Read on Twitter
1/ One of the final pieces of my PhD is out in @NatMachIntell. Let’s explore chemical reaction space using language models!
Article: https://rdcu.be/cenmd
Blog: https://ibm.co/3t0kzPw
Code: https://github.com/rxn4chemistry/rxnfp/tree/master/
#ChemTwitter #CompChem
@IBMResearch @reymondgroup @DCBunibern
Article: https://rdcu.be/cenmd
Blog: https://ibm.co/3t0kzPw
Code: https://github.com/rxn4chemistry/rxnfp/tree/master/
#ChemTwitter #CompChem
@IBMResearch @reymondgroup @DCBunibern
2/ This work is an exciting collaboration between @IBMResearch and @unibern with outstanding contributions from Daniel ( @skepteis), Alain ( @acvaucher), Vishnu, David ( @DavidKreutter), Teo ( @teodorolaino) and Jean-Louis ( @jrjrjlr).
3/ We trained classification models (Seq2Seq Transformers and BERT) on chemical reactions (test accuracy 98.2%).
Using the encoder, we developed data-driven reaction fingerprints, enabling similarity searches and almost perfect clustering of chemical reaction space. -> rxnfp
Using the encoder, we developed data-driven reaction fingerprints, enabling similarity searches and almost perfect clustering of chemical reaction space. -> rxnfp
4/ An interactive reaction atlas made using our fingerprints and #TMAP is available on: https://rxn4chemistry.github.io/rxnfp//tmaps/tmap_ft_10k.html
5/ You want to try it out? Check out the GitHub repo https://github.com/rxn4chemistry/rxnfp/tree/master/ or install it directly with `pip install rxnfp`, assuming you have already installed RDKit. What would we do without @RDKit_org :)
Tutorials can be found on: https://rxn4chemistry.github.io/rxnfp/
Tutorials can be found on: https://rxn4chemistry.github.io/rxnfp/
6/ Jorner ( @kjelljorner) et al. successfully applied our reaction fingerprints to predict activation energies of nucleophilic aromatic substitution reactions. Awesome work!
( https://pubs.rsc.org/en/content/articlehtml/2021/sc/d0sc04896h)
( https://pubs.rsc.org/en/content/articlehtml/2021/sc/d0sc04896h)
7/ The encoder (fingerprint) model can easily be fine-tuned on other chemical reaction tasks.
For example, we used it to learn to predict chemical reaction yields.
-> https://chemrxiv.org/articles/preprint/Prediction_of_Chemical_Reaction_Yields_using_Deep_Learning/12758474
For example, we used it to learn to predict chemical reaction yields.
-> https://chemrxiv.org/articles/preprint/Prediction_of_Chemical_Reaction_Yields_using_Deep_Learning/12758474
8/ As the encoder has already learned the reaction SMILES grammar, it can be efficiently fine-tuned even in the low-data regime. -> https://chemrxiv.org/articles/preprint/Data_augmentation_strategies_to_improve_reaction_yield_predictions_and_estimate_uncertainty/13286741
