Undirected Graph - Recursive Neural Networks for Chemoinformatics (in TensorFlow)

This is a Tensorflow implementation based on the models proposed in "Deep Architectures and Deep Learning in Chemoinformatics: The Prediction of Aqueous Solubility for Drug-Like Molecules" (A. Lusci et al., Journal of Chem. Inf. Modeling, 2013) [1]. It implements recursive neural networks operating on molecular graphs (SMILES) of arbitrary size for chemical property prediction (e.g. solubility).

Short summary of the method/paper:

Properties of a molecule with N atoms are predicted by averaging the predictions of an inner-recursive neural network on N different views of the molecule. In each of the N views, the undirected graph representing the molecule is transformed into a N directed acyclic graphs where all eges are oriented towards one atom - a different atom each time. A small MLP crawls along these directed edges operating on atom- and bond features, and when reaching the 'terminal' atom it emits a vector. There are N of these vectors which are averaged and then used to compute the overall prediction for the molecule. The same MLP is used for each of the N views of the molecule - i.e. the weights are shared.

Requirements:

Python 2 or 3, Numpy -- preferrably using Anaconda Tensorflow -- get version 0.12.0 or newer RDkit -- the easiest way to install it when using Anaconda is "conda install -c https://conda.anaconda.org/rdkit rdkit" networkx scipy scikit-learn (sklearn)

Usage:

run ugrnn/train.py to train models

--model_name

  Name of the model

--max_epochs

  Number of epochs to run the training for

--batch_size

  Batch size (choose a small number on the order of ~ 5)

--model_params

  Model Parameters: a tuple of three nubers (e.g. 7,7,5)
  The first entry is the size of the hidden layer of the network that crawls along the directed 
  edges of a view on the molecule (Mg hidden units in the paper).
  The second entry is the size of the internal output layer of the network that transfers the 
  assembled information from of a view on the molecule to the averaging step (Mg output units in the paper).
  The third entry is the size of the hidden layer of the 'external' output network that makes the 
  final prediction for the molecule after the averaging step (Mo hidden units in the paper).

--learning_rate

  Initial learning rate (defaults to 1e-3)

--output_dir

  Directory for storing the trained models

--training_file

  Path to the csv file containing training data set (e.g. data/delaney/train_delaney.csv)

--validation_file

  Path to the csv file containing validation data set (e.g. data/delaney/validate_delaney.csv)

--smile_col

  Column in the csv file that represents the inputs for the UG-RNN -- i.e. the SMILES of the molecules

--logp_col

  Column in the csv file that holds logP information 
  (if present in data set and if --add_logp option is used)

--target_col

  Column in the csv file that represents the prediction target for the UG-RNN

--contract_rings

  Boolean (see paper for details) [default False]

--add_logp

  Boolean (see paper for details) [default False]

--clip_gradient

  Boolean, specify whether to clip gradients during training, by default False as it is 
  not needed with careful setting of other hyperparameters.

run ugrnn/predict.py to make predictions models

--model_names [can be a list]

  Name of the models used for prediction (one or more)

--model_params [can be a list]

  Model Parameters (same as the ones used for training. 3-tuples, one for each model)