TOUCAN: Supervised learning for fungal BGC discovery

A supervised learning framework to predict Biosynthetic Gene Clusters (BGCs) in fungi based on a combination of feature types (k-mers, Pfam protein domains, and GO terms).

How to: Classification

Make a copy of /src/config.init.DEFAULT, and rename it to /src/config.init. Update the [default] home to the current project root path.

Configure

At the [prediction] section in the config.init file, specify the minimum parameters accordingly:

• the task: train, validation, or test
• indicate the corpus location in source.path
• (if using sequences) indicate the source.type: nucleotide or aminoacid
• specify the positive instances % in pos.perc
• indicate the feat.type as kmers, domains or go (if combining multiple features, separate them with a -, as in go-kmers-domains)
• set the minimum occurrences to consider a feature in feat.minOcc
• set the k-mer length in feat.size
• select a classifier: logit, mlp, linearsvc, nusvc, svc, randomforest

Run

To run the classification task from the project virtualenv simply:

(.env) user@foo:~fungalbgcs/src$ python -m pipeprediction.ML

Output

The train task will generate a /metrics folder, with:

• the (re-load-able) model file (classifier)_(featuretype).model.pkl
• a list of features file (featuretype).feat

The validation task will also generate in the /metrics folder:

• a performance file (classifier)_(featuretype).valid with P, R, F-m and a confusion matrix
• a list of {valid_instance_IDs, predicted label} file (classifier)_(featuretype).IDs.valid

The test task requires either train or validation to have been performed, since it will read from the model *.model.pkl and feature *.feat files. It generates in the /metrics folder:

• a performance file (classifier)_(featuretype).test with P, R, F-m and a confusion matrix
• a list of {test_instance_IDs, predicted label} file (classifier)_(featuretype)_(testfolder).IDs.test, used as input for evaluation against gold clusters

Resources

Datasets: Openly available fungal BGC datasets to train and validate models (details here).

External software: To set up Pfam for protein domain annotation locally, please refer to the steps on /extSoftware/.