Thread by @pschwllr

Taking chemical reaction prediction models one step further in a great collaboration with Giorgio (@Giorgio_P_), a brilliant organic chemist!

A thread ⬇️1/N

A major limitation of current deep learning reaction prediction models is stereochemistry. It is not taken into account by graph-neural networks and a weakness of text-based prediction models, like the Molecular Transformer (doi.org/10.1021/acscen…).
How can we improve? 2/N

In this work, we take carbohydrate reactions as an example. Compared to the reactions in patents (avg. 0.4 stereocentres in product), carbohydrate contain multiple stereocentres (avg. >6 in our test set), which make reactivity predictions challenging even for human experts. 3/N

Another difficulty is the availability of good quality data. Typically, there is not enough data available to train a reaction prediction model solely on one reaction class. Inspired by #NLProc (e.g. the work @seb_ruder), we explored different transfer learning techniques. 4/N

With transfer learning, we can leverage the knowledge extracted from large general reaction data sets (e.g open-source USPTO @dan2097 @nmsoftware), to train better models for the prediction of specific complex reaction (here, carbo reactions). 5/N

We explore different settings: Multi-task, where we train on the generic and specific data sets simultaneously, and sequential transfer learning, where a model trained on the generic data is adapted to the specific data in a subsequent training run. 6/N

While the first scenario ensures good performance not only on the specific but also on the generic data, the second scenario is particularly interesting because of the reduced computational cost and the fact, that the generic (potentially proprietary) data is not disclosed. 7/N

We evaluated our models in numerous ways:
- random and time-split test sets of carbohydrate reactions
- recent @J_A_C_S total syntheses
- an in-house 14-step synthesis of a lipid-linked oligosaccharide (LLO) by @Giorgio_P_
8/N

The transfer-learned models show intriguing performance across all test sets using only small specific data sets for training.

Moreover, it is the first deep learning reaction prediction work including an experimental validation. 9/N

The methods were implemented with #OpenNMT and are straightforward to adapt for any reaction class of interest. Canicalisation was done using @RDKit_org. Code and trained models are available from github.com/rxn4chemistry/… 10/N

Further explanations can be found in our blog post: chemistrycommunity.nature.com/posts/transfer… and the article 11/N

If you have questions, how to adapt our work to your reaction domain of interest. Feel free to reach out! 12/N

At the same time, this study is my first peer-reviewed work with the @reymondgroup. I’m very grateful to my two supervisors Prof Jean-Louis Reymond (@jrjrjlr) and Teodoro Laino (@teodorolaino)! 13/N

Excited to see our carbohydrate transformer out in @NatureComms, an awesome open-access journal!
Stay tuned, more will follow!
#compchem #glycotime #AI4Chem 14/N