Can Generative AI Replace Immunofluorescent Staining Processes? A Comparison Study of Synthetically Generated CellPainting Images from Brightfield

Xiaodan Xing, Siofra Murdoch, Chunling Tang, Giorgos Papanastasiou, Jan Cross-Zamirski, Yunzhe Guo, Xianglu Xiao, Carola‐Bibiane Schönlieb, Yinhai Wang, Guang Yang

Abstract

Cell imaging assays utilizing fluorescence stains are essential for observing sub-cellular organelles and their responses to perturbations. Immunofluorescent staining process is routinely in labs, however the recent innovations in generative AI is challenging the idea of IF staining are required. This is especially true when the availability and cost of specific fluorescence dyes is a problem to some labs. Furthermore, staining process takes time and leads to inter-intra technician and hinders downstream image and data analysis, and the reusability of image data for other projects. Recent studies showed the use of generated synthetic immunofluorescence (IF) images from brightfield (BF) images using generative AI algorithms in the literature. Therefore, in this study, we benchmark and compare five models from three types of IF generation backbones, CNN, GAN, and diffusion models, using a publicly available dataset. This paper not only serves as a comparative study to determine the best-performing model but also proposes a comprehensive analysis pipeline for evaluating the efficacy of generators in IF image synthesis. We highlighted the potential of deep learning-based generators for IF image synthesis, while also discussed potential issues and future research directions. Although generative AI shows promise in simplifying cell phenotyping using only BF images with IF staining, further research and validations are needed to address the key challenges of model generalisability, batch effects, feature relevance and computational costs.

Introduction: The Motivation for In Silico Staining

In drug discovery and cell biology, immunofluorescence (IF) is the standard for visualizing specific organelles. However, the conventional workflow is often inefficient, costly, and destructive, typically requiring cell fixation which kills the samples. Brightfield (BF) imaging offers a non-destructive, fast, and inexpensive alternative, but lacks the contrast needed to distinguish specific biological structures.

Generative AI presents a potential solution: "in silico" staining. By training models to translate BF images into synthetic IF images, researchers hope to achieve the benefits of fluorescence without the wet-lab drawbacks. This study benchmarks various AI architectures to determine if they can truly replace traditional staining.

Overview of the motivation, image synthesis models, and evaluation pipeline.

Methodology and Benchmarking

The researchers evaluated five models across three major generative families:

  1. CNN (UNet): A standard baseline for image-to-image translation.
  2. GANs (Pix2Pix and SPADE): These utilize a discriminator to improve texture realism.
  3. Diffusion Models (Palette and SPADE-Diffusion): Modern state-of-the-art models known for high-quality image generation.

Beyond standard pixel-level metrics like MSE, PSNR, and SSIM, the study introduced a "biological meaning" evaluation pipeline. Using CellProfiler, the team extracted over 3,000 morphological features to assess if synthetic images preserve the biological signals necessary for downstream tasks like Mechanism of Action (MoA) prediction.

The study utilized the cpg0000 dataset, comprising over 3 million image pairs. The data includes two distinct cell lines—A549 (lung cancer) and U2OS (bone cancer)—treated with hundreds of chemical compounds and CRISPR knockouts. This diversity allowed for a rigorous test of model generalizability across different biological contexts.

Results: Visual Fidelity vs. Biological Utility

The benchmarking revealed a nuanced landscape where visual quality does not always equate to biological accuracy.

GAN Supremacy in Visuals

Pix2Pix emerged as the top performer for visual quality, capturing fine organelle details that CNNs often blurred. It avoided the "hallucinations" or structural artifacts sometimes observed in Diffusion-based models.

Visual comparison of synthetic IF images generated by different models across five channels (Mito, AGP, RNA, ER, DNA).

The Diffusion Trade-off

While Diffusion models like Palette produced noisier images and were significantly slower (taking 240 seconds per image vs. 0.07 seconds for Pix2Pix), they were surprisingly effective at predicting a drug's MoA. This suggests that these models capture subtle biological signals that may not be visually "clean" but are analytically relevant.

The Generalizability Wall

A critical finding was the lack of generalizability. Models trained on A549 cells saw a performance drop of over 60% in tasks like cell counting when applied to U2OS cells. This "generalizability gap" indicates that current models are highly specific to the cell lines they were trained on.

Discussion and Future Directions

While generative AI can achieve roughly 70% accuracy in MoA identification—proving its utility for high-throughput screening—it is not yet a total replacement for IF staining. Key challenges remain:

  • Trustworthiness: Feature correlations between real and synthetic images remain relatively low (under 0.3).
  • Batch Effects: Models are sensitive to specific lab conditions.
  • Computational Cost: Diffusion models are currently too slow for massive-scale screening.

Future research should focus on "Foundation Models for Biology"—large-scale AI trained on diverse datasets to overcome the generalizability barrier and provide a more robust tool for in silico staining.

Summary updated

Abstract

Cell imaging assays utilizing fluorescence stains are essential for observing sub-cellular organelles and their responses to perturbations. Immunofluorescent staining process is routinely in labs, however the recent innovations in generative AI is challenging the idea of IF staining are required. This is especially true when the availability and cost of specific fluorescence dyes is a problem to some labs. Furthermore, staining process takes time and leads to inter-intra technician and hinders downstream image and data analysis, and the reusability of image data for other projects. Recent studies showed the use of generated synthetic immunofluorescence (IF) images from brightfield (BF) images using generative AI algorithms in the literature. Therefore, in this study, we benchmark and compare five models from three types of IF generation backbones, CNN, GAN, and diffusion models, using a publicly available dataset. This paper not only serves as a comparative study to determine the best-performing model but also proposes a comprehensive analysis pipeline for evaluating the efficacy of generators in IF image synthesis. We highlighted the potential of deep learning-based generators for IF image synthesis, while also discussed potential issues and future research directions. Although generative AI shows promise in simplifying cell phenotyping using only BF images with IF staining, further research and validations are needed to address the key challenges of model generalisability, batch effects, feature relevance and computational costs.
BibTeX
@article{xing2024can,
  author = {Xiaodan Xing and Siofra Murdoch and Chunling Tang and Giorgos Papanastasiou and Jan Cross-Zamirski and Yunzhe Guo and Xianglu Xiao and Carola‐Bibiane Schönlieb and Yinhai Wang and Guang Yang},
  title = {{Can Generative AI Replace Immunofluorescent Staining Processes? A Comparison Study of Synthetically Generated CellPainting Images from Brightfield}},
  journal = {Computers in Biology and Medicine},
  year = {2024},
  doi = {10.1016/j.compbiomed.2024.109102},
}
Figure