Project Summary Alzheimer’s disease (AD) lacks effective treatment, primarily due to our limited scientific understanding of the early cellular pathways leading to end-stage pathologies like amyloid-β (Aβ) and tau. To bridge this knowledge gap, we propose significant advancements in XAI that can expedite data-driven discoveries in AD. The growing availability of multimodal single-cell data from donor cohorts, covering the entire spectrum of AD pathology, underscores the need for machine learning (ML) models to learn low-dimensional embeddings, a crucial step in interpreting these datasets. Current XAI technology addresses the opacity of supervised ML models by calculating feature attributions, indicating the importance of individual features like gene expression levels to the model’s output, such as a Aβ level, for specific samples such as cells. However, a significant gap exists between model explanations and biological insights: (1) Current XAI techniques face limitations in the context of unsupervised embedding learning and the integration of prior knowledge. (2) The computation of accurate feature attributions poses challenges due to its exponential complexity. (3) Validating hypothesized causal factors requires interventional experiment. IDEAL-XAI addresses these limitations by focusing on the following objectives: Aim 1: Generate biologically informed explanations of AD progression. To bridge the gap between gene-level attributions and systems-level explanations, our innovative XAI methods operate within a unified latent embedding space, incorporating diverse biological concepts from prior knowledge. These methods interpret patterns within the embedding space, including disease progression, to identify potential driver genes. Aim 2: Compute accurate feature attributions to identify putative AD drivers. We propose theoretically grounded techniques to rigorously compute Shapley values for modern large deep models such as transformers and graph neural networks, handle feature correlations and multimodality, and evaluate feature attribution methods in a principled manner to assist investigators discern the most effective techniques for their applications. Aim 3: Validate computational hypotheses in human neurons and microglia. To validate our computational hypotheses regarding AD drivers, we will perform experiments in human neurons and microglia, two critical cell types in AD, across various stages. By modulating the expression of potential drivers in these cells, we aim to measure their impact on various AD-related phenotypes, providing essential insights into therapeutic targets. The successful IDEAL-XAI project will significantly advance XAI principles and techniques, with broader applications in biomedical research areas. Moreover, it will enhance our understanding of AD and expedite the discovery of potential therapeutic targets. IDEAL-XAI represents a crucial step toward unraveling the mysteries of AD and translating complex ML models...