We report a method to convert discrete representations of molecules to and
from a multidimensional continuous representation. This model allows us to
generate new molecules for efficient exploration and optimization through
open-ended spaces of chemical compounds. A deep neural network was trained on
hundreds of thousands of existing chemical structures to construct three
coupled functions: an encoder, a decoder, and a predictor. The encoder
converts the discrete representation of a molecule into a real-valued
continuous vector, and the decoder converts these continuous vectors back to
discrete molecular representations. The predictor estimates chemical
properties from the latent continuous vector representation of the molecule.
Continuous representations of molecules allow us to automatically generate
novel chemical structures by performing simple operations in the latent space,
such as decoding random vectors, perturbing known chemical structures, or
interpolating between molecules. Continuous representations also allow the use
of powerful gradient-based optimization to efficiently guide the search for
optimized functional compounds. We demonstrate our method in the domain of
drug-like molecules and also in a set of molecules with fewer that nine heavy
atoms.
We report a method to convert discrete representations of molecules to and from a multidimensional continuous representation. This model allows us to generate new molecules for efficient exploration and optimization through open-ended spaces of chemical compounds. A deep neural network was trained on hundreds of thousands of existing chemical structures to construct three coupled functions: an encoder, a decoder, and a predictor. The encoder converts the discrete representation of a molecule into a real-valued continuous vector, and the decoder converts these continuous vectors back to discrete molecular representations. The predictor estimates chemical properties from the latent continuous vector representation of the molecule. Continuous representations of molecules allow us to automatically generate novel chemical structures by performing simple operations in the latent space, such as decoding random vectors, perturbing known chemical structures, or interpolating between molecules. Continuous representations also allow the use of powerful gradient-based optimization to efficiently guide the search for optimized functional compounds. We demonstrate our method in the domain of drug-like molecules and also in a set of molecules with fewer that nine heavy atoms.
Notes
-------
This is currently an imperfect reproduction of the paper. One difference is
that teacher forcing in the decoder is not implemented. The paper also
discusses co-learning molecular properties at the same time as training the
encoder/decoder. This is not done here. The hyperparameters chosen are from
ZINC dataset.
This network also currently suffers from exploding gradients. Care has to be taken when training.
TODO(LESWING): Teacher Forcing
"""
def__init__(self,
input_tokens,
output_tokens,
max_output_length,
encoder_layers=4,
decoder_layers=4,
embedding_dimension=512,
dropout=0.0,
reverse_input=True,
embedding_dimension=196,
variational=True,
annealing_start_step=5000,
annealing_final_step=10000,
filter_sizes=[9,9,10],
kernel_sizes=[9,9,11],
decoder_dimension=488,
**kwargs):
"""
TODO(LESWING) have qm9/zinc hyper params and construct via static method
Parameters
----------
filter_sizes: list of int
Number of filters for each 1D convolution in the encoder
kernel_sizes: list of int
Kernel size for each 1D convolution in the encoder
decoder_dimension: int
Number of channels for the GRU Decoder
"""
self.filter_sizes=[9,9,10]
self.kernel_sizes=[9,9,11]
iflen(filter_sizes)!=len(kernel_sizes):
raiseValueError("Must have same number of layers and kernels")
