RepeatModeler stuck at rmblastn with a small fungal genome at "Refining 2 families"

[ishengtsai]

Hi, Repatmodeler have been great since the very first version (been using it from 10Mb to 5Gb genomes which is fine). However, for some odd reason when annotating this fungal genome (˜42Mb), the rmblastn just got stuck at round 2. We have been trying this on:

rmblastn precompiled 2.14.1
rmblastn compiled from source 2.14.1
rmblastn compiled from source 2.14.0

sh -c /home/ijt/bin/rmblast-2.14.1/bin//rmblastn -num_alignments 9999999 -db /mnt/nas2/ijt/fungi/Ryder_fungi_repeat_test/RM_159647.SunJul211952592024/round-2/family-22-cons-2.fa -query /mnt/nas2/ijt/fungi/Ryder_fungi_repeat_test/RM_159647.SunJul211952592024/round-2/family-22.fa -gapopen 20 -gapextend 5 -mask_level 80 -complexity_adjust -word_size 7 -xdrop_ungap 300 -xdrop_gap_final 150 -xdrop_gap 75 -min_raw_gapped_score 150 -dust no -outfmt="6 score perc_sub perc_query_gap perc_db_gap qseqid qstart qend qlen sstrand sseqid sstart send slen kdiv cpg_kdiv transi transv cpg_sites qseq sseq" -num_threads 1 -mt_mode 1 -matrix comparison.matrix 2>/mnt/nas2/ijt/fungi/Ryder_fungi_repeat_test/RM_159647.SunJul211952592024/round-2/ncResults-1721599401-186349-2050.75563326076.err

I actually do not know what seems to be causing the problem so I am attaching the two fasta below. Some suggestions would be appreciated.

family-22-cons-2.fa
family-22.fa

[rmhubley]

This is a 30kb TA-rich region. There are some internal tandem repetitions but no overall discernible pattern. Owing to the size (and the several copies present) this may be a satellite region. It's certainly not a TE sequence. But your original question is why does it take so long to align. This is a 30kb sequence of mostly two bases being searched against a handful of 30kb sequences with similar composition. At word-size of 7 this really blows up the search space. In essence almost every position can align to almost any other position.

This is the first time I have seen (or someone has reported) that RECON produced such a long low-complexity region as a family. I suppose we could add a low-complexity filter to the families returned by RECON (as is done for RepeatScout) -- I'll look into this for a future release. Another possibility would be to apply a low-complexity filter to seeding words - as is done in Phil Green's crossmatch. Although that would be a larger undertaking. In this case, rmblast will finish it's work, albeit after much much longer processing time than the rest of the families. On my machine running in a single thread this search took 1 hr 22 minutes.

BTW...thanks for tracking this down to those files and attaching them. That really helped in figuring this one out.

[ishengtsai]

Thanks. Look forward to the future release! For this case I would remove the contig / region in question and rerun RepeatModeler?

[rmhubley]

That would ensure that it didn't happen again. Although, as I said, it should finish given more time.