High-Dimensional Pharmacology

Kaggle Solubility competition is over. It did draw the attention of 100 participants with some strong issues underlined by the participants. I decided to share my dive into this... Data: 70 711 compounds in train set: A very good and diverse dataset.   30 308 compounds in test set . Quality m etric: Kappa coefficient  (was it changed later?). Results Initial winner (by leaderboard) was a Ohue laboratory of Tokyo Institute of Technology with a combination of the BART transformer language model and GNN with focal loss  ( solution ).  The dramatic difference between the winner and the rest of the participants was noticeably big and raised a lot of questions. Bernhard Rohde (Novartis) was the second leaderboard winner. He underlined the same issue and showed that “plate/row/column” has a drastic effect on model quality (explicit IDs were provided: EOS12286). Still, the Bayesian Instance-Based Offset Learning solution he used draw my attention, because it can be app...

 T his is a first of the “lets-test-this” posts: Google Colab notebook is here . TLDR : Jazzy features does not have positive impact on solubility prediction. Intro: hydrogen bonds drive all kind of interactions, and no doubt are important for drug design process. A nice package from AstraZeneca was published recently that predict of hydrogen-bond strengths and free energies of hydration of small molecules. It looks like that the main application of Jazzy is for local settings – protein-ligand strength estimation, SAR studies, but my interest was to test it on compound solubility prediction. Goal : testing if free energy of hydration (without free parameters) from Jazzy has an impact on prediction of compound solubility. Setup: Data was taken from SolCuration, that's a derivate from Meng et al paper (AqSol) + additional curated datasets. In case of duplicated measurements, a median was taken and compounds with SD > 0.5 were removed (~1500 drop). Three types of descriptor...