CDDD-ONNX

This package provides CDDD (Continuous and Data-Driven Descriptors) models in ONNX format with automatic model downloading capabilities. This is an ONNX runtime version of the original CDDD package.

Limitations

• Currently, only the encoder model is implemented. The decoder uses TensorFlow-specific opcodes that are not supported by ONNX. I am working on a solution.

Installation

pip install cddd-onnx

Usage

Command Line Interface

Extract molecular descriptors from SMILES using the command line interface:

cddd-onnx --input smiles.smi --output descriptors.csv

For CSV files with a custom SMILES column header:

cddd-onnx --input molecules.csv --output descriptors.csv --smiles_header smiles_column

Python Interface

The format is the same as in the original CDDD package. Just import cddd_onnx instead of cddd:

from cddd_onnx import InferenceModel
# Create model instance
model = InferenceModel()

smiles_list = ["CCCCO", "CCCN", "CC1=CC=CC=C1"]
embeddings = model.seq_to_emb(smiles_list)

• The preprocessing stage is inside the seq_to_emb function, so you do not need to run it separately.
• Be aware that if the SMILES is out of AD, the preprocessor returns None, resulting in a row of NaNs for such compounds.

Input Formats

Supported input formats:

• CSV files with SMILES column
• SMI files (one SMILES per line)

Models

The models are automatically downloaded to ~/.cddd_onnx/models/ directory when first used.

Requirements

• onnxruntime
• numpy
• pandas
• tqdm
• rdkit

License

MIT License

Citation

If you use this software, please cite the original CDDD paper:

@article{Winter2019,
  title = {Learning continuous and data-driven molecular descriptors by translating equivalent chemical representations},
  volume = {10},
  ISSN = {2041-6539},
  url = {http://dx.doi.org/10.1039/C8SC04175J},
  DOI = {10.1039/c8sc04175j},
  number = {6},
  journal = {Chemical Science},
  publisher = {Royal Society of Chemistry (RSC)},
  author = {Winter,  Robin and Montanari,  Floriane and Noé,  Frank and Clevert,  Djork-Arné},
  year = {2019},
  pages = {1692–1701}
}