SuzukiConditions

Here are scipts for data processing, inputs generation and logs from training for all models described in the paper entitled Machine Learning May Simply Capture Literature Popularity Trends: The case of Suzuki-Miyaura coupling. by Wiktor Beker, Rafał Roszak, Agnieszka Wołos, Nicholas H. Angello, Vandana Rathore, Martin D. Burke, and Bartosz A. Grzybowski. Link and DOI to the paper will be added when will be provided by publisher. DOI to this repo is 10.5281/zenodo.4652819

Contents

The repo consists of 4 parts placed in separated directories:

• dataPreparation/ - scripts used for preparation and filtration data. To avoid any legal doubt we are not publish a reaction set which was extracted from reaxys database. One can download the dataset from reaxsys using published reaxsysID or can prepare own dataset - see readme.md file in dataPreparation/ for details
• classification/ - scripts used for classification task with input and logs obtained from training
• regression/ - scripts used regression task and related input and logs
• solvent_classification/ - scripts and data for advanced model for solvent classification.
• uspto/ - all scripts and data related to experiments with USPTO

How to start

1. Unfortunately we cannot publish our reaction dataset instead one need to download it manually from reaxsys or use (or generate) own dataset using information in dataPreparation/readme.md
2. When dataset is present, it can be filtered and transform into format suitable for AI learining, see dataPreparation/readme.md for details
3. Once proper input is generated one can train any of tested NN models using scripts from classification/, regression/ or gcnn_and_dan/

Requirements

1. Python
2. RDkit
3. TensorFlow
4. Keras
5. Numpy
6. Scikit-learn
7. additional requirements for advanced models are described in solvent_classification/readme.md