This repository contains the data and code needed to test a similarity-based approach to one-step retrosynthesis.
Please note that rdchiral is a work-in-progress. The current version as of June 19, 2017 has been copied into this repository for result reproducibility. An up-to-date version can be found at the public repo http://github.com/connorcoley/rdchiral
The set of 50k reactions comes from http://pubs.acs.org/doi/abs/10.1021/acs.jcim.6b00564. Each reaction is pre-labeled with a class number (1-10). The dataset is further cleaned following Liu et al. (2017) (https://arxiv.org/pdf/1706.01643.pdf) so that each reaction has a single product and trivial products are excluded. Atom maps are removed for reactant atoms that do not contribute atoms to the product of interest. data_processed.csv is a Pandas dataframe and is meant to work with the functions in get_data.py.
All of the "heavy lifting" occurs inside the scripts folder. extract_templates is just used for examining the templates corresponding to the training data. Likewise, analyze_templates looks at the some trends and the most common templates, but is not needed in the workflow.
After an initial data processing using proc_data, the test_similarity script actually applies the similarity method using the training data as a corpus. The Jupyter notebook is meant to look at a single condition (i.e., class, fingerprint type, similarity metric) at a time. The standalone script can test the whole suite of conditions. Results are written into results.txt and are saved in separate files.
The notebook process_results reads from results.txt and examines the validation performance visually. This is how the metric was selected for use on the test data, which required a simple modification of the test_similarity script. Test results are also read using process_results and output in a tabular form at the end of the notebook.
For any questions, feel free to email ccoley@mit.edu