Reward certain substructure matching

[frank-stonybrook]

Dear Reinvent4 developer,

Thanks for developing such an amazing tool!

I did a quick look at the code as well as the manuscript and am wondering whether it is possible to reward the generation of certain substructures. It seems to me that both CustomAlerts and MatchingSubstructure will penalize the matching of substructures.

Thanks

[halx]

Hi,

many thanks for your interest in REINVENT and welcome to the community!

Yes, those components would be suitable for that as both take SMARTS. Just be aware that they work differently. CustomAlerts is a filter and MatchingSubstructure a penalty which technically runs last and multiplies the total score with 0.5 if the pattern does not match.

Many thanks,
Hannes.

[frank-stonybrook]

Thanks Hannes for the clarification!

Am I understand it correctly:

For CustomAlerts, a design matches any of the substructure SMARTS pattern, it will be excluded immediately for any subsequence scoring.

For MatchingSubstructure, depends on how the transform is setup, it can be either a penalty or reward. For example, the following toml setup will penalize the matching of "CO" by zeroing the score if any CO is found in the design:

[[scoring.component]]
[scoring.component.MatchingSubstructure]
[[scoring.component.MatchingSubstructure.endpoint]]
name = "Matching SMARTs substructure"
weight = 1
params.smarts = "CO"

transform.type = "reverse_sigmoid"
transform.high = 3
transform.low = 1
transform.k = 0.5

while removing the transform section will turn it to a reward.

[halx]

Hi again,

Yes, CustomAlerts is a filter and that is how a filter is supposed to work i.e. avoid unnecessary calculation if you already know that the total score will be zero.

You can transform any component, so also MatchingSubstructure, but a sigmoid may not be that ideal as you would have to find out how the two only values, 1.0 and 0.5, would best reverse map. We will update the value mapping transform to handle this better.

Cheers,
Hannes.

[frank-stonybrook]

Thanks for the detailed explanation!

[halx]

I just realize that you should also be able to use the left step transform with a lower value, say 0.5, to invert a binary input.