Python 3 Quickstart

Go to website directory and start the Django app:

cd ../website
python manage.py runserver

Go to http://localhost:8000/visualize/test in a browser to use the interactive visualization tool on the test dataset predictions Go to http://localhost:8000/visualize/validation in a browser to use the interactive visualization tool on the validation dataset predictions

rexgen_direct

Template-free prediction of organic reaction outcomes using graph convolutional neural networks

Described in A graph-convolutional neural network model for the prediction of chemical reactivity

Dependencies

• Python (trained/tested using 2.7.6, visualization/deployment compatible with 3.6.1)
• Numpy (trained/tested using 1.12.0, visualization/deployment compatible with 1.14.0)
• Tensorflow (trained/tested using 1.3.0, visualization/deployment compatible with 1.6.0)
• RDKit (trained/tested using 2017.09.1, visualization/deployment compatible with 2017.09.3)
• Django (visualization compatible with 2.0.6)

note: there may be some issues with relative imports when using Python 2 now; this should be easy to resolve by removing the periods preceding package names

Instructions

Looking at predictions from the test set

cd into the website folder and start the Django app using python manage.py runserver. Go to http://localhost:8000/visualize in a browser to use the interactive visualization tool

Using the trained models

You can use the fully trained model to predict outcomes by following the example at the end of rexgen_direct/rank_diff_wln/directcandranker.py

Retraining the models

Look at the two text files in rexgen_direct/core_wln_global/notes.txt and rexgen_direct/rank_diff_wln/notes.txt for the exact commands used for training, validation, and testing. You will have to unarchive the data files after cloning this repo.

Python 3 Instructions

Copy 1976_Sep2016_USPTOgrants_smiles.rsmi into rexgen_direct/data.

Run data preprocessing script

cd rexgen_direct/data
python prep_data.py

Create cbond_detailed file

cd ../core_wln_global

python nntest_direct.py --test ../data/custom_filtered.rsmi.proc --hidden 300 --depth 3 --model model-300-3-direct --checkpoint ckpt-140000 \
--verbose 1 --detailed 1 > model-300-3-direct/new_data.cbond_detailed

Sample Output:

After seeing 200, acc@12: 0.725, acc@16: 0.745, acc@20: 0.755, acc@40: 0.810, acc@80: 0.850
After seeing 300, acc@12: 0.727, acc@16: 0.743, acc@20: 0.760, acc@40: 0.817, acc@80: 0.860
After seeing 400, acc@12: 0.730, acc@16: 0.750, acc@20: 0.765, acc@40: 0.812, acc@80: 0.855
After seeing 500, acc@12: 0.736, acc@16: 0.752, acc@20: 0.764, acc@40: 0.810, acc@80: 0.846
After seeing 600, acc@12: 0.735, acc@16: 0.752, acc@20: 0.765, acc@40: 0.817, acc@80: 0.853
After seeing 700, acc@12: 0.743, acc@16: 0.759, acc@20: 0.770, acc@40: 0.819, acc@80: 0.851
After seeing 800, acc@12: 0.744, acc@16: 0.760, acc@20: 0.774, acc@40: 0.824, acc@80: 0.855
After seeing 900, acc@12: 0.746, acc@16: 0.761, acc@20: 0.773, acc@40: 0.821, acc@80: 0.852
After seeing 1000, acc@12: 0.744, acc@16: 0.763, acc@20: 0.775, acc@40: 0.826, acc@80: 0.857
After seeing 1100, acc@12: 0.750, acc@16: 0.768, acc@20: 0.782, acc@40: 0.831, acc@80: 0.862
After seeing 1200, acc@12: 0.748, acc@16: 0.766, acc@20: 0.780, acc@40: 0.827, acc@80: 0.858
...

Get bond predictions - includes reactivity scores in output

cd ../rank_diff_wln

python nntest_direct_useScores.py --test ../data/custom_filtered.rsmi.proc --cand ../core_wln_global/model-300-3-direct/new_data.cbond_detailed --hidden 500 --depth 3 --ncand 1500   --ncore 16 --model model-core16-500-3-max150-direct-useScores --checkpoint ckpt-2400000 --verbose 1 > model-core16-500-3-max150-direct-useScores/new_data.cbond_detailed_2400000

python ../scripts/eval_by_smiles.py --gold ../data/custom_filtered.rsmi.proc --pred model-core16-500-3-max150-direct-useScores/new_data.cbond_detailed_2400000 --bonds_as_doubles true

Go to website directory and start the Django app:

cd ../website
python manage.py runserver

Go to http://localhost:8000/visualize/test in a browser to use the interactive visualization tool on the test dataset predictions Go to http://localhost:8000/visualize/validation in a browser to use the interactive visualization tool on the validation dataset predictions