Label-Free AI-Classification of Subcellular Organelles Based on Optical Photothermal Infrared Images

Cells maintain homeostasis by dynamically reorganizing their organelles to tune metabolism in response to stress. Fluorescence microscopy maps organelle locations with subcellular resolution but provides limited information on their chemical composition. Infrared (IR) imaging offers a label-free alternative for probing intrinsic molecular vibrations that report on lipids, carbohydrates, and nucleic acids. However, its broader application to subcellular biology has been limited by spatial resolution and hyperspectral data complexity. Here, we combine submicron optical photothermal IR imaging with machine learning to classify subcellular structures in fixed U-2 OS cells. Using fluorescent-labeled organelles as ground truth, we trained and evaluated random forest (RF) classifiers and U-Net convolutional neural networks to identify organelles from IR spectra. The RF model converged rapidly, requiring fewer than 75 spectra per class per cell and fewer than 25 cells, indicating that models trained on small cellular regions can be extended to classify whole-cell images. The resulting classifiers accurately identified multiple organelles, including the endoplasmic reticulum, Golgi apparatus, mitochondria, nucleus, nucleolus, and stress granules. In contrast, classification was unsuccessful for nuclear speckles, actin, and microtubules, suggesting that some structures lack sufficiently distinct IR signatures under these conditions. Classifiers trained in U-2 OS cells generalized to HEK 293 cells, consistent with conserved organelle biochemical composition across cell types. However, the classifiers failed under cellular stress, indicating sensitivity to stress-induced changes in organelle state. Together, these results establish a scalable, label-free strategy for high-resolution mapping of organelle biochemical composition and provide a foundation for subcellular biomarker discovery and disease-state diagnostics.