Handling stereochemistry in training data

[bishalbudha]

Hello,
I am working on a drug design project using REINVENT4. My active dataset was downloaded from ChEMBL and many compounds have stereochemistry in their canonical SMILES. Since the REINVENT prior does not support stereochemical tokens (@, @@, /, \), I am unsure how to handle this.

Should I:

1. Remove stereochemistry consistently from all SMILES before training, or
2. Keep actives with stereo but strip the stereo information only during preprocessing?

I am concerned that removing stereochemistry may introduce ambiguity (as some enantiomers are inactive). Could you please share what approach is recommended, or what has been done in previous REINVENT workflows?

Thank you very much!

[halx]

For priors that do not support stereochemistry you will have to remove stereochemistry for all inputs and deduplicate. All transformer models support stereochemistry .

[bishalbudha]

Thank you, Halx, for your constant support. I really appreciate it.

However, even after removing stereochemistry before transfer learning, I keep encountering token errors when running the process. These were active compounds extracted from ChEMBL after stereochemistry removal.

Here is the error I get:

File "/usr/local/envs/reinvent4/lib/python3.10/site-packages/reinvent/utils/config\_parse.py", line 204, in validate\_tokens
f"Tokens {invalid\_tokens} in {smiles} are not supported by the model\n"
UnboundLocalError: local variable 'invalid\_tokens' referenced before assignment

Could you please help me with this?

[bishalbudha]

Halx sir, any help in this issue?

[halx]

The error message contains a bug which is fixed now. But you will need to ensure that the tokens in your input SMILES will be supported by the prior you are using.

[bishalbudha]

Hi @halx sir — thanks for the quick fix and guidance.

After updating, my transfer learning run completed. The only unsupported token I found in my ChEMBL actives (after removing stereochemistry) was [n-]. I’ve added a preprocessing step to drop any SMILES containing this token, and the run is now stable.

One question: instead of discovering unsupported tokens only after a failed TL step, is there a programmatic way to query the prior’s tokenizer vocabulary (e.g., from the prior checkpoint/config) so I can pre-filter my dataset for unsupported tokens before training? If there’s an API or a stored vocabulary file I can access, I’d like to integrate that into my data-cleaning pipeline.

Thanks again for your help!

[bishalbudha]

Thank you sir 🙏 — I already found the list of permitted tokens in Supplementary 10 of the official REINVENT4 publication. This solves my problem, and I’ll use that vocabulary to pre-filter my dataset before training. Really appreciate your guidance!