Question about Custom Alerts scoring in Reinvent4

[bishalbudha]

Hi everyone,

I’m running reinforcement learning in Reinvent4 and using Custom Alerts as part of my scoring function. As I understand, Custom Alerts work like a global filter and apply a binary penalty (0 or 1) depending on whether a generated SMILES contains an unwanted SMARTS pattern.

However, when I look at the training logs/plots (see attached images), I see a continuous score curve for Unwanted SMARTS (raw) and Unwanted SMARTS. For example, the graph shows values like 0.94 or 0.98.

👉 My question is:

• Does a value like 0.94 mean that 94% of the generated SMILES in that batch contain at least one Custom Alert SMARTS?

• Or is it representing something else (e.g., an averaged penalty, smoothed loss, or probability)?

If anyone has deeper insight into how Reinvent4 computes and logs the Custom Alert score in RL (especially the difference between the “raw” and “smoothed” plots), I’d really appreciate your explanation.

Thanks in advance!

[halx]

What you see in the TB output is the average over the batch so, yes, it is the proportion of molecules not having the unwanted pattern. The final score (vs the raw score) also takes into account the validity of the SMILES; basically multiplying the transformed score with the valid mask.

[bishalbudha]

Thank you so much!