Cell morphology reveals the structure and function of cells, and observing it can help us understand the mechanisms that drive and prevent disease. Advances in microscopy allow us to observe and capture cellular morphology under many experimental conditions. However, cellular morphology can be difficult to characterize because the underlying phenotypes are too complex or too subtle. In addition, microscopy imaging is high-dimensional and subject to technical noise and unwanted variation. For these reasons, robust computational methods are required to properly quantify cellular morphology and make it accessible and useful for biological studies. Our work aims to develop computational frameworks for transforming and processing microscopy images to advance biological research.

Learning representations of cellular morphology

Obtaining measurements of cellular morphology is a critical step in most microscopy-based studies. We use representation learning methods to extract useful features from microscopy images in a data-driven manner. One of the main challenges of bioimage analysis is the lack of ground truth manual annotations from experts; after all, these images are acquired to discover unknown phenomena. Therefore, we use self-supervised learning for creating foundation models of cell morphology that can reveal biologically meaningful cellular variation in any microscopy experiment. Our research has advanced the state-of-the-art in image-based profiling, and we continue to investigate efficient and robust ways of universal models of cellular morphology to accelerate biological research.

Generative modeling of cellular variation

Microscopy images are acquired under rigorous experimental protocols that include control conditions and systematic interventions. The goal of experimental interventions is to reveal the effect of treatments (e.g. gene editing) on cells. However, experiments can be expensive to execute (time and resources), and the intervention space is too large to conduct all the desired experiments. We investigate the use of generative modeling of cellular images to explore interventions that have not been part of the original experimental design. In this way, we aim to augment the ability of researchers to reason about the biological systems that they study using microscopy images and generative AI. 

Spatial biology and tissue organization

The next frontier in cellular biology is to understand the spatial organization of cells in tissues, and how their interactions lead to more complex behaviors and functions. Imaging is an excellent approach to observe cells organized in tissues, and it is at the core of many novel spatial technologies, including spatial transcriptomics and spatial proteomics. Our research work aims to reveal the spatial patterns hidden within high-resolution images of tissues to facilitate the quantification of intervention effects in clinical studies and basic biology research. We develop machine learning methods for characterizing the multi-resolution and multi-dimensional structure of tissues.

3D imaging and volumetric microscopy

Novel techniques for observing organisms at the cellular level in 3D are rapidly being developed and used to investigate complex biological systems. The three-dimensional structure of cells and the way they interact can help us uncover the mechanisms underlying certain diseases. As a rich source of information, volumetric 3D microscopy also poses many computational challenges, from the big data generated for a single sample to the mathematical models required to guide the analysis. We leverage advances in machine learning to process 3D imaging accurately and efficiently.

Sup-topic artistic images generated with Google’s Gemini AI. The text was 100% written by humans.