skyclub3/yak

Yet another k-mer analyzer

IMPORTANT: Since 3ace4ff, the format of binary k-mer dump is incompatible with the previous versions. You have to rerun yak count to generate k-mer dump in the new format.

Getting Started

# Download and compile
git clone https://github.com/lh3/yak
cd yak && make

# build k-mer hash table for assembly; count singletons
./yak count -K1.5g -t32 -o asm.yak asm.fa.gz
# build k-mer hash tables for high-coverage reads; discard singletons
./yak count -b37 -t32 -o ccs.yak ccs-reads.fq.gz
# for paired end: to provide two identical streams
./yak count -b37 -t32 -o sr.yak <(zcat sr*.fq.gz) <(zcat sr*.fq.gz)

# compute assembly or reads QV
./yak qv -t32 -p -K3.2g -l100k sr.yak asm.fa.gz > asm-sr.qv.txt
./yak qv -t32 -p sr.yak ccs-reads.fq.gz > ccs-sr.qv.txt
# compute k-mer QV for reads
./yak inspect ccs.yak sr.yak > ccs-sr.kqv.txt
# evaluate the completeness of assembly
./yak inspect sr.yak asm.yak > sr-asm.kqv.txt

# print k-mer histogram
./yak inspect sr.yak > sr.hist

# partition chrX/Y in human de novo assembly
wget -O- 'https://zenodo.org/record/7882299/files/human-chrXY-yak.tar?download=1' | tar tf -
./yak sexchr -K2g -t16 chrY-no-par.yak chrX-no-par.yak par.yak hap1.fa hap2.fa > cnt.txt
./groupxy.pl cnt.txt|awk '$4==1'|cut -f2|seqtk subseq -l80 <(cat hap1.fa hap2.fa) - > new-hap1.fa
./groupxy.pl cnt.txt|awk '$4==2'|cut -f2|seqtk subseq -l80 <(cat hap1.fa hap2.fa) - > new-hap2.fa

Introduction

Yak is initially developed for two specific use cases: 1) to robustly estimate the base accuracy of CCS reads and assembly contigs, and 2) to investigate the systematic error rate of CCS reads. It achieves the goals by comparing sequences to the k-mer spectrum of short reads or by comparing spectra. No reference genome or truth data is needed.

It is worth noting that estimating base accuracy is tricky. When the accuracy approaches Q50, both unsampled and erroneous k-mers in short reads may interfere with a naive estimator. Yak introduces an empirical model to address this issue. Its estimate is less affected by the coverage and the quality of short reads.