What?

A model + task specification to obtain molecules that we can synthesize in real life.

Why?

Generating a molecule with a model tells us what to synthesize. Knowing how to do that would be very useful. Moreover, the model might generate a molecule that is not possible to get in the lab. How can avoid that?

How?

source: original paper

The authors consider two problems:

• G1: generating molecules for further screening and scoring;
• G2: finding molecules satisfying some requirements.

The main idea is to build a synthesis task in a way that gives us an idea of how to synthesize a molecule in the real-life as well as reduce the probability of getting molecules impossible in the real life. They turn it into a sequential decision making problem:

• A model can do three types of actions:
  • add a node
    • a building block
    • a product node
  • choose a building block (if a previous action was added, we need to select a molecule to add)
  • connect nodes
    • After a product node was selected, we need to choose which nodes participate in making the product.
    • Stop making a product and stop synthesis actions are also available.
  • masking is applied to avoid choosing the same molecule block again.

The process above can be represented as a DAG, and, gives us an idea of how to synthesize the molecule in the lab.

How do the authors solve the problem:

• General model:

  

  source: original paper

  • RNN to get a latent $z$
  • An RNN takes previous action embedding as an input.
  • Action means to pick a molecule (which is a graph), GGNN is used to get an embedding.
  • Output of a RNN goes into different heads (one head per action type)
  • When building a product, we need to know the result of it. The authors assume access to an oracle for that.

• G1 (generate a ton of molecules)

  • Using VAE here (DAG of molecular graphs (DOG-AE))
  • Wasserstein AEs are used here: $\min_{\phi, \theta}\mathbb{E}{\mathcal{M}\sim p(\mathcal{M})}\mathbb{q\phi(z\mid \mathcal{M})}[-\log{p}\theta(\mathcal{M}\mid z)]+\lambda \mathcal{D}(q\phi(z), p(z))$, where $\mathcal{D}$ is MMD.
  • Encoder is interesting here. The authors to hierarchical message passing:

    • first level is molecular level to get an embedding of the blocks

    • second level is message passing on the synthesis DAG

      

• G2 (generate a molecule with given properties)

  • Hill climbing from GuacaMol (a cross-entropy method that can be view as REINFORCE with some reward shaping).
  • Train a generative model.
  • Fine-tune the decoder (sample a lot of candidates and fine-tune on top-K).

And?

• It's folklore already, but SMILES is such a strong baseline!
• While SMILES shows good performance, it gives molecules which we cannot synthesize. And the proposed approach >> SMILES judging by this metric.