Mitchell Pavlak mpavlak1@jhu.edu
Brandon Liu brandonmliu@gmail.com
Emmanuel Osikpa eosikpa1@jhu.edu
Abbey Thrope Abbey.thrope@gmail.com
The program is separated into four pieces. The primary components are Main.py, getSimilarity.py, and saveMethods.py. plotting.py generates a matplotlib plot to help show the results of the classifiers over all ADR tags.
Running steps vary based on whether or not a save file is available. If not,
run python getSimilarity.py. Keep all the folder structure the same as upon download.
Then run python Main.py. You can also run python plotting.py.
All code used in this project should be our own work or open-sourced code. Some specific python libraries used include scikit-learn, subprocess, multiprocessing, numpy, tqdm, pandas, cPickle.
Additionally, LSAlign was used to source PCScore and get rigid small molecule alignment. LSAlign is open for any academic reuse or modification. For more information about LSAlign, see: https://zhanglab.ccmb.med.umich.edu/papers/2018_1.pdf
See also the following: https://academic.oup.com/bioinformatics/article/32/15/2338/1744048
The goal of this project was to leverage 3D molecular structure data as a potential area for improvement of ADR prediction. We used LSAlign to align the small molecules. Instead of taking more detailed measures of 3D shape, we used a PCScore as developed in the top paper above (LSAlign). This metric allowed us to quantify the similarity in 3D structure, bond activity etc between various molecules. By training with a distance matrix of these similarities in addition to gene expression data, we were able to see slight improvements in accuracy over using the full 2d fingerprint of the molecule with GE data. Further, our 3D model required little over a third of the training time comparable 2D models we created required.