Presentation: Rule-based ChEBI extension

[sfluegel05]

Context

Last year, we have presented ChemLog, a framework for rule-based classification of molecules into ChEBI classes. We had successfully applied ChemLog to the peptide-branch of ChEBI by manually developing rules in tandem with domain experts.

Expected outcome of the workshop

This talk proposes 2 paths for extending rule-based extension to the rest of ChEBI: (a) Inductive Logic Programming (ILP), a learning framework that can infer first-order logic rules from samples, and (b) ChEBI2FOL, an LLM-based methodology for formalising natural language definitions in ChEBI. Both approaches have different strengths: ILP works directly on the existing classification in ChEBI and needs less data than Deep Learning methods. ChEBI2FOL exploits chemical background knowledge and iteratively refines its definition based on feedback. We will discuss how to best represent chemical structures for both applications (including functional groups, ring structures and targeted chemical properties) and show preliminary results.
Our goal is to integrate these new rules with ChemLog and Deep Learning methods as part of the Chebifier ensemble.

Prerequisites issues/tasks

Implementation issues/tasks

~20 minutes for the presentation and 10 minutes for the subsequent discussion.

[StroemPhi]

Thank you for this proposal @sfluegel05! We'd love to have you present this.

[StroemPhi]

I'm wondering if this might be helpful in the context of #6 where we will probably need to use things like functional groups and ring structures within the axioms of the reaction ontology (e.g. RXNO/MOP/OntoKin) to infer things.