Release v0.4.0

[kmaziarz]

This PR edits the CHANGELOG to mark the v0.4.0 release. After it's merged, I will tag the resulting commit, and then push the release to PyPI.

In particular, this release includes several changes that make molecule_generation compatible with the latest versions of rdkit and numpy.

[thk9178]

Thank you for publishing your great work.
In addition to VAE, I would like to supervised train my SMILES-property data, do you have a manual for the program?
If not, how can I train it with my property data?

[kmaziarz]

In addition to VAE, I would like to supervised train my SMILES-property data, do you have a manual for the program? If not, how can I train it with my property data?

Thank you for your question @thk9178, and sorry for the slow response. This package only provides the code to train the VAE, which would typically be done in a property-agnostic way. If you have some property data, there are essentially two routes:

• Train a separate supervised property predictor, and use it to perform optimization in the latent space of a trained MoLeR model (for the optimization part itself you could use things like Bayesian Optimisation, Molecular Swarm Optimisation, Genetic Algorithms, etc).
• Take MoLeR trained in a property-agnostic way and then fine-tune it further on samples with high property values. Then "property optimization" can be performed by sampling from the prior directly and ranking the results. We haven't explored this direction much though.

If you take the first route, then you need to train a supervised property prediction model, for which you'd have to look elsewhere. A simple starting point would be to just train a shallow MLP on molecular fingerprints.

[thk9178]

In addition to VAE, I would like to supervised train my SMILES-property data, do you have a manual for the program? If not, how can I train it with my property data?

Thank you for your question @thk9178, and sorry for the slow response. This package only provides the code to train the VAE, which would typically be done in a property-agnostic way. If you have some property data, there are essentially two routes:

• Train a separate supervised property predictor, and use it to perform optimization in the latent space of a trained MoLeR model (for the optimization part itself you could use things like Bayesian Optimisation, Molecular Swarm Optimisation, Genetic Algorithms, etc).
• Take MoLeR trained in a property-agnostic way and then fine-tune it further on samples with high property values. Then "property optimization" can be performed by sampling from the prior directly and ranking the results. We haven't explored this direction much though.

If you take the first route, then you need to train a supervised property prediction model, for which you'd have to look elsewhere. A simple starting point would be to just train a shallow MLP on molecular fingerprints.

I'm so grateful for your help. I can definitely do that as you recommended.
I'll have to pray that MoLer's latent space shows a smooth gradient for my property data. 😁
Thank you very much for your kind reply.