Project allows training chemical networks using deep learning. It trains a neural network and translates it to an equivalent chemical network. This is done based on the tight connection, between chemical and neural models of computation, that we discovered.

Two publications are result of this project:

@article{pnas22TrainingCRNs,
  title = {Programming and Training Rate-Independent Chemical Reaction Networks},
  author = {Vasic, Marko and Chalk, Cameron and Luchsinger, Austin and Khurshid, Sarfraz and Soloveichik, David},
  journal = {Proceedings of the National Academy of Sciences},
  year = {2022}
}

@inproceedings{icml20DeepMolecularProgramming,
  title={{D}eep {M}olecular {P}rogramming: {A} {N}atural {I}mplementation of {B}inary-{W}eight {R}e{L}{U} {N}eural {N}etworks},
  author = {Vasic, Marko and Chalk, Cameron and Khurshid, Sarfraz and Soloveichik, David},
  booktitle = {International Conference on Machine Learning},
  year = {2020},
}

If you would like to reference them in an academic publication please cite the previous papers.

Our YouTube Presentations of associated papers:

• Programming and Training Rate-Independent Chemical Reaction Networks
• Deep Molecular Programming

• Following software is needed to run the code:
  • Theano (0.7.0 or higher)
  • Lasagne (0.1 or higher)
  • pylearn2 (0.1.dev0)
  • Mathematica (11.2 or higher)

• Note that neural network models (used in the publications mentioned above) as well as translated chemical networks are included in the repo and are ready to be used. Pretrained neural networks are saved under data-repo/models directory while chemical networks obtained by translating those are saved under data-repo/mathematica directory. Thus, one can skip steps 1 and 2 below and go directly to step 3 of simulating existing chemical networks.

• Navigate to the src directory.

• Step 1: Training. To train a neural network run: python -m crn.subject --train; where currently supported subjects are: iris, virus, pattern_formation*, mnist-subset, mnist.

• Model will be saved under data-repo/models directory with pkl extension.

• Step 2: Translation. After training a model you can translate it to a CRN by running python -m crn.subject --translate.

• Translated CRN will be stored under data-repo/mathematica directory with wls extension (wls is a Mathematica script file).

• Step 3: Chemical Simulations. Finally, you can run Mathematica simulations by navigating to data-repo/mathematica, and executing the produced wls file.

• Kinetics simulations of the produced CRN will be stored under data-repo/kinetics directory.

• Possible issues: Note that MNIST neural network model (mnist.pkl) might fail to translate on some systems due to compatibility issues. This shouldn't prevent you to use CRN obtained from translating that neural network, which we saved in mnist.wls file. We are working on translating model files to a new, more portable format.

Some of the graphics in pattern formation dataset (data-repo/datasets/pattern_formation/graphics) are created by Joseph Wain and licensed under CC BY 3.0 US.

To train binary neural networks we adapt BinaryConnect code. We augment BinaryConnect to support zero weights. BinaryConnect code resides in src/binary_connect directory.