Effect of Kmer size on quantification

[Jayesh-Kumar-Sundaram]

My data: RNA-seq: single embryo CEL-Seq (3' bias data); 35bp Single End reads; Total reads: 361K
Annotation: I have two transcriptome assembly with no genome information.

Aligner and the alignment details

Aligner: Transcriptome-1, Transcriptome-2
Bowtie2 default: 54K, 41K
Hisat2 default: 47K, 34K
Kallisto, index -k 31: 7K, 17k (My usual default setting)
Kallisto, index -k 21: 17K, 30k
Kallisto, index -k 15: 102K, 100K
Kallisto, index -k 7: 118K, 102K
Kallisto --single-overhang, index -k 31: 40K, 30K
Kallisto --single-overhang, index -k 21: 77K, 64K
Kallisto --single-overhang, index -k 15: 154K, 128K
Kallisto --single-overhang, index -k 7: 128K, 109K

With my usual default kallisto setting, my alignment was poor. Then I realized that my data has 3' bias and is of short read length. So, I tried using different kmer length (21,15,7) for index creation to account for small read length and enabled --single-overhang to account for 3' bias. I am not sure what might a good setting to use. Any suggestions are welcome.
Note: The sample has a lot of spike-in reads. In the publication where the Transcriptome-1 assembly was used, they have reported only 16k reads aligned to Transcriptome-1, 173k reads to spike-in, 156k has no alignment (using bowtie2).