Paper: Maus et al., "Scalable Deep Bayesian optimization over Structured Inputs," arXiv:2201.11872.
A nice and interesting presentation about finding synthetizable molecules that bind to a given protein with Bayesian optimization.
An autoencoder is trained on a database of molecules. The latent space of the autoencoder is fed as input space to an optimizer. The key idea is training the autoencoder together with the surrogate model used by optimizer, such that the latent space makes the target function to maximize simpler.
This technique works well. As expected, the powerful optimizer hacks the target function designed by humans and finds weird molecules that produce huge scores.
Paper: Maus et al., "Scalable Deep Bayesian optimization over Structured Inputs," arXiv:2201.11872.
A nice and interesting presentation about finding synthetizable molecules that bind to a given protein with Bayesian optimization.
An autoencoder is trained on a database of molecules. The latent space of the autoencoder is fed as input space to an optimizer. The key idea is training the autoencoder together with the surrogate model used by optimizer, such that the latent space makes the target function to maximize simpler.
This technique works well. As expected, the powerful optimizer hacks the target function designed by humans and finds weird molecules that produce huge scores.