Training REINVENT Prior on Large Cyclic Peptides (≈1200 Da)

[meshoh52-design]

Hello REINVENT team,

I’m currently trying to train the reinvent.prior on a dataset of a relatively large cyclic peptides (~1200 Da, 80-100 heavy atoms) essentially polymyxin-like macrocycles.
These molecules are fully parsable and valid in RDKit (they pass sanitization and are handled correctly by my external activity scorer), but they are all being rejected by the default input filter during transfer learning with messages like:

"default" filter: CC(C)C[C@@h]1NC(=O)C@@HNC(=O)C@HNC(=O)C@@HCCNC(=O)C@HNC(=O)C@HNC(=O)C@HNC1=O is invalid

I believe this happens because the built-in RDKitFilter has a hard-coded max_heavy_atoms limit, so my macrocyclic peptides fall slightly outside the training domain of the standard prior and get dropped before training starts.

What I’d like to ask Is there a way to turn off the default molecule filter completely or adjust its thresholds and is reinvent a suitable prior for that or should i switch to a different prior ?

Thanks for your time and for maintaining such a great generative framework

[halx]

Hi,

You can switch off the standardization step with

[parameters]

standardize_smiles = false

Cheers,
Hannes.