Chemical VAE

I have implemented the variational autoencoder (VAE) described in the paper entitled "Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules."¹ I have compared performance using either SELFIE or SMILE strings as inputs.²

References

[1] R. Gómez-Bombarelli et al., Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules. ACS Central Science 4, 268-276 (2018).

[2] M. Krenn, F. Häse, A. Nigam, P. Friederich, A. Aspuru-Guzik, Self-referencing embedded strings (SELFIES): A 100% robust molecular string representation. Machine Learning: Science and Technology 1, 045024 (2020).

I have referred to the following implementations of VAEs for molecule generation for guidance:

• https://github.com/aspuru-guzik-group/chemical_vae
• https://github.com/molecularsets/moses/tree/master/moses/vae
• https://github.com/cxhernandez/molencoder
• https://github.com/bayeslabs/genmol
• https://github.com/Ishan-Kumar2/Molecular_VAE_Pytorch
• https://github.com/aksub99/molecular-vae
• https://github.com/topazape/molecular-VAE
• http://bits.csb.pitt.edu/mscbio2066/assign7/train_VAE.py