Lecture
Generating Ground-Truth for Cell State Inference by Pairing Molecular and Imaging Data at the Single-Cell Level
- at -
- ICM Saal 2
- Type: Lecture
Lecture description
J. Bues*, J. Pezoldt*, C. Lambert*, B. Deplancke*, et al.
*Laboratory of Systems Biology (LSBG), École polytechnique fédérale de Lausanne (EPFL)
A major paradigm shift in modern bioanalysis is the use of Artificial Intelligence (AI) to infer cellular identity and function directly from morphology. However, a critical bottleneck in training AI models is the scarcity of ground-truth datasets where high-resolution imagery is directly paired with high-dimensional molecular signatures.
To resolve this, we have engineered a fully automated framework that combines highcontent imaging with the molecular precision of single-cell RNA-sequencing per individual cell. Our platform, IRIS (Interconnected Robotic Imaging and Single cell transcriptomics), utilizes machine-vision and microfluidics to deterministically position cells for multimode imaging. By incorporating focal stacking, IRIS acquires highresolution morphological data of every processed cell. Subsequent molecular capture in droplets yields single cell transcriptomics data that approach state-of-the-art gene detection sensitivity, directly linking subcellular features to high-quality gene expression profiles at a scalable throughput of 1,000 cells/h.
We validated IRIS by probing the fundamental process of the cell cycle. By matching canonical morphological phases with their respective transcriptional programs on a per cell basis, we demonstrate that IRIS provides high-fidelity training data necessary for teaching AI models to uncover how general cellular processes manifest across different modalities. Furthermore, we show its discovery power by molecularly resolving two distinct nuclear-endoplasmic reticulum (ER) architectures within naïve CD8+ T cells, which were found to be driven by gene expression programs, translating into distinct activation potential.
Thus, IRIS establishes a single-cell phenomics framework for AI-driven biology, shifting the paradigm from extracting descriptive features to deriving molecular insights directly from cellular morphology. Our technology is designed with the goal of predicting molecular processes in live cells, enabling the non-invasive identification and sorting of cells with desirable properties or the quantitative measurement of drug responses at scale. IRIS therefore provides the data infrastructure required for the AI-native biotechnology era, closing the long-standing gap between cellular form and molecular function.
*Laboratory of Systems Biology (LSBG), École polytechnique fédérale de Lausanne (EPFL)
A major paradigm shift in modern bioanalysis is the use of Artificial Intelligence (AI) to infer cellular identity and function directly from morphology. However, a critical bottleneck in training AI models is the scarcity of ground-truth datasets where high-resolution imagery is directly paired with high-dimensional molecular signatures.
To resolve this, we have engineered a fully automated framework that combines highcontent imaging with the molecular precision of single-cell RNA-sequencing per individual cell. Our platform, IRIS (Interconnected Robotic Imaging and Single cell transcriptomics), utilizes machine-vision and microfluidics to deterministically position cells for multimode imaging. By incorporating focal stacking, IRIS acquires highresolution morphological data of every processed cell. Subsequent molecular capture in droplets yields single cell transcriptomics data that approach state-of-the-art gene detection sensitivity, directly linking subcellular features to high-quality gene expression profiles at a scalable throughput of 1,000 cells/h.
We validated IRIS by probing the fundamental process of the cell cycle. By matching canonical morphological phases with their respective transcriptional programs on a per cell basis, we demonstrate that IRIS provides high-fidelity training data necessary for teaching AI models to uncover how general cellular processes manifest across different modalities. Furthermore, we show its discovery power by molecularly resolving two distinct nuclear-endoplasmic reticulum (ER) architectures within naïve CD8+ T cells, which were found to be driven by gene expression programs, translating into distinct activation potential.
Thus, IRIS establishes a single-cell phenomics framework for AI-driven biology, shifting the paradigm from extracting descriptive features to deriving molecular insights directly from cellular morphology. Our technology is designed with the goal of predicting molecular processes in live cells, enabling the non-invasive identification and sorting of cells with desirable properties or the quantitative measurement of drug responses at scale. IRIS therefore provides the data infrastructure required for the AI-native biotechnology era, closing the long-standing gap between cellular form and molecular function.
