Crippen's logP as a Quantitative Molecular Benchmark for Explainable AI Heatmaps

Explore the use of Crippen's logP as a quantitative molecular benchmark for explainable AI heatmaps in this 54-minute conference talk by Jan Jensen from Valence Labs. Delve into the challenges of quantitatively benchmarking machine learning prediction methods for chemical properties, particularly in regression tasks. Learn how the Crippen logP model serves as an excellent benchmark for atomic attribution and heatmap approaches, especially when ground truth heatmaps are adjusted to reflect molecular representation. Discover examples of how this benchmark can be applied to better understand ML models and determine the most effective techniques for generating XAI heatmaps. The talk covers topics such as heatmaps in image classification, validation methods, comparing XAI methods, and future outlooks in the field.

- Intro
- Heatmaps: An Idea From Image Classification
- Uses of Heatmaps and How We Validate Them
- Crippen logP Atomic Contributions as Heatmap Ground Truth?
- Q+A
- Comparing XAI Methods
- Comparing Heatmaps
- Conclusions/Outlook
- Q+A