You are here:
Molecular Representations to Improve Chemical Modeling and Ontology in Relation to Environmental Chemistry Properties & Outcomes
Citation:
Charest, N., C. Lowe, AND A. Williams. Molecular Representations to Improve Chemical Modeling and Ontology in Relation to Environmental Chemistry Properties & Outcomes. SERMACS, Durham, NC, October 25 - 28, 2023.
Impact/Purpose:
N/A
Description:
The success of any effort to model environmental chemistry outcomes, be they fate, transport, transformation, or bioactivity, relies upon suitable cognition and representation of the underlying and causative chemistry. The way in which chemistry is encoded, which can include traditional structure representations such as spectra, SMILES, InChIKeys, or fingerprints, or experimental attributes like spectra, physical-chemical properties, or bioactivity profiles, can significantly impact the results of any modeling effort, be it machine learning-based or otherwise. Concrete examples of chemical encodings include the development of representations for both targeted (endpoint-specific) and untargeted (endpoint non-specific) similarity comparisons such as those found in read-across workflows and clustering methodologies. The choice of these representations significantly impacts the performance, interpretability, and capabilities of models built upon them for predictive or ontological purposes. This symposium invites direct scrutiny of the advantages, disadvantages, and methodologies associated with the various means of representing chemical structure pertaining to modeling and organizing chemical data. Topics can include modern approaches to chemical characterization such as string encoding, fingerprinting, graph representation, analytical spectra, substructure characterization, or novel basis sets. Emphasis is placed on the interpretability of these representations, and how they can help chemists harness modern cheminformatics and computational approaches while maintaining intuitive accessibility in the resulting models. We are also interested in new ontological schemas for organizing chemical classes, which may emerge from unsupervised machine learning or expert-driven efforts to better categorize chemicals of contemporary interest. Such chemicals would include pharmaceuticals, known environmental contaminants, and chemical families associated with high potential health hazard. A successful symposium would inspire a conversation on the benefits of higher model performance against the benefits of mechanistic interpretation and where middle ground between the two goals may exist.