Exploring Chemical Space with Machine Learning
|15:15 › 16:15
|Originally from Ukraine, Ganna (Anya) Gryn’ova received her BS and MSc in chemistry summa cum laude from Oles Honchar Dnipro National University. In 2014 she received a PhD in computational chemistry from Australian National University. Her doctoral thesis gathered a number of awards, including the IUPAC-Solvay International Award for Young Chemists for one of the five most outstanding PhD theses in the general area of the chemical sciences worldwide. Dr. Gryn’ova continued her research career at École Polytechnique Fédérale de Lausanne as a postdoctoral researcher working on in silico modeling of organic semiconductors. In 2016 she won the Marie Skłodowska-Curie Actions individual fellowship and focussed on the non-conventional architectures of single-molecule junctions. In 2019, Dr. Gryn’ova started her independent scientific career leading the junior research group “Computational Carbon Chemistry” (CCC) at the Heidelberg Institute for Theoretical Studies (HITS gGmbH) and Interdisciplinary Center for Scientific Computing (IWR) at Heidelberg University, Germany. The CCC group uses state-of-the-art computational chemistry and data science to explore and exploit diverse functional organic materials for applications in organocatalysis and environmental remediation. In 2021, Anya received the prestigious ERC Starting Grant for her project “PATTERNCHEM: Shape and Topology as Descriptors of Chemical and Physical Properties in Functional Organic Materials”; she is also a principal investigator in the Collaborative Research Centre SFB1249 “N-Heteropolycycles as Functional Materials” and the SIMPLAIX strategic research initiative on bridging scales from molecules to molecular materials by multiscale simulation and machine learning.
|Conferences - Seminars
Chemical (molecular, quantum) machine learning relies on representing molecules in unique and informative ways. Here, we introduce two new representations – a quantum-inspired representation called matrix of orthogonalised atomic orbital coefficients (MAOC) [S. Llenga, G. Gryn’ova, J. Chem. Phys., 2023, 158, 214116], and a fragmentation-based technique called matrix of fragment similarity representation (MFSR) [in prep.]. MAOC is uniquely suitable for representing monatomic, molecular, and periodic systems, and can distinguish compounds with identical compositions and geometries but distinct charges and spin multiplicities. MFSR is a dimensionality reduction and representation technique for mapping and exploring the chemical space based on specific building blocks. Most industrially and biologically relevant macromolecules are formed as a combination of finite building blocks (e.g., all proteins are a combination of just 20 aminoacids), and MFSR can predict their properties in less than a fraction of a second and with the quantum-chemical accuracy. More broadly, MFSR can be applied to any chemical system using either atoms or “ghost” centroids as representative fragments and allows even the most entangled deep learning models to be decodable in a chemically intuitive form.
- General public
- Yannick Calvino Alonso, Andres M Bran, Rebecca Neeser, Philippe Schwaller