• Classification based (supervised)
  
  • Limited by database, length of reads, conflicting taxonomies
    
  • What value are names?
  • Sequencing errors not as critical
• Operational taxonomic units (OTUs; unsupervised)
  • Assign sequences to OTUs based on similarity to each other
  • Computationally demanding
  • Sensitive to sequencing errors
    
• Should not have to justify the need high quality data

• Stand-alone, platform-independent, open source, community supported (-ive) software package
• Builds upon popularity of DOTUR, which was published in 2005 (1473 citations; as of 8/13/14)
• The design is very PAUP*/R-ish

AltMetrics:

• 2952 mailing list recipients
• 313 wiki articles
• 915 forum members
• 7486 forum posts for 2252 topics

1. Data quality
2. Number of reads
3. Sequence length
4. Cost
   

• Titanium: 250 bp reads with an 0.02% error rate
• GSFLX+: 600 bp reads with an 0.02% error rate
• Others...
  • PacBio: Full length reads, but error rates of 0.5%
    
  • IonTorrent: Disaster - short and high error rates
    

• Single read sequencing (Ugh...)
• Sequence through primer and index (Illumina's protocol)
  • Waste base calls
  • Requires two PCR steps to get adapters on (chimeras?)
• Caporasso/Knight (EMP/AG/etc.)
  • Single index - large number of long primers
  • Generally being used with HiSeq (?)
• Kozich/Schloss
  • Dual index - considerably fewer primers required
  • Wet-lab SOP available

• Does a dual-index strategy work?
• Can we create a pipeline like we have for 454 that results in low error rate data?
• Is it possible to move beyond the Knight approach now that reads have gone from 150 to 250 to 300 bp to sequence longer regions?
• What are the principles of primer design so we can amplify other regions and genes?

* assumes: 15x10⁶ reads per run

• Amplified V34, V4, and V45 from mock community DNA and DNA isolated from human feces, mouse feces, and soil
  
• All samples amplified in triplicate with different barcodes (12 total samples)
  
• Have used same design/samples with other sequencing platforms (454, PacBio, IonTorrent)
  
• Sequencing runs...
  
  • Cluster density ~ 800 k/mm2
  • %PhiX ~ 5-10%
    

• DNA from 21 bacterial isolates with genome sequences:
  Acinetobacter baumannii, Actinomyces odontolyticus, Bacillus cereus, Bacteroides vulgatus, Clostridium beijerinckii, Deinococcus radiodurans, Enterococcus faecalis, Escherichia coli, Helicobacter pylori, Lactobacillus gasseri, Listeria monocytogenes, Neisseria meningitidis, Porphyromonas gingivalis, Propionibacterium acnes, Pseudomonas aeruginosa, Rhodobacter sphaeroides, Staphylococcus aureus, Staphylococcus epidermidis, Streptococcus agalactiae, Streptococcus mutans, Streptococcus pneumoniae (Available through BEI: v3.1, HM-278D)
• Since we know the true 16S rRNA gene sequences we can measure the error rates
• Can remove chimeras a priori

• Trust MiSeq software to call indices
• Form contigs
  • If there is a mismatch, the base with a qual score that is greater than or equal to a threshold wins
  • If it is less than threshold the base becomes an N and the read will ultimately get rejected
  • Contig length must make “sense”
• Remove anything with…
  • Ambiguous base calls
  • Homopolymer length greater than 8
  • Doesn’t map to correct region

• The V4 region yields the best error rates and retains the most reads
  
• Require Q to be at least 6 points higher to break a mismatch
  

"V3-V5": Based on our 454 SOP using PyroNoise

All rarefied to 5,000 reads

(XX) – number of OTUs w/ no errors

Mock-PC has no chimeras

• "Richness" is correlated with sequencing depth / crap
• Error rates have a significant impact on number of observed OTUs

"V3-V5": Based on our 454 SOP using PyroNoise

Meh.

• Note that the quality scores crap out around cycle 500
• Similar results with varying cluster density/%PhiX

• Sequence reads must fully overlap with each other to get good error remediation
• If they don't...
  • Inflated number of unique reads / OTUs
  • Samples will look more different than they really are
  • Significant computational hurdles await
  • Cheaper to throw data away and resquence correctly than to buy sufficient RAM/time
• Staggered primers and high %PhiX aren't necessary
• The V3 MiSeq kit isn't going to cut it for longer reads