Overview

CHAMOIS is a fast method for predicting chemical features of natural products produced by Biosynthetic Gene Clusters (BGCs) using only their genomic sequence. It can be used to get chemical features from BGCs predicted in silico with tools such as GECCO or antiSMASH. It is developed by the Zeller Lab at LUMC and EMBL.

This record contains the training data for CHAMOIS version 0.2.0, as well as the data and scripts supporting the analyses presented in the paper. The code for the CHAMOIS tool can be found in the zellerlab/CHAMOIS repository on GitHub. The various HDF5 files are intended to be opened with the anndata library (Virshup 2024) using the anndata.read_h5ad function. Alternatively, the observations metadata, variables metadata, and data tables are given as tab-separated-values (TSV) files in the same folder under the name obs.tsv, var.tsv and X.tsv respectively.

Contents

Each dataset is an archive containing the following files:

• features.hdf5 and features folder : The Pfam v36.0 domain vectors for each BGC of the dataset.
• classes.hdf5 and classes folder: The predicted ChemOnt classes for a selected compound of each BGC of the dataset.
• compound.json : A JSON file listing all compounds per BGC (as the BGCs in classes.hdf5 only have the classification for a single compound).
• types.tsv : The MIBiG types (Polyketide, NRP, RiPP, etc.) for each BGC of the dataset (if any).
• taxonomy.tsv : The taxonomy for the host of each BGC of the dataset (if any).

CHAMOIS can be trained and evaluated on these domains directly using the chamois train and chamois cv commands:

• chamois train -f features.hdf5 -c classes.hdf5 -o model.json
• chamois cv -f features.hdf5 -c classes.hdf5 -o report.tsv
• chamois cvi -f features.hdf5 -c classes.hdf5 -o report.tsv

Datasets

MIBiG 2.0

This dataset contains 1,499 annotated BGCs released in MIBiG 2.0 (Kautsar 2019), excluding some records from a manually curated list, and with manual corrections in BGC coordinates and compound assignment. It also excludes the BGCs that were deprecated or removed in MIBiG 3.1 to avoid low-quality entries.

MIBiG 3.1

This dataset contains 1,940 annotated BGCs released in MIBiG 3.1 (Terlouw 2023), excluding some records from a manually curated list, and with manual corrections in BGC coordinates and compound assignment.

MIBiG 4.0

This dataset contains 1,915 annotated BGCs released in MIBiG 4.0 (Zdouc 2024), excluding some records from a manually curated list, and with manual corrections in BGC coordinates and compound assignment.

Benchmark 

This dataset contains 70 annotated BGCs, found in literature in their native context (the complete host genome) and used in the BGC screening benchmark of the CHAMOIS paper. The BGCs are distinct from the MIBiG 2.0 and 3.1 datasets, so it can be used as an external validation set if needed, although some clusters still exhibit moderate similarity. The dataset also contains the complete sequences of the 65 genomes containing the BGCs.

PRISM 4

This dataset contains 1,267 annotated BGCs from the "Gold Standard BGCs" published in PRISM 4 (Skinnider 2020). It overlaps with the MIBiG datasets.