Spec2Tax: Predicting sample organism taxonomy using machine learning on repository-scale untargeted mass spectrometry data
Classifier.prepare_datasets.ipynb prepares datasets from the spec2vec embeddings of the data from GNPS. The output of this notebook is four pairs of data (300-dimensional spec2vec embeddings of the raw mass spec data) and label files, named X and y respectively. The interclass dataset is placed in a directory called "class". The intraclass datasets are placed in directories named after the class ("Mammalia", "Gammaproteobacteria", and "Magnoliopsida").
Classifier_inter_and_intraclass_classification.ipynb assumes there are datasets at the paths from the "prepare_datasets" notebook, and trains and tests an elastic net model with a logistic regression estimator. It saves the results from the tested model in a "results" directory in the dataset's directory.
Running as a docker "definitions", "src", ".dockerignore", Dockerfile, requirements.txt, and Makefile were used to run the training and testing steps via AWS Sagemaker.
Interclass Classification This dataset samples 5000 spectra from each of these 5 taxonomic classes: Mammalia, Actinomycetia, Coscinodiscophyceae, Anthozoa, Insecta. Raw mass spec data were mapped to a 300-dimension dense embedding space with a spec2vec model. Failed conversions (amounting to less than <~2% of the total dataset) were excluded. The datasets are saved in as numpy arrays in X (spec2vec embeddings) and y (class labels converted to ints of the embeddings) pickle files.
Intraclass Mammalia Classification These dataset samples 5000 spectra from each of these 3 taxonomic families within the class Mammalia: Hominidae, Muridae, Rhinocerotidae. Data processing and filenames are as described above.
Intraclass Gammaproteobactera Classification These dataset samples 1000 spectra from each of these 2 taxonomic families within the class Gammaproteobactera: Enterobacteriaceae and Morganellaceae. Data processing and filenames are as described above.
Intraclass Magnoliopsida Classification These dataset samples 500 spectra from each of these 3 taxonomic families within the class Magnoliopsida: Malvaceae, Euphorbiaceae and Moraceae. Data processing and filenames are as described above.