QSAR/QSPR using descriptor-free molecular embedding

This code relies on Keras for the machine learning framework, Theano for computations as its back-end, and RDKit for parsing molecules from SMILES strings. Plotting is done in matplotlib. All other required packages should be dependencies of Keras, Theano, or RDKit.

This code implements the tensor-based convolutional embedding strategy described in placeholder for QSAR/QSPR tasks. The model architecture, training schedule, and data source are defined in a configuration file and trained using a cross-validation (CV). The basic architecture is as follows:

• Pre-processing to convert a SMILES string into an attributed graph, then into an attributed adjacency tensor
• Convolutional embedding layer, which takings a molecular tensor and produces a learned feature vector
• Optional dropout layer
• Optional hidden densely-connected neural network layer
• Optional dropout layer
• Optional second hidden densely-connected neural network layer
• Linear dense output layer

Models are built, trained, and tested with the command

python conv_qsar/main/main_cv.py conv_qsar/inputs/<input_file>.cfg

Numerous example input files, corresponding the models described in placeholder are included in inputs. These include models to be trained on full datasets, 5-fold CVs with internal validation and early stopping, 5-fold CVs without internal validation, models initialized with weights from other trained models, and multi-task models predicting on multiple data sets. Note that when using multi-task models, the output_size must be increased and the loss function must be custom to ensure NaN values are filtered out if not all inputs x have the full set of outputs y.

There are four available data sets in this version of the code contained in data:

1. Abraham octanol solubility data, from Abraham and Admire's 2014 paper.
2. Delaney aqueous solubility data, from Delaney's 2004 paper.
3. Bradley double plus good melting point data, from Bradley's open science notebook initiative.
4. Tox21 data from the Tox21 Data Challenge 2014, describing toxicity against 12 targets.

Because certain entries could not be unambiguously resolved into chemical structures, or because duplicates in the data sets were found, the effective data sets after processing are exported using scripts/save_data.py as coley_abraham.tdf, coley_delaney.tdf, coley_bradley.tdf, coley_tox21.tdf, and coley_tox21-test.tdf.

This version of the code contains the general method of non-linear model interpretation of assigning individual atom and bond attributes to their average value in the molecular tensor representation. The extent to which this hurts performance is indicative of how dependent a trained model has become on that atom/bond feature. As long as the configuration file defines a model which loads previously-trained weights, the testing routine is performed by

python conv_qsar/main/test_index_removal.py conv_qsar/inputs/<input_file>.cfg

It is assumed that the trained model used molecular_attributes, as the indices for removal are hard-coded into this script.

The range of possible architectures (beyond what is enabled with the current configuration file style) can be extended by modifying build_model in main/core.py. See the Keras documentation for ideas.

Additional .csv data sets can be incorporated by adding an additional elif statement to main/data.py. As long as one column corresponds to SMILES strings and another to the property target, the existing code can be used with minimal modification.

Additional atom- or bond-level attributes can be included by modifying utils/neural_fp.py, specifically the bondAttributes and atomAttributes functions. Because molecules are already stored as RDKit molecule objects, any property calculable in RDKit can easily be added.