Speaker
Description
Providing clear explanations is crucial in interdisciplinary research fields like bioinformatics where non-experts in machine learning (ML) must understand model decisions to foster trust in the system. Interactive visualisation can help in enabling the active exploration of model behaviour. In this paper,we present an approach to interpreting compound potency predictions by using RuleSense, a visual analytics system that integrates rule-based modeling, topic modeling, and interactive visual tools to support deeper interpretation of predictive models. RuleSense groups related rules into coherent topics, identifies key feature patterns, and allows users to trace how these patterns contribute to predictions. This helps to reduce the complexity of the model and helps to gain understanding on the most meaningful parts of a model. We demonstrate the approach in the context of compound potency prediction—a critical task in drug discovery—showing how RuleSense reveals relationships between chemical structure and activity, supports hypothesis generation, and bridges the gap from prediction to scientific insight.
