Note!

This repository provides data and examples that were used for development of DeepBGC and its evaluation with ClusterFinder and antiSMASH.

See https://github.com/Merck/deepbgc for the DeepBGC tool.

Note!

DeepBGC development & evaluation code

Reproducing data

Reproduction and storage of data files is managed using DVC (development version 0.22.0). Each data file has a .dvc history file that contains the command that was used to generate the output along with md5 hashes of its dependencies.

Installation

• Install python 3, ideally using conda
• Run pip install -r requirements.txt to download DVC and other requirements

Downloading a file

• Run the AWS config script to generate temporary AWS credentials in ~/.aws/credentials:
  • generate-aws-config --account lab --insecure
• Run dvc pull data/path/to/file.dvc to download required file.

High-level overview

Main folders

• bgc_detection/ all the code
• data/ all the data
  • bacteria/ 3k reference bacteria
    • candidates/ novel detected BGC candidates
  • clusterfinder/ ClusterFinder (Cimermancic et al.) datasets
  • evaluation/ Cross-validation, Leave-Class-Out and Bootstrap evaluation
  • features/ Pfam2vec and other protein domain features
  • figures/ Paper figures
  • mibig/ MIBiG BGC database samples
  • models/ Model configurations and trained models
  • pfam/ Pfam repository files
  • training/ Negative and positive training data and t-SNE visualizations
• notebooks/ Jupyter (iPython) notebooks

Training a model

• Define a JSON config file, see data/models/config for reference.
• Run bgc_detection/run_training.py with given config and path to training data. See DVC files in data/models/trained for reference.
• Trained model will be presented as Python pickle file.

Predicting using trained model

• Prepare a protein FASTA file, e.g. using Prodigal (see data/bacteria/proteins.dvc for reference) or extract it from an annotated GenBank file using bgc_detection/preprocessing/proteins2fasta.py.
• Detect protein domains using Hmmscan (see data/bacteria/domtbl.dvc for reference)
• Convert the Hmmscan domtbl file into a Domain CSV file using bgc_detection/preprocessing/domtbl2csv.py (see data/bacteria/domains.dvc for reference)
• Predict BGC domain-level probability using bgc_detection/run_prediction.py (see data/bacteria/prediction/128lstm-100pfamdim-8pfamiter-posweighted-neg-10k.dvc for reference)
• Threshold and merge domain-level predictions into a BGC candidate CSV file using bgc_detection/candidates/threshold_candidates.py (see [data/bacteria/candidates/128lstm-100pfamdim-8pfamiter-posweighted-neg-10k-fpr2/candidates.csv.dvc] for reference)

Bootstrap validation on 9 Fully-annotated genomes

See notebooks/LabelledContigBootstrap.ipynb.

Leave Class Out validation and Cross validation

See data/evaluation/lco-neg-10k (TODO).

See data/evaluation/cv-10fold-neg-10k (TODO).

Random Forest classification

See notebooks/CandidateClassification.ipynb and notebooks/CandidateActivityClassification.ipynb

Novel BGC candidates generation

See notebooks/NovelCandidates.ipynb.