List of things working on for current paper tables and figures. Original Post: 2014-11-21

Github respository + R markdown file + Paper draft

Github repository for this paper is found here. In particular, the R markdown file named “abxD01_analysis.Rmd” (or the pdf version) contains info on all figures and how they were generated.

Figures

Figure 1: Barchart, Differences in community structures following array of antibiotic treatments

figure PDFs here

  • change the “0” to be min of 0.001 instead of 0.0001 relabund
    • completed 11/20/14, just make sure future files are changed for other graphs using the “abxD01.barcharts.xOTU.test.r”
  • add the 1/2 dashed lines
    • completed 11/20/14
  • log trans anova to compare the CFUs across abx treatments to see if there are signif. differences
    • 1/21/15, without the untreated group included it is still significantly different, likely because of the ciprofloxacin
    • 1/21/15, removing the ciprofloxacin group and then doing the anova again showed no significant differences in CFU level among the remaining abx treatments
  • update the graph on github to reflect new min
    • 1/22/15
  • currently shows genera phylotype tx1, do family too tx2
    • 1/22/15, both up on github
  • color shade the bars by phylum
    • 2/15/15

Figure 2: Stripchart, Correlation analysis of bacterial species present on Day 0 with C. difficile levels on Day 1

figure PDF here

  • graph is first filtered by 16min tot requirement, so at least one sample must have 16min sequences of that otu
  • then based on the avg abundance for each otu, if that otu wasn’t greater than 0.001 relabund of the community in at least one of orig/cef/vanc/strep tit, then was eliminated
  • make for classification level family
    • 1/25/15, fixed graph by family, showed break down by phylum

Figure 3: Barchart, Differences in community structures between abx-titration treated mice

figure PDF here

  • add possibly a right y axes with cdiff CFU, or make new trio graph alongside

  • I want to change the graph so that the OTUs are their own graphs and the titrations are side by side
    • 11/20/14, completed most of the graphical parameters.
  • once code is working then put up the other titration data files
    • 1/17/15, added to github
  • change min relabund from 0.0001 to 0.001 for cef and strep
    • 11/20/14, changed the vanc titrations min
    • 1/17/15, changed the cef and strep titrations min
  • eliminate the untreated groups
    • 1/25/15
  • Possibly work on getting plots on top of grid lines
    • 1/25/15
  • do the stats to compare the mid/low to the highest dose
    • need to know how to graph
    • 1/27/15, used letter naming scheme
  • determine the tx’s with statistical/interesting differences for each titabx, aggregate, use same for 3 paneled graph
    • make forlogscale.csv that combines all three cef/strep/vanc data
      • edit barchart function so that it takes the combo file and makes it paneled by the 3 abx groups
    • make one common ids file
    • after use the combo file, make sure that there are sig diff in the add-on OTUs
      • 2/22/15

Figure 4: Heatmap, Correlation analysis results compared across abx experiments

figure PDF here

  • for all the significant values for each pair of original + abx rows that are both significant, quantify that significance for each column (abx treatment)… see personal notes for more
  • use classification names, possibly the genus if have it

  • figure out what want to use the side panel for? possibly highlight otus included in the model

  • Add “ns” to the graph where not significant
    • 12/8/14
  • Is this the order of drugs I want? -keep consistent
    • 12/8/14
  • change xlabel names
    • 12/8/14
  • change code to get the key back
    • 12/8/14

Figure 5: Barchart, Differences in the community with extra microbiome recovery time following abx treatment

figure PDF here

  • put metro/amp file together with all tx2 otus
  • pick subset of otus that important
  • run with divide=TRUE
  • possibly look at diffs between strong individual otus or genera
  • Do for ampicillin too

  • calculate differences in thetayc between individual days w/recovery

  • random forest regression to do family level feature selection, use these to inform selected tx2’s
    • 1/26/15
  • family graph: get rid of OTU12, 14, 9
    • 1/26/15
  • lactos decrease with recovery… are these the streptococcus? No
  • use the barcharts by graph/sort by phylum side by side like did for titration data… show family differences..
    • 1/26/15
  • stats for difference, wilcox for difference
    • 1/28/15

