This is a work in progress. The idea is simple: develop a machine-learning pipeline based on the principle of stacking that can work with count data as generated via high-throughput platforms in various areas of biological research.
Using a three-model approach (based on Poisson, Negative Binomial and Zero-inflated negative binomial regression models) to find significant OTUs between two treatment groups is one way of analysing the dataset. In this approach, the treatment groups are explanatory variables and OTU counts are response variables. It is implemented in NegBinSig-Test.
Another approach is to use logistic regression modelling based on penalized regression (as implemented via Lasso and Elastic Nets). In these approaches, OTU counts become explanatory variables and treatment groups become response variable as implemented in MicrobeNets.
If one, from a biological standpoint, considers statistical models as nothing more than screening tools, one might be interested in knowing whether there is any biological convergance between various models. In such a case, it will be useful to see if we find a handful of predictors as significant, regardless of the test used. This is what this pipeline attempts to do from the output produced by the two analysis scripts described earlier.
For help regarding the input parameters to run the pipline, which is present in src/ folder, please type the following:
python bio_convergence.py -h
The pipeline accepts an OTU table file, generated via QIIME 1.8.0 (stable public release), (in tab-delimited format) as input and a standard mapping/metadata file compatible with QIIME.
The publicly available dataset of the lean-obese study, as obtained via qiita, is provided with the pipeline as an example.
The main output of the script is called convergence_trt1_trt2.txt, where trt1 and trt2 are the levels of the metadata variable used for prediction.