Efficient molecular feature learning for low data QSAR

Official repository of the ISCB paper

Requirements and installation

A new virtual environment with all the requirement can be created automatically by running the installation script as follows:

bash install.sh

Otherwise, we highly recommend you to use conda for package management and make sure you met at least these requirements:

• RDKit (version >= 2017.09)
• Python (version >= 3.4)
• PyTorch (version >= 0.2)

For a more complete list of requirements, please look inside the install.sh

Datasets and methods

All the datasets used in our experiments are available in the datasets folder :

• Pubchem toxicity dataset collection at datasets/pubchemtox/
• MetaQsar dataset collection at datasets/chembl/
• MHC dataset collection at datasets/mhc/

All the methods are available as well :

• MetaKRR: metalearn/models/metakrr_singlekernel.py
• RF+ECFP4 and KRR+ECFP4: metalearn/models/fp_learner.py
• Seq2Seq: metalearn/models/seq2seq_fingerprint.py
• IterRefLSTM: metalearn/low_data/deepchem_meta.py
• MANN: metalearn/models/mann.py

Experiments and training

All the experiments in the iscb paper are available in the folders expts_iscb and expts_low_data. The latter only contains the experiments related to the IterRefLSTM algorithm. The train file in each folder should allow to train a model and save the trained weights as well as its performances on the meta-test.

To actually train a model, please

1. Create a train.json where you specify which config list you want to use and an id indicating the config position in that list. Example of train.json

{ 'config_file': 'config_krr.json', 'config_id': 0 }

2. Execute the training file

python train --config_file train.json --input_path ../datasets --output_path ./results

Contact

Prudencio Tossou