Interpretable individuaL Latent neUral Mapping for post-hoc Explanations of high-dimensional tabular data applied to science

Objective

Explainable Artificial Intelligence (XAI) is essential for deploying trustworthy AI systems, especially in critical fields where transparency is key. Our XAI ERC project (2019-2025) has focused on developing post-hoc “explainers” for black-box machine learning models. A key outcome of this project is ILLUME, a meta-explainer that generates a complex global surrogate model to mimic a black-box system. When applied to individual inputs, this model becomes interpretable and can produce multiple explanation formats, including feature relevance, decision rules, and counterfactuals, for tabular data. Compared to competitors, ILLUME has shown superior accuracy, robustness, and quality, establishing itself as a breakthrough post-hoc XAI approach. Our current proposal aims to consolidate ILLUME's capabilities in two main intertwined directions: i) Demonstrate Scientific Impact: We will prove ILLUME's effectiveness as an “XAI4Science” tool by applying it to complex biomedical problems, specifically in pharmacology and cancer research. By collaborating with expert biomedical scientists, we will show how ILLUME can provide valuable insights and new hypotheses where explanation is often more crucial than prediction. We will also implement training materials to enable scientists to use ILLUME independently. ii) Enhance Usability: We will transform ILLUME into a widely accessible toolset with a smooth user interface, expanding its support to high-dimensional data like genetic and proteomic data, as well as new modalities such as time series and images. This will include adapting the conversational user interface developed in the XAI project to work with ILLUME, broadening its practical use. By combining advanced explanation capabilities, support for diverse data types, and an accessible interface, the enhanced ILLUME toolset is poised to become a unified, scalable framework for post-hoc explainability with the potential to set a new standard for transparent and human-centered AI.