Modèles génératifs profonds pour la détection de phénotypes subtils dans des images de microscopie

Detecting cellular phenotypic variations in biological microscopy images is instrumental in advancing fundamental research and drug discovery. However, this task remains challenging in practical scenarios due to inherent biological variability, data scarcity, and the need for interpretability. In this work, we develop deep learning-based strategies to effectively detect and interpret these phenotypic variations. First, we demonstrate that deep discriminative models, combined with classical interpretability methods, can successfully identify discriminative regions within microscopy images. Nevertheless, these discriminative models alone often fall short in managing biological variability. To address this limitation, we explore Generative Adversarial Networks (GANs) and illustrate how they can effectively reduce biological variability while preserving meaningful phenotypic differences. Additionally, we show that self-supervised learning techniques can mitigate the requirement for large annotated datasets, enhancing model performance in data-constrained environments. Furthermore, we investigate efficient conditioning mechanisms for GANs, enabling controlled generation of biologically relevant phenotypes. We leverage conditional GANs specifically to uncover distinct phenotypes within microscopy images. Building upon this foundation, we propose a novel framework based on diffusion models to detect subtle phenotypic variations. Our experiments demonstrate that diffusion models exhibit robustness in identifying both apparent and imperceptible phenotypic changes. Importantly, we show how pretrained diffusion models, originally trained on extensive natural image datasets, can be adapted effectively to biological domains. This transfer learning strategy is particularly valuable in contexts where biological data is scarce and challenging to obtain. Finally, we introduce Diffex, a unified framework that bridges generative and discriminative models. Diffex allows the explanation of classifier decisions on microscopy images by identifying disentangled directions within diffusion model latent spaces, thus providing interpretable insights into model predictions. Overall, our work offers significant advancements in biological interpretability, unlocking new insights and accelerating more efficient drug discovery.