This is a short description of our latest submission to the OpenADMET + ExpansionRx Blind Challenge.

UCTPrague submission

# OpenADMET + ExpansionRx Blind Challenge

This is a short description of our latest submission to the OpenADMET + ExpansionRx Blind Challenge.

This submission was made on behalf of the UCT Prague cheminformatics group, and this specific
submission was done by @HunzallahX and @dehaenw.

## short description
The approach used is based on ensembles of TabPFN models.
Ensembling is done using a weighted sum of individual sum.
The models are trained on "bespoke features"
- MOE descriptors (Molecular Operating Environment is a commercial package)
- MORDRED Descriptors (2D only)
- RDkit Descriptors
- FCFP
- ECFP
- RDkit Fingerprint
- AtomPairs fingerprint

The TabPFN models are not finetuned, the ensembling is done based on the performance
of each individual endpoint + individual feature type. So far, we did not do any multitask regression
at any stage.

When the challenge ends all code and an extended description of our approach will be made public at:
https://github.com/lich-uct/openADMET-challenge

Performance comments:
- We tried finetuning TabPFN, no increased performance (but no deterioration)
- No difference in models training using random split vs butina
- No difference in non-MOE features calculated from dominant protomers instead of SMILES as provided
- MAE for internal metrics is lower than on leaderboard, but better MAE internal leads to better MAE external