Periodic Reporting for period 2 - DeepEmbryo (Reverse-engineering the development of embryos with physics-informed machine learning)
Período documentado: 2022-07-01 hasta 2023-12-31
The first steps of animal embryo development involves a succession of cell divisions, cell rearrangements, signaling events and gene regulation. The principles of self-organization underlying early embryogenesis remain poorly understood, as they involve unknown feedbacks between geometry, mechanics, and signaling. The DeepEmbryo project aims to develop theoretical and computational methods to reverse engineer the very first stages of embryonic development from biological data, in particular from fluorescence microscopy imaging data. Beyond the experimental dissection of the molecular pathways underlying the cellular mechanisms, artificial intelligence methods such as convolutional neural networks, combined with physical modeling, are proposed here as a new tool to discover unknown couplings between geometry, mechanics and signalling. The project aims to 1) infer the forces that shape early embryos, (2) to decipher the couplings between forces, divisions and signaling, and 3) to uncover some principles underlying developmental robustness. The project should have potential applications in reproductive medicine, to help select embryos with the best potential for implantation in the context of assisted reproduction techniques.
Halfway through the project, the first objective was achieved. We have developed and validated a powerful technique to segment three-dimensional microscopy images of early embryos into surface meshes and to infer mechanical maps of relative cellular forces in the embryo from these meshes. Although our method does not rely on deep learning, we are working on an extension of the method which fundamentally relies on automatic differentiation as in neural networks. We have developed a new artificial microscopy image rendering technique that takes a mesh resulting from a mechanical simulation to create a realistic fluorescence image in a fully differentiable way. This allows us to consider training model parameters - such as forces - to match an artificial image with a real image of an early embryo.
While 2D force inference methods were developed, the 3rd dimension had remained out of reach. Our 3D force inference method fills this important gap and will allow the creation of spatio-temporal mechanical atlases of various early animal embryos, which can be combined with other data such as gene expression maps obtained through recent techniques of single-cell sequencing, in the hope of deducing couplings between the mechanics and the gene regulation.
Labeling 3D microscopy image data to train neural networks is a daunting – and sometimes impossible – task that limits the possibilities offered by deep learning to advance task automation in biology. Our method for creating artificial images should become a generic tool to generate large datasets to train neural networks with naturally labeled data as they come directly from physical simulations.
Until the end of the project, we hope to be able to learn the couplings between cell divisions and cell geometry/mechanics directly from time series of 3D microscopy images of early embryonic development. We believe that we will also be able to infer potential causal relationships between mechanics, geometry and gene expression from the combined analysis of gene expression and mechanical maps.
Labeling 3D microscopy image data to train neural networks is a daunting – and sometimes impossible – task that limits the possibilities offered by deep learning to advance task automation in biology. Our method for creating artificial images should become a generic tool to generate large datasets to train neural networks with naturally labeled data as they come directly from physical simulations.
Until the end of the project, we hope to be able to learn the couplings between cell divisions and cell geometry/mechanics directly from time series of 3D microscopy images of early embryonic development. We believe that we will also be able to infer potential causal relationships between mechanics, geometry and gene expression from the combined analysis of gene expression and mechanical maps.