Major questions:

1. How specific & sensitive is the our current HR-SIP analytical method?

2. What are some major factors affecting accuracy of our current HR-SIP analytical method?

3. How can we increase the accuracy of HR-SIP?

Figure run down

• Figure1
  • taxon abundance example
    • control vs treatment facets (100% atom 13C; 10% incorporators)
    • 3 plots: absolute counts, absolute subsampled counts, relative abundance subsampled counts
• Figure2
  • Bray-curtis nmds plots
    • 100% atom 13C
    • % incorporators: 0, 10, 25, 50
    • showing variance between replicates (elipses?, error bars?, procrustes?)
      • procrustes: dendrogram of plots (reps & treatments)
        • all replicates should be very similar vs treatments
• Figure3
  • accuracy ~ atom % 13C + % incorporators
  • boxplots
• Figure4
  • accuracy ~ atom % 13C + evenness
• Figure5
  • BD-window
• Figure6
  • comparing methods
    • HR-SIP
    • qSIP
    • old-SIP
  • x-axis: atom % 13C
  • y-axis: values
  • color/fill: method
  • different plots: % incorporators

• Figure S1:
  • simulation scheme diagram
• Figure SX:
  • validation figs
• Figure S2:
  • accuracy ~ abundance correlation (figure2)
• Figure S3:
  • accuracy ~ abundance correlation (figure3)

Simulation methodology validation

Diagram of simulation methodology

• Figure 1

Supplemental example plots of key aspects in simulation

Genomic fragment simulation

• Sup Figure:
  • 3 taxa of differing GC
  • Fragment length distributions
    • 'realistic' = skewed normal
    • 'large' = ~50 kb
    • 'small' = ~1-2 kb
  • faceted heatmap:
    • taxa ~ (fragment length distribution)

Diffusion calculations

• Sup Figure:
  • 3 taxa of differing GC (same as last)
  • histograms:
    • BD distribution for pre & post diffusion
  • Shows effect of diffusion & how it relates to fragment length

Real HR-SIP dataset validation

• Dataset: Ashley's priming experiment
• Select OTUs with close genome representative
  • BLASTn of OTU 16S vs genome 16S blast db
• Simulation of just genomes of interest
• Simulation of all genomes
• Plot abundance distributions: real vs simulated

Mock community validation?

• Goal: validate the simulation model vs a ground-truth real dataset
• Is this necessary???
• 10-20 isolates
• 'Control' community:
  • All DNA is 12C
• 'Treatment' community:
  • 2 to 4 isolates grown with 13C (~100% 13C incorporation)
• 2 gradients (full HR-SIP pipeline)
  • add to MiSeq run
• Marquessa project???

• Downsides:

  • Costly in time and money
  • Would takes months to complete

• (Sup) Figure:

  • Taxon abundances of real & simulated communities
• Table:
  • Accuracy of detecting incorporators (simulation vs real dataset)

Example figure: absolute & relative taxon abundances

• All taxa simulated (n=1210); just top 100 plotted

• Figure:

  • 1 community
  • faceted:
    • (absolute abundance or relative abundance) ~ (total community or subsampled community)

1) How specific & sensitive is the our current HR-SIP analytical method?

2) What some major factors affecting accuracy of identifying incorporators?

Accuracy as a function of isotope incorporation

Variable parameters:

• % isotope incorporation
  • 0, 25, 50, 100
• % taxa that incorporate
  • 1, 5, 10, 25, 50
• n-reps (stocastic: taxon abundances & which incorporate)
  • 10 (20?)
• total simulations
  • 4 * 3 * 10 = 120
  • time = 120 * 1.6 [hr] = 192 [hr] = 8 [days]

Set parameters:

• Community rank-abundance:
  • lognormal
• Total community abundance:
  • 1e9
• 'heavy' BD range:
  • 1.71 - 1.75

Analysis:

• Figure:
  • faceted boxplot
    • x-axis = percent isotope incorporation
    • color/group = percent of taxa to incorporate
    • facet = sensitivity/specificity (and balanced accuracy?)

Possible extensions:

• % isotope incorporation
  • normal distribution, varying standard deviation
    • mean = 50
    • sd = (1, 10, 25)
  • uniform distribution

Accuracy as a function of community evenness and isotope incorporation

Variable parameters:

• Community rank-abundance:
  • lognormal abundance distribution
    • very uneven
    • moderately uneven
    • slightly uneven
• % incorp
  • 0:100:5
• XCommunity richness:
  • X100 taxa
  • X300 taxa
  • X900 taxa
• n-reps (stocastic: taxon abundances & which incorporate)
  • 20
• total simulations

Set parameters:

• % taxa that incorporate
  • 10%?
• Total community abundance:
  • 1e9
• 'heavy' BD range:
  • 1.71 - 1.75

Analysis:

• Figure:
  • pointrange (percent_incorp ~ accuracy_value; color by community evenness)

Accuracy as a function taxon abundance

Variable parameters:

• Partition communities into 'dominant' and 'rare' subsets
  • dominant/rare cutoff: 1% relative abundance?
• Assess sensitivity & specificity for each subset
• Note:
  • Much less time to perform if using data from past simulation run datasets

Analysis:

• Sup Figures:
  • Same as above simulation runs, but split into 'dominant' vs 'rare'

Accuracy as a function of BD window

Variable parameters:

• Note:
  • Much less time to perform if using data from past simulation run datasets
• 'Heavy' BD range:
  • BD-min
    • range = [1.67:1.77:0.02]
  • BD-max
    • BD-min + 0.04
  • product: BD-min ~ BD-max

Analysis:

• Figure:
  • Plot showing sensitivities/specificities for each 'heavy' BD range
• Figure:
  • Taxon abundances of true positives & false negatives
  • Why: displays the issue with selecting one 'heavy' BD window

How can we increase the accuracy of our method?

Multiple 'heavy' fraction windows

• General sliding window approach

• Binning OTUs by log2fc ~ BD

• Extrapolating GC for each OTU, with 2-3 different 'heavy' BD windows

Time series analysis

• MetagenomeSeq method