Tweets

2022-10-05T16:04:16+00:00 | 🔗

sort of MoleculeNet work, but it's been a mess trying to figure out (A) which split everyone else uses and (B) which split is actually the most informative. A larger endeavor likely needs to be made there. Super excited though, to see what the team accomplished on TDC! 6/n

2022-10-05T16:04:16+00:00 | 🔗

good results quickly, and be confident that the predictors we make are the best they can be! I'll probably make a thread at some point analyzing these models more. And when we have time we'll (@OlorenAI) make an effort at submitting to the rest of the benchmarks + some 5/n

2022-10-05T16:04:16+00:00 | 🔗

and copy and paste the desired model parameter dictionaries in: oce.create_BC(model_parameter_dictionary). create_BC stands for create BaseClass. When we work on a new QSAR project, we start by trying all the models we benchmarked + a few special ones :) so that we can get 4/n

2022-10-05T16:04:15+00:00 | 🔗

CYP2D6_Veith (CYP P450 2D6 Inhibition) With the others at comparable levels to prior submissions. Full code to replicate these results: https://t.co/zzA8bLx4X1 And to use these models in your own projects, just install and import @OlorenChemEng 3/n

2022-10-05T16:04:15+00:00 | 🔗

Pgp_Broccatelli (Pgp (P-glycoprotein) Inhibition) Clearance_Hepatocyte (Human Hepatocyte intrinsic clearance) Clearance_Microsome (Human Microsome intrinsic clearance) AMES (Ames Mutagenicity) Half_Life_Obach (Half life) CYP3A4_Veith (CYP P450 3A4 Inhibition) 2/n

2022-10-05T16:04:15+00:00 | 🔗

Oloren (Andrew Li) new set of results published on @ProjectTDC! It's remarkable how far gradient-boosting+stacking diverse predictors together gives performance gains across the board!!! Best-on-leaderboard performance on: HIA_Hou (Human Intestinal Absorption) 1/n

2022-10-05T14:37:26+00:00 | 🔗

Lots of people with AI drug discovery companies charge wayyyy too much for not enough https://t.co/aLDXzUvrMs

2022-10-04T15:59:11+00:00 | 🔗

This is all possible using hyperopt, and a few neat Python tricks of abstraction to make everything mesh nicely together in a coherent framework in Oloren ChemEngine. P.S. Our goal is to have all molecular property prediction methods in OCE so let us know if we are missing smth https://t.co/pkFPnvBvD8

2022-10-04T15:59:10+00:00 | 🔗

Some scripts for hyperparameter optimization of various graph neural networks (ChemProp, GINEPLUS, pre-trained models from Strategies for Pre-training Graph Neural Networks)... fairly comprehensive actually, and all in a few lines of Python! https://t.co/5YnYfUgqtp

2022-10-01T04:58:49+00:00 | 🔗

and more than that should there be a specific set of graph edit operations allowed, perhaps something grounded in a chemical rationale? The obstacle I see is that this might make edit paths needlessly complicated. Anyways, related references + suggestions much appreciated! n/n