Simplex Consensus

fgumi simplex and fgumi filter accept a single --min-reads value.

For Duplex Consensus

fgumi duplex and fgumi filter accept one, two, or three --min-reads values. If fewer than three values are supplied, the last is repeated (e.g. 80 40 becomes 80 40 40, 10 becomes 10 10 10).

The values control:

1. First value: minimum total raw reads across both single-strand consensuses for the final duplex read
2. Second value: minimum reads for the single-strand consensus with more support
3. Third value: minimum reads for the single-strand consensus with less support

If values two and three differ, the more stringent value must come first.

Example: --min-reads 7 3 1 requires:

• At least 7 total raw reads supporting the duplex consensus
• At least 3 raw reads for the better-supported single-strand consensus
• At least 1 raw read for the other single-strand consensus