April 2, 2025
AI-based technology developed in Dr Liat Keren’s lab at the Weizmann Institute of Science has shown it enables an unprecedented view of processes in body tissues.
Artificial intelligence systems are working magic in many areas of the life sciences – they help decipher protein structures, uncover hidden patterns in the genome, and process vast amounts of biological and medical data.
In a study published in the scientific journal Nature Biotechnology Weizmann Institute of Science scientists harness AI capabilities to provide an unprecedented glimpse into the body’s tissues.
The new technology, developed in Dr Liat Keren ‘s research group in the Department of Molecular Cell Biology, makes it possible to observe more proteins than ever before in a tissue sample, and it represents a significant leap forward in studying the cellular composition of tissues.
To get a snapshot of a tissue and understand how it functions, one must measure as many proteins as possible that function in it at the same time,” explained Dr Keren.
“This is essential for understanding the activity of tissues and essential for understanding diseases. In cancer, for example, understanding the composition of tumour tissue and how the different cell types in it react to each other may affect the effectiveness of treatment or predict which patients have a better chance of recovery.”
In contrast to currently accepted methods that provide information on only three or four proteins at a time, the new technology, called CombPlex, significantly increases the number of proteins that can be studied in a tissue sample: Already in the ‘proof of feasibility’ phase of the technology, it demonstrated the ability to simultaneously measure about twenty proteins at once in a single cell, but in the future this number may reach hundreds of proteins. In addition to improved imaging capabilities, the new technology does not require additional equipment beyond that already available in laboratories, and therefore may be accessible for wide-scale use.
Barcode for each protein
Labelling proteins with fluorescent tags has catalysed research in the field in recent decades, but it also has clear optical limitations: When several proteins are labelled with different colours in the same tissue sample, the labels can overlap, making the image unreadable. This limitation can be circumvented by repeated imaging—that is, washing out the existing labels and attaching new ones—but this approach, called cyclic fluorescence, is time-consuming and ultimately allows us to label a few dozen proteins at most, out of millions of possible protein variants produced based on the approximately 20,000 recipes in human DNA.
“We wanted to develop an imaging method that could capture more proteins at the same time, and ultimately all of them,” Dr Keren said as she described the vision and added a metaphor from the field of photography:
“Imagine that you want to document a room but can only photograph three or four objects at a time, say tables, chairs and screens. To include windows, carpets and lamps, you would have to take dozens of pictures, and eventually combine them to get an idea of the entire room. Our goal is to provide a means that allows you to photograph the entire room, with all its objects, with one click.”
To do this, the researchers first adopted a combinatorial approach: They labelled each molecule with several fluorescent tags, so that each protein is identified by a unique combination of colours that constitutes a kind of barcode. These barcodes make it possible to label a large number of molecules using a small number of colours – and as the number of colours increases, the number of possible barcodes jumps exponentially. This makes it possible to label more proteins in the tissue using fewer colours and use mechanical means to scan the barcodes and measure the proteins. But the challenges did not end there: above a certain number of barcodes, the colours mix and it is no longer possible to distinguish between the different proteins, even by mechanical means.
The researchers hypothesised that artificial intelligence might be able to overcome this hurdle. In collaboration with Dr Shai Begon of the Weizmann Institute’s Center for Artificial Intelligence, they developed an AI algorithm trained to distinguish proteins in tissue using fluorescence microscopy images from laboratories around the world.
The research team, which included researchers in the fields of biochemistry, bioinformatics, and mathematics, was amazed by the degree of success: images that appear to the human eye as a nightmarish fluorescent tangle were successfully decomposed by the algorithm into their components. This is how CombPlex was developed – an AI-assisted imaging method to measure many proteins at the single-cell level in a tissue sample.
Because the new technology can be used with common fluorescent microscopes, it could greatly advance tissue research in both research laboratories and hospitals. In addition, the method not only provides a more complete picture compared to existing methods, it also does so faster: within a day or two instead of several weeks.
“We hope that CombPlex will replace existing methods in the future, as it offers a comprehensive and in-depth look at tissues and can lead to more accurate clinical insights,” says Dr Keren.
The Bina Pre-Application Research Unit guided and funded the development process of CombPlex.
“When we consulted with various experts in the field, they were all very enthusiastic about the possibilities offered by the new technology,” said Dr Sharon Fairman, who heads the unit, which was founded in 2021 to identify projects with application potential by the institute’s scientists in the early stages.
From top to bottom: Raz Ben-Uri (right), Dr Shai Begon, Lior Ben-Shabbat, Dr Ofer Elhanani, Dana Sheinshein and Dr Liat Keren
An illegible microscopy image with overlapping fluorescent labels of seven proteins (left) was translated with great precision using CombPlex into an image that allows the different proteins to be distinguished (right)
Dr Liat Keren
