4 ms·

Chainer Chemistry: A Library for Deep Learning in Biology and Chemistry

59 points by wei_jok 9y ago

Pulcinella 9y ago
This seems like the development of an artificial chemical intuition[0][1] except since its a computer it can intuit concrete numbers much more comfortably than humans can. I do wonder how its accuracy compares to just running the DFT calculations. [0]wavefunction.fieldofscience.com/2016/09/what-is-chemical-intuition.html [1]https://news.ycombinator.com/item?id=12624036 https://news.ycombinator.com/item?id=12624036
- lilleswing 9y ago
  There is a lot of research on attempting to approximate DFT level energy calculations with simple Neural Networks. https://arxiv.org/abs/1610.08935 https://arxiv.org/abs/1610.08935 https://github.com/deepchem/deepchem/blob/master/deepchem/models/tensorgraph/models/symmetry_function_regression.py#L113 https://github.com/deepchem/deepchem/blob/master/deepchem/mo... http://aip.scitation.org/doi/pdf/10.1063/1.4986081 http://aip.scitation.org/doi/pdf/10.1063/1.4986081 It is a pretty straight-forward application of ML to the space that is starting to have solid results in practice.
brilee 9y ago
Excellent. I'd been wanting to implement something like this but these guys beat me to the punch. Most existing research on NNs in chemistry uses the SMILES string representation and feeds it into a recurrent neural network. The problem with SMILES in my opinion, is that SMILES is not unique, and there are a gazillion ways to represent the same molecule. Thus, the RNN is probably overfitting to the quirks of the SMILES serialization algorithm used by the US Patent Office and/or ChemDraw and/or whatever program was used to generate the library of molecules. Convolutional graph networks are the obviously correct way to handle molecules.
- dekhn 9y ago
  they use canonical SMILES. Problem: there's more than one way to make canonical SMILES :(
lilleswing 9y ago
Really cool stuff here, a lot along the same lines of what we have been trying to do with DeepChem and MoleculeNet. https://deepchem.io/ https://deepchem.io/ https://github.com/deepchem/deepchem https://github.com/deepchem/deepchem http://moleculenet.ai/ http://moleculenet.ai/ We seem to be a little further along and stable than this repo. We are also based on Tensorflow instead of raw cuda which means users can play around with the networks more easily.
- akapajama 9y ago
  This is based on Chainer, not raw cuda, meaning its also very flexible for the user
corochann 9y ago
Links of Chainer Chemistry - blog: https://preferredresearch.jp/2017/12/18/chainer-chemistry-beta-release/ https://preferredresearch.jp/2017/12/18/chainer-chemistry-be... - github: https://github.com/pfnet-research/chainer-chemistry https://github.com/pfnet-research/chainer-chemistry - document: http://chainer-chemistry.readthedocs.io/en/latest/index.html http://chainer-chemistry.readthedocs.io/en/latest/index.html

59 points by wei_jok 9y ago

This seems like the development of an artificial chemical intuition[0][1] except since its a computer it can intuit concrete numbers much more comfortably than humans can. I do wonder how its accuracy compares to just running the DFT calculations. [0]wavefunction.fieldofscience.com/2016/09/what-is-chemical-intuition.html [1]https://news.ycombinator.com/item?id=12624036 https://news.ycombinator.com/item?id=12624036

lilleswing 9y ago
There is a lot of research on attempting to approximate DFT level energy calculations with simple Neural Networks. https://arxiv.org/abs/1610.08935 https://arxiv.org/abs/1610.08935 https://github.com/deepchem/deepchem/blob/master/deepchem/models/tensorgraph/models/symmetry_function_regression.py#L113 https://github.com/deepchem/deepchem/blob/master/deepchem/mo... http://aip.scitation.org/doi/pdf/10.1063/1.4986081 http://aip.scitation.org/doi/pdf/10.1063/1.4986081 It is a pretty straight-forward application of ML to the space that is starting to have solid results in practice.

brilee 9y ago

Excellent. I'd been wanting to implement something like this but these guys beat me to the punch. Most existing research on NNs in chemistry uses the SMILES string representation and feeds it into a recurrent neural network. The problem with SMILES in my opinion, is that SMILES is not unique, and there are a gazillion ways to represent the same molecule. Thus, the RNN is probably overfitting to the quirks of the SMILES serialization algorithm used by the US Patent Office and/or ChemDraw and/or whatever program was used to generate the library of molecules. Convolutional graph networks are the obviously correct way to handle molecules.

dekhn 9y ago
they use canonical SMILES. Problem: there's more than one way to make canonical SMILES :(

lilleswing 9y ago

Really cool stuff here, a lot along the same lines of what we have been trying to do with DeepChem and MoleculeNet. https://deepchem.io/ https://deepchem.io/ https://github.com/deepchem/deepchem https://github.com/deepchem/deepchem http://moleculenet.ai/ http://moleculenet.ai/ We seem to be a little further along and stable than this repo. We are also based on Tensorflow instead of raw cuda which means users can play around with the networks more easily.

akapajama 9y ago
This is based on Chainer, not raw cuda, meaning its also very flexible for the user

corochann 9y ago

Links of Chainer Chemistry - blog: https://preferredresearch.jp/2017/12/18/chainer-chemistry-beta-release/ https://preferredresearch.jp/2017/12/18/chainer-chemistry-be... - github: https://github.com/pfnet-research/chainer-chemistry https://github.com/pfnet-research/chainer-chemistry - document: http://chainer-chemistry.readthedocs.io/en/latest/index.html http://chainer-chemistry.readthedocs.io/en/latest/index.html