FeatureCreature is an open-source software tool for structural visualization of machine learned Quantitative Structure Activity Relationships (QSARs) explanations.
QSARs are often black-box models, principally built using difficult to interpret Artificial Neural Networks, Random Forests (and related techniques), or other machine learning techniques. It is common to learn a QSAR that has very high accuracy, precision and recall; but little interpretability. The goal of FeatureCreature is to visualize locally explanations for QSAR predictions or classifications in 2D chemical space.
FeatureCreature relies heavily on our BioCompoundML work.
Whitmore, L. S., Davis, R. W., McCormick, R. L., Gladden, J. M., Simmons, B. A., George, A., & Hudson, C. M. (2016). BioCompoundML: a general biofuel property screening tool for biological molecules using Random Forest Classifiers. Energy & Fuels, 30(10), 8410-8418.
Parts of this work was presented at the 2016 NIPS Workshop on Interpretable Machine Learning in Complex Systems
Whitmore, L.S., George, A. and Hudson, C.M., 2016. Mapping chemical performance on molecular structures using locally interpretable explanations. arXiv preprint arXiv:1611.07443. arxiv
Clone this github repo by running the FeatureCreature Jupyter Notebook (http://jupyter.org)
Visualization functions require the Indigo Toolkit Python bindings (http://lifescience.opensource.epam.com/download/indigo/index.html).
Additionally FeatureCreature requires a scientific python environment, including Numpy, Scipy and scikit-learn. The easiest way to do this is to download and use a conda environment (https://www.continuum.io/downloads)