New Series 4 candidates based on generative model - EOSI

[miquelduranfrigola]

Hello @mattodd @edwintse,

At @ersilia-os we have tried to generate new Series 4 candidates. In short, we provide two tables:

• A list of >100k molecules obtained with a generative model: download 100k
• A relatively diverse selection of 1k molecules: download 1k

For a first assessment of the results, you can check this dynamic visualization of the selected 1k candidates. If a cluster is of particular interest, please refer to the full results to discover other similar molecules. You can also check a tree map of all molecules.

Our generative model approach is based on Reinvent 2.0. We have implemented several reinforcement-learning agents, aimed at optimizing activity and other desirable properties. This GitHub Repository contains more detailed information and source code.

This is the first time we run a generative model, so please bear with us. We will be more than happy to optimize further runs based on your feedback.

Thanks!
@GemmaTuron @miquelduranfrigola

[edwintse]

Hi @miquelduranfrigola, thanks for all these great suggestions!

• Ideally we would be looking for suggestions that are a bit different from our existing compounds (e.g. we're less interested in single-point changes like adding halogens). Would it be possible to optimise to narrow down the list further?
• We were also wondering if, in the dynamic visualisation page, you were able to add a sliding filter for LogP? That would be helpful for us as a guide for solubility.

[miquelduranfrigola]

Hi @edwintse thanks for the feedback.

To answer your points

• Yes, it is possible to further narrow down the list. We have put here the results of several batches/runs of small molecule generation, some of which, as you have noticed, were focused on a very narrow chemical space (e.g. batch 1). In subsequent batches we forced some diversity by penalizing molecules that looked too similar to existing compounds. This generated a decent chemical diversity in batches >3. In the tree map you will see a field called "Tanimoto" that precisely measures this feature (the similarity to existing series 4 molecules). The dynamic representation (1k compounds) that we provide is just a cluster-based selection. We are happy to do alternative ones and come up with a richer (and shorter :) ) list of candidates.
• Meanwhile, we have added a sliding filter for sLogP in the dynamic visualisation page.

Suggested way forward

We will:

• Do 1 or 2 extra batches based on new constraints (see below)
• Do a better selection (100 candidates? 10?)

Now the most important is that we define the constraints for the generative model. These 3 are clear:

1. Activity
2. Chemical diversity / distance to known series 4 compounds
3. SLogP

Anything else? For example:

• Synthetic or retrosynthetic accessibility
• Molecular weight
• Max number of rings
• Max/min number of heteroatoms
• Max number of halogens
• Number of rotatable bonds
• Preferred substituents
• Forbidden/undesired substituents

It will be very useful to know precisely your ideal profile of properties. Please let us know and we will try to implement it.

[mattodd]

Hi @miquelduranfrigola. Very interesting, and a great way to visualise the suggestions.

I'd definitely agree with you re the top three filters. For logP we tend to want to focus on compounds <3.5, very roughly.
Synthetic accessibility is a "nice to have". We can of course do this as humans, but it can take a while with so many suggestions, and as we mention in the preprint about the previous competitions, we had to ditch some interesting possibilities since they would have taken too long to do. So that's a soft "yes" as a feature.
Molecular weight would always generally be below 500, but that's just typical.
For the others, they're not so important, since we can always try to engineer out problematic motifs (like aromatic amines) if they are found to be potent.

To consider the shortlist, though, it would seem that, with those constraints applied, we (me, @edwintse, other potential contributors) need to simply browse through and isolate structures that, well, take our fancy. I mean, there are a lot of possibles here.

Pinging @drc007 purely because I think you'll get a kick out of this.

[drc007]

@miquelduranfrigola, This is really interesting. It would also be useful to have a measure of the "confidence" in the predicted activity model.

Can you also identify which new molecules would add the greatest amount of new information to your predicted activity model?

[edwintse]

Hi @miquelduranfrigola, if you're able to incorporate all the filters @mattodd mentioned that'd be great. I suppose we'd be looking at getting the list down to 100 compounds, then we can quickly look through and pick a couple out to make.

Just on the tree map visualisation tool, we were just curious about how the compounds were placed throughout the tree and if there was anything particularly significant about the different red clusters? I guess it's a bit hard to pick out certain compounds within the clusters or to know which are the "best".

[miquelduranfrigola]

Thanks, @mattodd @drc007 and @edwintse for your comments! This will help moving forward. In the following days, @GemmaTuron and I will give it a push.

We will try to:

• Identify 100 candidates (following your filtering preferences).
• Add a confidence score to the activity prediction.

In addition, I will provide a deeper explanation of the TMAP. I guess we will select the top 100 candidates based on "red regions", so hopefully this will address @edwintse's good point about how to navigate this map.

As for @drc007's suggestion to identify what molecules would add more information to future models... very interesting, didn't think of this! I don't have an immediate answer, but we will try to address the point. Perhaps, to start with, we could see what molecules would expand more efficiently the applicability domain?

[edwintse]

Hi @miquelduranfrigola, we were just wondering whether it's possible to add a substructure search function to the dynamic visualisation tool? I guess we'd want to be focusing on structures that have meta or para substituents on the RHS phenyl ring as the more interesting ones to pursue.

[GemmaTuron]

hello @edwintse, I was actually just having a closer look at the most desirable substituents according to the information on the wiki and series 4 paper. We are trying to refine the molecule generator these days, it would be great if you can give us some hints about the most desirable substituents, also taking into account what you have observed in terms of HLM and RLM.
As for the display, can try to play a bit using SMARTS structures and see if we can incorporate this in the visualization tool. Will let you know if it works.

[GemmaTuron]

Hello @edwintse, sorry for the delay! I have updated the app visualization to provide some substructure search capabilities. As you mentioned you are interested in the RHS substituent I have added the following select-boxes:
Heteroaryl: when selected, displays all RHS substituents composed of an aryl (including phenyl)
Phenyl: when selected, displays RHS substituents containing strictly phenyls (no heteroatoms)
Para/Meta/Orto options allow to select compounds with para-, meta- or orto- substituents on the phenyl.
Let me know if this is useful or you were thinking of different filters.

[edwintse]

@GemmaTuron Wow, that's amazing and super useful to narrow things down!