Presentation: Explainability of chemical language models

[schnamo]

Context

Why this topic matters
Chemical language models are being widely adopted in research, yet they are often still considered a "black box". While interpretability and explainability approaches have been developed for machine learning models, we are still missing innovative approaches to explain the predictions of chemical language models at the class level and to connect such explanations with prior knowledge ontologies.

Relevant tools, standards, or datasets
We are currently developing a novel method to find explanations for ChEBI ontology class classification predictions of chemical language models that can find statistically significant patterns across molecules that are associated with specific chemical classes.

Expected outcome of the workshop

To showcase our novel method and discuss the results as well their usefulness with audience members. We would also like to gather feedback on what the community would benefit from and if we could achieve that with our method.

Prerequisites issues/tasks

n/a

Implementation issues/tasks

How much time do you need for your talk and how much for the subsequent discussion?
Ca. 20 minutes talk and 10 minutes discussion

[StroemPhi]

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