Velevt - add reads -> ruin N50. why? need advise.

[MikhailFokin]

Thank you for reading this question, in general I understand how Velvet works, but can not explain 10 fold decrease of N50 when adding more reads to the dataset.

DETAILS

MiSeq v3 ~300 bp reads, mate-pair libraries 3-12kb inserts,
(I have also ~5% of paired-end 800bp insert library used in both assemblies).

Assembly 1.
Nextclip -> A only files (Junction Adapter in both reads) -> RevCompl -> Velvet k=91

results Assembly 1
Estimated Coverage = 36.798895
Pre-graph has 623415 nodes and 21026066 sequences 53626987 kmers found
Final graph has 3170 nodes and n50 of 1948774, max 3890197, total 35875507, using 14957532/21026066 reads

Assembly 2.
Nextclip -> A, B (JA in read2), C(JA in read 1), E(JA in both with relaxed cond) -> join A,B,C,E by "cat" -> RevCompl -> Velvet k=91

results Assembly 2
Pre-graph has 1937511 nodes and 31230404 sequences 110486096 kmers found
Estimated Coverage = 44.494662
Final graph has 8571 nodes and n50 of 225231, max 909759, total 35949035, using 21595103/31230404 reads

PS I've checked that no JA left in final assemblies.
PPS My guess now - adding many bad reads complicates the graph, so playing with filtering (by Trimmomatic) now.

[Brian Bushnell]

PPS My guess now - adding many bad reads complicates the graph, so playing with filtering (by Trimmomatic) now.

Seems like the most likely reason. Though also, Velvet doesn't necessarily do well with too high coverage (not that 44x is too high; that should be fine). I encourage you to look at or post a FastQC analysis of the data.

Also, is it possible that your read pairing got messed up at some point? Check to make sure the names still match at every line.