Speaker: Mr. Philippe SCHWALLER
Institution: Researcher, IBM Research, Zurich
Hosted by: Professor Haibin SU
Zoom Link: https://hkust.zoom.us/j/93043529088?pwd=T3VHQmhIUGViM0dHemJlb2RrTGphZz09
In organic chemistry, we are currently witnessing a rise in artificial intelligence (AI) approaches, which show great potential for improving molecular design, facilitating synthesis and accelerating the discovery of novel molecules. Based on an analogy between written language and organic chemistry, I built linguistics-inspired neural network models for chemical reaction prediction [1, 2] and synthesis planning . Those advances led to the developments of IBM RXN for Chemistry  and RoboRXN , the first AI-driven, cloud-connected, and automated synthesis platform.
To make my chemical language models more explainable, I studied methods for chemical reaction classification and fingerprints. By finding a mapping from discrete reactions to continuous vectors, I enabled efficient chemical reaction space exploration . Moreover, I specialised similar models for yield predictions and achieved state-of-the-art performance, also in the low-data regime [7, 8].
Intrigued by the remarkable performance of chemical language models, I visually inspected what the models learned during training on vast sets of unlabelled chemical reactions collections. Most strikingly, I discovered that my models had captured how atoms rearrange during a reaction, without supervision or human labelling. This atom rearrangement is also called atom-mapping. Knowing the atom-mapping makes chemical reactions better machine-accessible and opens up the possibility of automatically extracting reaction centres and rules. Using the atom-mapping signal from my models, I developed the open-source RXNMapper , which an independent group recently benchmarked as the best atom-mapping tool - even better than commercially available ones . My work on atom-mapping demonstrates the first extraction of organic chemistry grammar from unsupervised learning of chemical reactions and provides the missing link between data-driven and rule-based chemical reaction approaches. My models have learned the language of organic chemistry - atom by atom.
 P Schwaller, T Laino, T Gaudin, P Bolgar, CA Hunter, C Bekas, AA Lee. Molecular Transformer: A Model for Uncertainty-Calibrated Chemical Reaction Prediction. ACS Cent. Sci. 2019, 5, 9, 1572–1583. (https://doi.org/10.1021/acscentsci.9b00576)
 G Pesciullesi•, P Schwaller•, T Laino, JL Reymond. Transfer learning enables the molecular transformer to predict regio- and stereoselective reactions on carbohydrates. Nat. Commun., 2020, 11, 4874. (https://www.nature.com/articles/s41467-020-18671-7)
 P Schwaller, R Petraglia, V Zullo, V H Nair, R A Haeuselmann, R Pisoni, C Bekas, A Iuliano, T Laino. Predicting retrosynthetic pathways using a combined linguistic model and hyper-graph exploration strategy. Chem. Sci., 2020,11, 3316-3325.
 IBM RXN for Chemistry platform (https://rxn.res.ibm.com)
 RoboRXN (https://rxn.res.ibm.com/rxn/robo-rxn/)
 P Schwaller, D Probst, AC Vaucher, VH Nair, D Kreutter, T Laino, JL Reymond. Mapping the Space of Chemical Reactions using Attention-Based Neural Net- works. Nat. Mach. Int., 2021 (https://www.nature.com/articles/s42256-020-00284-w)
 P SchwallerAC Vaucher, T Laino, JL Reymond. Prediction of Chemical Reaction Yields using Deep Learning. Mach. Learn.: Sci. Technol.In press. Preprint: (https://doi.org/10.26434/chemrxiv.12758474)
 P Schwaller, AC Vaucher, T Laino, JL Reymond. Data augmentation strategies to improve reaction yield predictions and estimate uncertainty. 2020 NeurIPS Workshop on Machine Learning for Molecules. (https://doi.org/10.26434/chemrxiv.13286741)
 P Schwaller, B Hoover, JL Reymond, H Strobelt, T Laino. Extraction of organic chemistry grammar from unsupervised learning of chemical reactions. Sci. Adv. In press. Preprint: (https://doi.org/10.26434/chemrxiv.12298559)
 A Lin, N Dyubankova, T Madzhidov, R Nugmanov, A Rakhimbekova, Z Ibragimova, T Akhmetshin, TR Gimadiev, R Suleymanov, J Verhoeven, JK Wegner, H Ceulemans, A Varnek. Atom-to-atom mapping: a benchmarking study of popular mapping algorithms and consensus strategies. (https://doi.org/10.26434/chemrxiv.13012679)
About the speaker
Mr. Philippe Schwaller is currently a PhD student in Chemistry and Molecular Sciences in the Reymond Group at the University of Bern and a researcher at IBM Research - Europe in Zurich. He received a bachelor’s and master’s degree in Materials Science and Engineering from EPFL (Lausanne, Switzerland) and an MPhil degree in Physics from the University of Cambridge (UK). Since he joined IBM Research in March 2017, his main focus is on machine learning for accelerating the synthesis and discovery of novel molecules and materials.