Summary: Unsupervised techniques are central to dissecting large volumes of data with minimal user input. Despite their variety, unsupervised approaches are almost invariably built on a similarity or difference metric that allows comparing pairs of molecules. However, this basic idea faces an apparently insurmountable problem: when we want to compare N molecules or conformations, we need to perform O(N2) operations. Hence, this scaling can greatly interfere with our need to model ever increasing molecular sets. We recently proposed an alternative way of building similarity indices that bypasses this problem. Our instant similarity (iSIM) indices are defined over an arbitrary number of objects, and can thus compare N molecules with an unprecedented O(N) scaling. This has paved the way to several applications, including more efficient diversity selection methods, and novel ways to explore and represent large sectors of chemical space. Here we will be discussing one of the most enticing applications based on our indices: the development of highly-efficient clustering algorithms. We will present our novel sampling and clustering methods, that allow tackling very large sections of chemical space.
Ramon Alain Miranda-Quintana majored in Radiochemistry in the Higher Institute of Technologies and Applied Sciences in 2011 and obtained his Ph.D. in Chemistry from the University of Havana. After a research appointment at McMaster University, he won a York Science Fellowship to work in York University as a Postdoctoral Scholar (where he won the 2019 Polanyi Prize in Chemistry). He then joined the Department of Chemistry at the University of Florida as an Assistant Professor in 2020, where he is also a member of the Quantum Theory Project. His research interests include the development of ab initio electronic structure methods to study strongly correlated systems, understanding how charge and spin transfer processes shape chemical reactivity and solvation processes, and developing efficient extended similarity-based tools for data science applications in chemistry and the biomedical sciences. At UF he was won an Oak Ridge Ralph E. Powe Junior Faculty Enhancement Award, the OpenEye Cadence Molecular Sciences Outstanding Junior Faculty Award in Computational Chemistry, and an NIH R35 grant.
Summary: Unsupervised techniques are central to dissecting large volumes of data with minimal user input. Despite their variety, unsupervised approaches are almost invariably built on a similarity or difference metric that allows comparing pairs of molecules. However, this basic idea faces an apparently insurmountable problem: when we want to compare N molecules or conformations, we need to perform O(N2) operations. Hence, this scaling can greatly interfere with our need to model ever increasing molecular sets. We recently proposed an alternative way of building similarity indices that bypasses this problem. Our instant similarity (iSIM) indices are defined over an arbitrary number of objects, and can thus compare N molecules with an unprecedented O(N) scaling. This has paved the way to several applications, including more efficient diversity selection methods, and novel ways to explore and represent large sectors of chemical space. Here we will be discussing one of the most enticing applications based on our indices: the development of highly-efficient clustering algorithms. We will present our novel sampling and clustering methods, that allow tackling very large sections of chemical space.
Ramon Alain Miranda-Quintana majored in Radiochemistry in the Higher Institute of Technologies and Applied Sciences in 2011 and obtained his Ph.D. in Chemistry from the University of Havana. After a research appointment at McMaster University, he won a York Science Fellowship to work in York University as a Postdoctoral Scholar (where he won the 2019 Polanyi Prize in Chemistry). He then joined the Department of Chemistry at the University of Florida as an Assistant Professor in 2020, where he is also a member of the Quantum Theory Project. His research interests include the development of ab initio electronic structure methods to study strongly correlated systems, understanding how charge and spin transfer processes shape chemical reactivity and solvation processes, and developing efficient extended similarity-based tools for data science applications in chemistry and the biomedical sciences. At UF he was won an Oak Ridge Ralph E. Powe Junior Faculty Enhancement Award, the OpenEye Cadence Molecular Sciences Outstanding Junior Faculty Award in Computational Chemistry, and an NIH R35 grant.
