Neuraldecipher

Implementation of the Paper "Neuraldecipher - Reverse-engineering extended-connectivity fingerprints (ECFPs) to their molecular structures" by Tuan Le, Robin Winter, Frank Noé and Djork-Arné Clevert.¹

Installation

Prerequisites: python==3.6.10

rdkit==2020.03.2
numpy==1.18.1
tqdm==4.46.1
h5py==2.10.0
jupyter==1.0.0

Conda

Create a new enviorment:

git clone URL
cd neuraldecipher
conda env create -f environment.yml
conda activate neuraldecipher

pytorch==1.4.0 (GPU with cuda10 or CPU)

conda install pytorch==1.4.0 torchvision==0.5.0 -c pytorch # GPU
# conda install pytorch==1.4.0 torchvision cpuonly -c pytorch # CPU

Dependency for encoding and decoding SMILES representations

• cddd
  To complete the reverse-engineering workflow, the decoder network from Winter et al. (see Workflow) is needed in the final evaluation. Note, it suffices to clone the cddd repository and start from the installation of tensorflow-gpu==1.10.0 without creating the environment. It is important to have the cddd module installed within the neuraldecipher environment for latter inference. To use tensordboard with pytorch, remove the tensorboard==1.10.0 from the cddd dependency
  pip uninstall tensorboard
pip install tensorboard==1.14.0
  We included this workaround to still be able to use the CDDD inference server and tensorboard to log the training of the Neuraldecipher.
  The CDDD server is also needed to compute the CDDD vector representation from the SMILES to train the Neuraldecipher.
  We provided a Jupyter Notebook in source/get_cddd.ipynb to compute the CDDD representations from the ChEMBL25 dataset.

Repository structure

The repository consists of several subdirectories:

• data consists of the training and test data.
• logs consists of the tensorboard log files for each training run
• params consists of the json parameter files for each run. See example.
• models consists of the saved models. In case the Neuraldecipher was trained on bit-ECFPs, the results are saved in models/bits_results. Otherwise the models are saved in models.
• source consists of all necessary python scripts for execution.

The provided data consists of:

• data/smiles.npy: List of SMILES from the filtered ChEMBL25 database saved as numpy array.
• data/smiles_temporal.npy: List of temporal SMILES from the filtered ChEMBL26 database saved as numpy array.
• data/cluster.npy: List of cluster assignment from the smiles.npy array. This array is needed to create train and validation datasets.

Getting started

Computing several extended-connectivity fingerprints (ECFPs) depending on length k and bond diameter d

The python script in source/get_ecfp.py computes the extended-connectivity fingerprints.
The options for the script are the following:

--all: Boolean flag whether or not all ECFP configurations as described in the paper¹ should be computed. Defaults to False. In this case on the ECFP with bond-diameter d=6 and fingerprint size k=1024 are computed for the binary and count representations.

--nworkers: Integer of number of parallel cpu-workers to use in order to compute the ECFP representations. Defaults to 1. In order to speed up the computation, it is recommended to use more workers.

Execution:

python source/get_ecfp.py -h # in order to see the information for the arguments
python source/get_ecfp.py --all False --nworkers 10 # only compute one ECFP setting and use 10 cpus for multiprocessing

Computing CDDD representations

The Jupyter Notebook in source/get_cddd.ipynb shows how to generate CDDD representations from the data/smiles.npy array.

Training the Neuraldecipher model

The python script in source/main.py excutes the training for the Neuraldecipher.
The options for the script are the following:

--config: String to the params.json file that consists the information for Neuraldecipher network architecture and training settings. Defaults to params/1024_config_bit_gpu.json

--split: String to select if the cluster or random split should be used (see reference ¹) for details.
Defaults to cluster.

--workers: Integer of number of parallel cpu-workers for the dataloader. Defaults to 5

--cosineloss: Boolean flag whether or not the cosineloss should be used within the training. Defaults to False. This flag can be set to True to additionally add the cosine similarity loss next to the difference loss (e.g. L2, or logcosh).

Execution:

python source/main.py -h # in order to see the information for the arguments
python source/main.py --config params/1024_config_bit_gpu.json --split cluster --workers 5 --cosineloss False

Monitoring the training

Since tensorboard-gpu==1.10.0 is installed within the neuraldecipher environment, we cannot run tensorboard==1.14.0 within the neuraldecipherenvironment. We merely included tensorboard==1.14.0 to the neuraldecipher environment to log the training of our Neuraldecipher. To monitor the training, please create a new environment tb and install tensorflow==1.14.0 (CPU version) which also includes tensorboard==1.14.0 in its installation.

conda create -n tb python=3.6.10 tensorflow==1.14.0
conda activate tb

Run tensorboard command in a new shell (here to localhost:8888):

tensorboard --logdir logs/ --port 8888 --host localhost

Evaluationg the trained model

We provide the model weights for the trained model on ECFP6 representations of length 1024 trained on the cluster split and show the performance on the cluster validation dataset and temporal dataset in in the Notebook source/evaluation.ipynb.

References

[1] T. Le, R. Winter, F. Noe and D. Clevert, Chem. Sci., 2020, DOI: 10.1039/D0SC03115A