18 May 2026
Researchers at Human Technopole have developed a novel machine learning-based method that transforms composite fluorescence microscopy images with overlapping signals into separate images revealing individual cellular structures. The tool is published in Nature Methods, with experimental models and training data openly available on GitHub.
Berlin, 1871. The orange-red powder dissolved in water, and a bright yellow-green glow appeared. Chemist Adolf von Baeyer had just created the first synthetic fluorescent molecule, later called fluorescein.
Since Baeyer’s groundbreaking discovery, fluorescent molecules have transformed the way scientists observe biological systems under the microscope. By tagging selected cellular structures, researchers can now track protein localisation, monitor gene expression, observe intracellular organisation, detect cell interactions, perform live-cell imaging, and more.
The challenge becomes greater when researchers want to visualise several cellular structures at once. Multiplexed fluorescence microscopy relies on combining fluorophores whose excitation and emission spectra can be clearly separated. When these spectra overlap, signals may spill into the wrong channel, producing undesired effects such as crosstalk, bleedthrough, and potentially misleading images. Selecting the right set of fluorophores is an art in itself: one must understand the properties of each colour to achieve the best visual effect, not unlike what artists do.
Even when an appropriate palette of fluorophores is selected, samples cannot be imaged indefinitely. Repeated or prolonged exposure can cause photobleaching, in which fluorophores lose their ability to emit light. Therefore, choosing shorter or less frequent exposure times is fundamental to avoid exhausting the fluorescence source.
Finding a balance between signal intensity, specificity, and sample preservation is a constant challenge for microscopists. Even Nobel laureate Eric Betzig pointed it out in his Nobel Prize lecture in 2014, highlighting the widespread recognition of this notorious issue.
The Jug Group at Human Technopole has now developed a way to facilitate this balancing act. Their novel method—called Micro𝕊plit—allows multiple fluorescently labelled cellular structures to be detected in a single channel—corresponding to a single collective colour. The resulting superimposed image is computationally unmixed into separate channels, each displaying one labelled structure.
Florian Jug, Group Leader in the Computational Biology Centre and senior author of this study, explains: “It’s somewhat similar to photographing a group of people in one shot instead of taking individual portraits, and then having a software return you the individual portraits, reliably.”
Micro𝕊plit can be trained using different types of microscopy data, making it adaptable to the biological material at hand. It can learn from previously acquired images of individual structures, multiplexed images, or from newly generated training datasets in which both separate and mixed images are acquired. Once trained, it can split a mixed input image into denoised images of the individual structures, even when the training data itself is noisy.
The unmixed denoised images generated by Micro𝕊plit are based on predictions of an artificial neural network. Starting from the provided training material, the machine learning-driven method predicts how the superimposed structures are spatially distributed and splits them into separate images, each highlighting one type of biological structure. In cases where the input images are ambiguous, multiple solutions can be generated, and by looking at their variability it is possible to determine the certainty versus error rate of the predictions. Low diversity is a sign of high accuracy, while high variability indicates ambiguous solutions and potentially elevated error rate. This intrinsic property of Micro𝕊plit enables the users to identify and discard entire or partial images with unreliable predictions, and at the same time it provides evidence that those unmixed images predicted with consistency are of high quality. This feature empowers the users to verify the AI-derived predictions and choose to rely on the system based on evidence rather than blind trust. “An AI method whose only output is a beautiful image is useless to a scientist,” emphasises Florian Jug. “Micro𝕊plit’s output is also an image but completed by a map of where it might be wrong.”
As Micro𝕊plit can separate different cellular structures, it can also identify imaging artifacts. Should any structured noise be present in the input image, Micro𝕊plit can help remove it, as long as it does not always overlap with structures of interest. This application is extremely useful and truly unique, distinguishing Micro𝕊plit from other content-aware image denoising methods.
In addition to the identification and unmixing of cellular structures and artifact removal, Micro𝕊plit can support a range of other downstream analyses. For example, it could be used to track the presence, absence, or exact number of specific structures, calculate their dimensional properties, or study the interactions between different cellular components in various experimental conditions. As it has only recently made available to the scientific community, its potential is not fully explored yet, and novel exciting applications still remain to be discovered.
To facilitate the further implementation and reuse of this method by others, the complete Micro𝕊plit package, including training and evaluation data, as well as example notebooks for all performed experiments, is publicly available. Every owner of a fluorescent microscope can instantly benefit from it, as Micro𝕊plit is compatible with any imaging device. Moreover, the Jug Group is now working on the development of user-friendly graphical tools to reach an even broader user community by providing full support from image acquisition to data visualisation.
From the first synthetic fluorescent molecule to machine learning-based image unmixing, scientific advances have continually expanded what researchers can see. Micro𝕊plit is set to contribute to pushing the boundaries even further by optimising fluorescence microscopy to extract more information from less data.
