Studying gene expression in a cancer patient’s cells can support clinical biologists understand the origins of cancer and predict the success of various treatments. However, cells are complicated and contain many layers, so how a biologist performs measurements affects what data they can obtain. For example, measuring proteins in a cell may provide different information about the effects of cancer than measuring gene expression or cell morphology.
It is significant where the information comes from in the cell. However, to obtain complete information about a cell’s state, scientists often need to perform multiple measurements using different techniques and analyze them one at a time. Machine learning methods can speed up this process, but existing methods combine all the information from each measurement method into one whole, making it arduous to determine which data comes from which part of the cell.
To overcome this problem, researchers at the Broad Institute of MIT and Harvard and the ETH Zurich/Paul Scherrer Institute (PSI) have developed an artificial intelligence-based platform that learns which cell state information is shared across different measurement modalities and which information is unique to a specific measurement type.
By precisely identifying which information comes from which parts of the cell, this approach provides a more holistic view of the cell’s state, making it easier for the biologist to see the full picture of cellular interactions. This could support scientists understand disease mechanisms and track the progression of cancer, neurodegenerative diseases such as Alzheimer’s disease, and metabolic diseases such as diabetes.
“When we study cells, one measurement is often not enough, so scientists are developing new technologies to measure different aspects of cells. Although we have many ways of looking at a cell, ultimately we only have one fundamental state of the cell. By combining information from all of these measurement methods in a smarter way, we could get a more complete picture of the cell’s state,” says lead author Xinyi Zhang SM ’22, PhD ’25, a former graduate student in MIT’s Department of Electrical Engineering and Computer Science (EECS) and a fellow at the Eric and Wendy Schmidt Center at Broad Institute at MIT and Harvard, who is currently group leader at AITHYRA in Vienna, Austria.
Zhang joins the article on the work of GV Shivashankar, professor at the Faculty of Health Sciences and Technology at ETH Zurich and head of the Multiscale Bioimaging Laboratory at PSI; and senior author Caroline Uhler, professor of EECS and the Institute for Data, Systems, and Society (IDSS) at MIT, member of MIT’s Laboratory for Information and Decision Systems (LIDS), and director of the Eric and Wendy Schmidt Center at the Broad Institute. Tests is published today in .
Manipulating multiple measurements
Scientists can utilize many tools to capture information about a cell’s state. For example, they can measure RNA to see if a cell is growing, or they can measure chromatin morphology to see if a cell is coping with external physical or chemical signals.
“When scientists perform multimodal analysis, they collect information using multiple measurement modalities and integrate it to better understand the underlying state of the cell. Some information is captured by only one modality, while other information is shared between modalities. To fully understand what is happening inside the cell, it is important to know where the information is coming from,” says Shivashankar.
Often, the only way for scientists to solve this problem is to conduct many individual experiments and compare the results. This leisurely and cumbersome process limits the amount of information they can collect.
In the novel work, the researchers built a machine learning framework that understands in detail what information overlaps across modalities and what information is unique to a particular modality but not captured by other modalities.
“As a user, you can simply enter your cell phone information and the system will automatically tell you which data is shared and which is modality-dependent,” Zhang says.
To build this framework, researchers rethought the typical way machine learning models are designed to capture and interpret multimodal cellular measurements.
Typically, these methods, called autoencoders, have one model for each measurement modality, and each model encodes a distinct representation of the data captured by that modality. A representation is a compressed version of the input that discards any irrelevant details.
The MIT method has a common representation space in which data overlapping across multiple modalities is encoded, as well as separate spaces in which unique data from each modality is encoded.
Basically, you can think of it as a Venn diagram of cellular data.
The researchers also used a special two-step training procedure that helps their model deal with the complexities involved in deciding what data is shared across multiple data modalities. After training, the model can identify which data is shared and which is unique when fed with cellular data it has never seen before.
Distinguishing data
When tested on synthetic datasets, the platform correctly captured known common and modality-specific information. When they applied their method to real single-cell datasets, it comprehensively and automatically distinguished gene activity recorded jointly by two measurement modalities, such as transcriptomics and chromatin accessibility, while correctly identifying which information came from only one of these modalities.
Additionally, the researchers used their method to determine which measurement method captured a specific protein marker that indicates DNA damage in cancer patients. Knowing where this information comes from would support clinical researchers determine the technique they should utilize to measure this marker.
“There are too many modalities in a cell and we can’t measure them all, so we need a prediction tool. But then the question becomes: Which modalities should we measure and which should we predict? Our method can answer this question,” says Uhler.
In the future, researchers want to enable the model to provide more interpretable information about the state of the cell. They also want to conduct additional experiments to ensure it correctly untangles the cellular information and applies the model to a broader range of clinical questions.
“It’s not enough to simply integrate information from all these modalities,” Uhler says. “We can learn a lot about the state of a cell if we carefully compare different modalities to understand how different cell components regulate each other.”
This research is funded in part by the Eric and Wendy Schmidt Center at the Broad Institute, the Swiss National Science Foundation, the U.S. National Institutes of Health, the U.S. Office of Naval Research, AstraZeneca, the MIT-IBM Watson AI Lab, the MIT J-Clinic for Machine Learning and Health, and the Simons Investigator Award.