Figure 6:

  • See how few otus i could use before the toptitdel drops off

  • rewrite up modeling results

  • RF model built on delay data, tested on other data sets

  • RF model on 20OTUs based on toptitdel: Tried 0.5% (82 OTUs), 1% (44OTUs), and no cutoff (299 OTUs). RF models based on 1% and toptitdel combined data sets poorly predict delay data set alone (r2=0.4). 0.5% however was able to predict in upper r2=0.9s. Took the top 20 OTUs from feature selection in a random forest model based on toptitdel and still predicted delay data with r2=0.92, which is WAY better than the 44 OTUs (1% cutoff).

  • RF models using 0.5% cutoff compared to 1% and 0 cutoff

  • The criteria for picking the otus, these results can be listed in supplementary tables possibly.
    • 12/1/14, made file with info for table, now make pretty
  • Consider rerunning the model building code using a bigger candidate list
    • 12/1/14 with the BIC included in new calcs
  • Calculate the BIC for each–change code for this
    • 12/1/14
    • 12/2/14, found a faster easier way to determine variable selection using the “leaps” package in r for multiple linear regression analysis
  • using the best models from the leaps analysis predict the outcome for the newtitration data and the predicted r^2
    • make file with newtitration data that contains the OTUs in the candidate list and run, got 0.67 for the 5 parameter model
      • 12/4/14
    • now try the best model from both methods of selection…12/4/14, think im just going to go with the exhaustive method for paper
    • then model with just the top 3 otus… seems the best trade off point between BIC and adjR2… 12/4/14, gives 0.67 too, but the anova between the 5parameter model and 3 parameter model is significantly different, meaning that the 5 parameter model does improve prediction overall -use random forest for delay data alone to see what features stand out
    • how do these OTUs differ from the model built based on the original data set?
      • 12/17/14 OTU11 has most influence, followed by 1, 5, 23, 21, 3, 6, 39 before dropping off in %IncMSE. I’m curious to see what inclusion of OTU11 (Ecoli) will have on the effect of the model.
  • Make figure! for modeling results, actual vs predicted, line? with error lines?
    • 1/29/15
  • try eliminating 283 from 5 model because is low in abundance, then see how performs
    • 1/29/15
  • correlation calculations for the delay data
    • done early february
  • force OTU11 in the model in addition to the others to test the delay data and see if makes a diff, also look up the results for the 5model+OTU11
    • done early february
  • do leaps with all the filtered otus, 1/28/15
    • also rerunning with 50 OTUs, which were picked from after filtering again the abundance, doing the rf, recalculate correlations
  • I have been semi frustrated with the process. There’s no one absolute way of building models and picking the right one. Trying out two different sets of guidelines – using a curated or an uncurated pool of OTUs from which to build the best models. I may use the results of this to next use with either 2/3-1/3 train-test sets + validation, or without this step? See http://math.furman.edu/~dcs/courses/math47/R/library/DAAG/html/cv.lm.html
  • 2/8/15 - waiting on results for the uncurated pool with all samples… 2/14/15 canceling, I don’t think it’s working
  • incorporate validation sets for all 3
    • 2/22/15

Supplemental Figure 1

  • make graph showing the inverse Simpson correlation plotted against C. difficile CFU

Tables

Tables are saved within the paper word document uploaded to github (see link at top).

Table 1: Description of antibiotics in first set of experiments.

There are probably other targets to include for a few abx: clinda, metro

  • change the order of the abx
    • 11/21/14
  • make class/mechanism uniform throughout table
    • 11/21/14
  • add an administration route column
    • 11/21/14

Supplemental Table 1: Titration amounts for each antibiotic treatment.

  • change the order of the abx
    • 11/21/14
  • add an administration route column
    • 11/21/14

Supplemental Table 2: Model OTU Selection Criteria

  • make pretty