"Machine learning in chemistry and beyond" (ChE-651) seminar by Gabe Gomes "Machine-Guided Catalyst Optimization"
Event details
| Date | 27.09.2022 |
| Hour | 15:30 › 16:30 |
| Speaker | Gabe is an assistant professor at the departments of Chemistry and Chemical Engineering at Carnegie Mellon University, in Pittsburgh, USA. Gabe was born and raised in the countryside of the state of Rio de Janeiro, Brazil, where he received his B.Sc. He earned his Ph.D. in Fall 2018 from Florida State University, where he was awarded the LASER Fellowship in 2014 and the 2016-2017 IBM Ph.D Scholarship. At FSU, Gabe's research was centered on the relationship between molecular structure and reactivity, focusing on the development and applications of stereoelectronic effects. For his work at FSU, in 2018, Gabe received several awards for his work in computational chemistry, including his selection for the CAS SciFinder Future Leaders Program. In 2019, Gabe joined the University of Toronto as a Postdoctoral Research Fellow in the Matter Lab, led by Professor Alán Aspuru-Guzik. In 2020, Gabe was awarded the prestigious NSERC Banting Postdoctoral Fellowship with the project “Designing Catalysts with Artificial Intelligence,” and has been featured on the “Next Great Impossible” series by Merck/Milipore-Sigma. Gabe joined the Journal of Chemical Information and Modeling as an Early Career Board member in 2021. |
| Location | Online |
| Category | Conferences - Seminars |
| Event Language | English |
The ability to forge difficult chemical bonds through catalysis has transformed society on all fronts, from feeding our ever-growing populations to increasing our life-expectancies by synthesizing new drugs. Not only has the rise in popularity of metal-catalyzed cross-coupling reactions enabled us to make existing processes more efficient, but it also has allowed us to synthesize novel and unexplored molecules and materials, unlocking the technologies of the future.
In metal catalysis, the choice of ligand often leads to the most significant impact on the reaction outcomes, such as yield or product selectivity. Identifying optimal metal-ligand combinations can be a laborious experimental process. This practice is often held back by the difficulty of meaningfully comparing results with different ligands. I will introduce our efforts to develop a platform for inverse design of catalysts utilizing high-throughput virtual screening and machine learning (ML), coupled with an extensive ligands database. One of the most used strategies in ML for metal catalysis is the development of models based on reactant parameterization. Given that the space of potential ligands is immense, this tactic can be very challenging.
We have developed a platform aimed at data-driven ligand optimization and their inverse design. At the center of this strategy is Kraken,[1] a database of descriptors for organophosphorus ligands that encompasses features that are most important for catalysis including conformational flexibility and ability for both coordinative and non-covalent bonding.
However, the development of ML solutions for chemical applications cannot be a simple application of “out-of-the-shelf” methods. Molecular representation is at the core of chemical understanding. To date, many ways to represent molecules for machines have been developed: different types of fingerprints, language model-based embeddings, graph representations, some physicochemical descriptors. Still, existing approaches lack quantum-mechanical information about molecular structure. The second part of this talk proposes a solution based on embedding stereoelectronic effects onto molecular graphs with graph neural networks.[2]
We developed an end-to-end pipeline that takes a molecular geometry input and returns a representation with stereoelectronic information. After enriching the standard molecular graphs with lone pairs (predicted by a separate model) and bond-orbital nodes, we developed a multi-task approach that learns all the interaction values provided by Natural Bond Orbital analysis. Finally, we use the resulting representations for downstream tasks, effectively displaying the approach’s vast applicability and increased performances.
References
[1]. T. Gensch, G. Gomes, P. Friederich, E. Peters, T. Gaudin, R. Pollice, K. Jorner, A. Nigam, M. L. D'Addario, M. S. Sigman, A. Aspuru-Guzik "A comprehensive platform for the discovery and understanding of organophosphorus ligands in catalysis", J. Am. Chem. Soc. 2022, 144 (3), 1205
[2]. D. Boiko, T. Reschützegger, B. Sanchez-Lengeling, S. Blau, G. Gomes* “Stereoelectronics-Aware Molecular Representation Learning”, ChemrXiv, 2022, doi: 10.26434/chemrxiv-2022-nz4pc
Practical information
- General public
- Free
Organizer
- Kevin Maik Jablonka, Solène Oberli, Puck van Gerwen, Andres M Bran
Contact
- Kevin Maik Jablonka, Solène Oberli, Puck van Gerwen, Andres M Bran