kallisto quant/pseudo counts differ

[jnotwell]

I'm trying the kallisto pseudo --quant --batch command that was used in the recent isoform-level analysis preprint.

I found that it produced different counts than the kallisto quant command - is there a reason for this difference? Steps to reproduce below:

Download the sequencing reads and kallisto index:

$ wget http:https://data.nemoarchive.org/biccn/lab/zeng/transcriptome/scell/SMARTer/raw/MOp/LS-15395_S48_E1-50.fastq.tar

$ tar -xvf LS-15395_S48_E1-50.fastq.tar

$ wget https://github.com/pachterlab/kallisto-transcriptome-indices/releases/download/ensembl-96/mus_musculus.tar.gz

$ tar -zxvf mus_musculus.tar.gz

Next, quantify with kallisto quant:

$ kallisto version
kallisto, version 0.46.2

$ kallisto quant -i mus_musculus/transcriptome.idx -o LS-15395_S48_E1-50 LS-15395_S48_E1-50_R1.fastq.gz LS-15395_S48_E1-50_R2.fastq.gz 

[quant] fragment length distribution will be estimated from the data
[index] k-mer length: 31
[index] number of targets: 118,489
[index] number of k-mers: 100,614,952
[index] number of equivalence classes: 433,624
[quant] running in paired-end mode
[quant] will process pair 1: LS-15395_S48_E1-50_R1.fastq.gz
                             LS-15395_S48_E1-50_R2.fastq.gz
[quant] finding pseudoalignments for the reads ... done
[quant] processed 2,604,792 reads, 2,289,341 reads pseudoaligned
[quant] estimated average fragment length: 182.784
[   em] quantifying the abundances ... done
[   em] the Expectation-Maximization algorithm ran for 996 rounds

$ cat LS-15395_S48_E1-50/abundance.tsv | tail -n+2 | datamash sum 4
2289340.9621364

Now, perform a similar quantification with kallisto pseudo:

$ ls LS-15395_S48_E1-50_R1.fastq.gz | sed "s/_/\t/3" | awk '{print $1 "\t" $1 "_" $2 "\t" $1 "_R2.fastq.gz"}' > batch.tsv

$ kallisto pseudo -i mus_musculus/transcriptome.idx -o LS-15395_S48_E1-50_batch --quant --batch=batch.tsv  

[quant] fragment length distribution will be estimated from the data
[index] k-mer length: 31
[index] number of targets: 118,489
[index] number of k-mers: 100,614,952
[index] number of equivalence classes: 433,624
[quant] running in paired-end mode
[quant] will process pair 1: LS-15395_S48_E1-50_R1.fastq.gz
                             LS-15395_S48_E1-50_R2.fastq.gz
[quant] finding pseudoalignments for all files ... done
[quant] processed 2,604,792 reads, 2,289,341 reads pseudoaligned
[quant] Running EM algorithm for each cell .. done

We can load the total counts in python:

$ import numpy as np

$ from scipy.io import mmread

$ np.sum(mmread('LS-15395_S48_E1-50_batch/matrix.abundance.mtx'))
2196198.893168244

The counts produced by kallisto quant match the number of pseudoalignments, but the number produced by kallisto pseudo do not.

[Yenaled]

This was a while ago but this was fixed a while back in version 0.48.0 so I'm closing this issue. Also, pseudo is now deprecated in favor of "kallisto bus" (which can now do the exact same preprocessing as pseudo).