Shanghai Jiaotong University Team Achieves First Frame-by-Frame Gene Regulation Decoding, Visualizing Embryo Development at 1-Minute Resolution}
A pioneering study by Shanghai Jiaotong University decodes embryonic gene regulation at 1-minute resolution, transforming our understanding of developmental dynamics through advanced deep learning techniques.
Embryonic development, the process of life formation from a single cell to a complex organism, is governed by intricate spatial-temporal gene regulation networks that determine cell fate and body structure.
Traditional imaging methods struggle to track multiple molecular fluctuations simultaneously over time. Solid embryo imaging offers sensitivity and capacity but lacks high temporal resolution.
Shanghai Jiaotong University proposed a multi-scale deep learning approach to precisely infer absolute developmental time from fixed Drosophila embryo images at a 1-minute resolution.
This research, titled Deep learning-based high-resolution time inference for deciphering dynamic gene regulation from fixed embryos, was published in Nature Communications on July 16, 2025.
Link to the paper: https://www.nature.com/articles/s41467-025-61907-7
Decoding Embryo Development
Embryonic development is orchestrated by complex gene regulatory networks. In early Drosophila embryos, although no visible body structures form yet, future segments along the anterior-posterior (AP) axis are already pre-defined.
Conventional real-time imaging is limited by complex interactions, and fixed imaging lacks fine spatial-temporal resolution. Deep learning helps address these issues. The team used a regression method to infer the absolute developmental time of early Drosophila embryos during stages nc11-14, with a resolution of 0.3 to 1 minute.
This regression employed three independent convolutional neural networks (CNNs) to capture morphological features across multiple spatial scales, accurately predicting developmental time from standard DNA images, unaffected by strain differences.
To relate developmental time to dynamic nuclear signals, over 30 transgenic Drosophila embryos were imaged for histone H2A-RFP, revealing significant differences in nuclear counts between cell cycles.
Figure 1: CNN models predict developmental time from live embryo nuclear histone images. (Source: Paper)
The three CNN models each output a regression prediction based on multi-scale morphological features, dividing the embryo image into various-sized windows.
These models' predictions are combined via median filtering, achieving high accuracy with an average error of only 0.25 minutes across stages nc11-14 (nc11: 100%, nc12: 98%, nc13: 100%, nc14: 87%).
The CNN-based approach demonstrates high value in precise time inference from comprehensive image features, outperforming traditional membrane-based methods, especially for nc14 stage.
DNA-based Experiments
Compared to histones, DNA can be labeled more easily with organic dyes. However, DNA and histone images differ significantly, requiring adaptation of models.
Researchers stained 160 fixed embryos for both DNA and histones, using CNNs trained on histone images to infer developmental time, then applying these models to DNA images. After calibration, predictions from DNA models closely matched histone-based inferences.
Figure 2: Transfer learning from fixed embryo histone images to DNA images for developmental time inference. (Source: Paper)
To extend this method to other Drosophila strains, especially wild-type, DNA images need appropriate scaling. With accurate timing, gene expression analysis becomes more precise.
For example, analyzing the gene Kr, the team quantified the spatial-temporal distribution of transcription and protein concentrations during nc11-13, revealing two distinct peaks in Kr expression during nc13, with early peaks being shorter, explaining previous observations of increased late Kr expression.
Figure 3: Dynamic regulation of Kr gene by multiple transcription factors during nc13. (Source: Paper)
This approach demonstrates powerful capabilities in revealing key gene transcription regulation at single-cell and single-molecule levels, with potential to integrate with single-cell sequencing, spatial transcriptomics, and organoid imaging—transforming static slices into high-definition time-lapse videos across species and tissues.
Although currently requiring extensive embryo imaging, this method offers an effective, gene-modification-free way to quantify developmental dynamics, with future integration with high-throughput RNA and protein data expected to advance the field further.