3. To assemble, or not to
assemble?
Goals: reconstruct phylogenetic content and predict
functional potential of ensemble.
• Should we analyze short reads directly?
OR
• Do we assemble short reads into longer contigs first,
and then analyze the contigs?
5. But! Assembly is…
• Morally frightening: don’t you mis-assemble
sequences?
• Computationally challenging: don’t you need big
computers?
• Technically tricky: don’t you need to be an expert?
6. Or… is it?
• Most assembly papers analyze novel data sets and
then have to argue that their result is ok (guilty!)
• Very few assembly benchmarks have been done.
• Even fewer (trustworthy) computational
time/memory comparisons have been done.
• And even fewer “assembly recipes” have been
written down clearly.
8. A mock community!
• ~60 genomes, all sequenced;
• Lab mixed with 10:1 ratio of most abundant to least
abundant;
• 2x101 reads, 107 mn reads total (Illumina);
• 10.5 Gbp of sequence in toto.
• The paper also compared16s primer sets & 454
shotgun metagenome data => reconstruction.
Shakya et al., 2013; pmid 23387867
9. Paper conclusions
• “Metagenomic sequencing outperformed most SSU
rRNA gene primer sets used in this study.”
• “The Illumina short reads provided a very good estimates
of taxonomic distribution above the species level, with
only a two- to threefold overestimation of the actual
number of genera and orders.”
• “For the 454 data … the use of the default parameters
severely overestimated higher level diversity (~ 20- fold
for bacterial genera and identified > 100 spurious
eukaryotes).”
Shakya et al., 2013; pmid 23387867
10. How about assembly??
• Shakya et al. did not do assembly; no standard for
analysis at the time, not experts.
• But we work on assembly!
• And we’ve been working on a tutorial/process for
doing it!
11. Adapter trim &
quality filter
Diginorm to C=10
Trim high-
coverage reads at
low-abundance
k-mers
Diginorm to C=5
Partition
graph
Split into "groups"
Reinflate groups
(optional
Assemble!!!
Map reads to
assembly
Too big to
assemble?
Small enough to assemble?
Annotate contigs
with abundances
MG-RAST, etc.
The Kalamazoo Metagenomics Protocol
Derived from approach used in Howe et al., 2014
13. Adapter trim &
quality filter
Diginorm to C=10
Trim high-
coverage reads at
low-abundance
k-mers
Diginorm to C=5
Partition
graph
Split into "groups"
Reinflate groups
(optional
Assemble!!!
Map reads to
assembly
Too big to
assemble?
Small enough to assemble?
Annotate contigs
with abundances
MG-RAST, etc.
The Kalamazoo Metagenomics Protocol => benchmarking!
Assemble with Velvet, IDBA, SPAdes
14. Benchmarking process
• Apply various filtering treatments to the data
(x3)
o Basic quality trimming and filtering
o + digital normalization
o + partitioning
• Apply different assemblers to the data for each
treatment (x3)
o IDBA
o SPAdes
o Velvet
• Measure compute time/memory req’d.
• Compare assembly results to “known” answer
with Quast.
15. Recovery, by assembler
Velvet IDBA Spades
Quality Quality Quality
Total length (>= 0 bp) 1.6E+08 2.0E+08 2.0E+08
Total length (>= 1000 bp) 1.6E+08 1.9E+08 1.9E+08
Largest contig 561,449 979,948 1,387,918
# misassembled contigs 631 1032 752
Genome fraction (%) 72.949 90.969 90.424
Duplication ratio 1.004 1.007 1.004
Conclusion: SPAdes and IDBA achieve similar results.
Dr. Sherine Awad
16. Treatments: some effect
IDBA
Quality Diginorm Partition
Total length (>= 0 bp) 2.0E+08 2.0E+08 2.0E+08
Total length (>= 1000 bp) 1.9E+08 2.0E+08 1.9E+08
Largest contig 979,948 1,469,321 551,171
# misassembled contigs 1032 916 828
Unaligned length 10,709,716 10,637,811 10,644,357
Genome fraction (%) 90.969 91.003 90.082
Duplication ratio 1.007 1.008 1.007
Conclusion: Treatments do not alter results much.
Dr. Sherine Awad
17. Computational cost
Velvet idba Spades
Time
(h:m:s)
RAM
(gb)
Time
(h:m:s)
RAM
(gb)
Time
(h:m:s)
RAM
(gb)
Quality 60:42:52 1,594 33:53:46 129 67:02:16 400
Diginorm 6:48:46 827 6:34:24 104 15:53:10 127
Partition 4:30:36 1,156 8:30:29 93 7:54:26 129
(Run on Michigan State HPC)
Dr. Sherine Awad
18. Need to understand:
• What is not being assembled and why?
o Low coverage?
o Strain variation?
o Something else?
• Effects of strain variation
• Additional contigs being assembled –
contamination? Spurious assembly?
• Performance of MEGAHIT assembler (a new assembler
that is very fast but still young).
19. Other observations
• 90% recovery is not bad; relatively few
misassemblies, too.
• This was not a highly polymorphic community BUT it
did have several closely related strains; more
generally, we see that strains do generate
chimeras, but not different species gen’ly.
• Challenging to execute even with a
tutorial/protocol :(
20. But! Assembly is…
• Morally frightening: don’t you mis-assemble
sequences? NO. (Or at least, not systematically.)
• Computationally challenging: don’t you need big
computers? YES. (But that’s changing.)
• Technically tricky: don’t you need to be an expert?
UNFORTUNATELY STILL YES BUT THERE’S HOPE.
21. Benchmarking &
protocols
• Our work is completely reproducible and open.
• You can re-run our benchmarks yourself if you want!
• We will be adding new assemblers in as time
permits.
• Protocol is open, versioned, citable… but also still a
work in progress :)
23. Primer bias against
Verrucomicrobia
Check taxonomy of reads causing
mismatch (A)
Verrucomicrobia cause
70% (117/168) of
mismatch
Current primers are not effective at amplifying
Verrucomicrobia
Jaron Guo
24. Thanks!
Please contact me at ctbrown@ucdavis.edu!
Everything I talked about is freely available.
Search for ‘khmer protocols’.