User Group: Researcher

Obesity and high body fat increase the risk of diabetes. But not every overweight person also develops the disease. The decisive factor is where the fat is stored in the body. If fat is stored under the skin, it is more harmless than fat in deeper areas of the abdomen (known as visceral fat). How fat is distributed throughout the body can be easily visualized with whole-body magnetic resonance imaging. “We have now investigated whether type 2 diabetes could also be diagnosed on the basis of certain patterns of body fat distribution in MRI,” said last author Prof. Robert Wagner, explaining the researchers’ approach.

Deep learning trained with over 2000 MRI images

To detect such patterns, the researchers used artificial intelligence (AI). They trained deep-learning (machine learning) networks with whole-body MRI scans of 2000 people who had also undergone screening with oral glucose tolerance testing (abbreviated oGTT). The oGTT, also called the sugar load test, can detect impaired glucose metabolism and diagnose diabetes. This is how AI learned to detect diabetes.

Fat accumulation in lower abdomen important clue to diabetes development

“An analysis of the model results showed that fat accumulation in the lower abdomen plays a crucial role in diabetes detection,” Wagner said. Further additional analysis also showed that a proportion of people with pre-diabetes, as well as people with a diabetes subtype that can lead to kidney disease, can also be identified via MRI scans.

Researchers are now working to decipher the biological control of body fat distribution. One goal is to identify the causes of diabetes through new methods, such as the use of AI, to find better prevention and treatment options.

Computer scientist Moritz Hardt, Ph.D., has accepted a call to the Max Planck Institute for Intelligent Systems (MPI-IS). He joined the institute part-time in July 2021, and will start working full-time in Tübingen as of January 2022. His newly established department “Social Foundations of Computation” aims to build foundations of computation that thrive on social context. The research is intended to contribute to a paradigm shift within computer science that treats computer science from the ground up as a social science that takes into account the role of society as a whole, as well as the actions and dynamic behavior of individuals – especially when algorithms influence people's life chances.

Computation as a form of social action

“Our research will examine the interplay between algorithms and society,” says Moritz Hardt. “Not only looking at isolated consequences of the technology, our research aims to develop computational methods that are built on social foundations from the outset. The goal is to develop a social perspective on computer science that extends the discipline itself to a social science.”

In a world increasingly permeated by algorithms, only recently, a research community in computer science has been growing, questioning how often sensitive data leads to decisions, and foregrounding people who produce and consume data. The topic of fairness, for example, became a central focus of research within the machine learning community only about five years ago. An active field of research emerged that is receiving ever more attention. With Hardt’s appointment, the Max Planck Society is strengthening basic research within computer science that considers the role of society as well as norms and values as an integral part of computation. Every statistical model, prediction or calculation should be based on a social foundation; especially when a computer, rather than a human, makes decisions, Hardt claims.

“Moritz Hardt has made fundamental mathematical and algorithmic contributions to the interaction between machine learning methods and their social embedding. We are very pleased that he will continue his research at the MPI-IS, working together to develop an artificial intelligence that brings to the fore its responsibility for shaping our future,” says Bernhard Schölkopf, Managing Director at the MPI-IS.

“Computer science had neglected society”

Hardt, who received his doctorate from the University of Princeton in 2011, became intrigued with the topic of privacy and fairness in sensitive data analysis early on in his doctoral work, when the topic had little to no support within the community. “I got into social issues through my work on differential privacy. I realized it wasn’t just about privacy, but many other questions that computer scientists had previously ignored over the decades. Computer science had neglected society. Even today, social problems are often categorized as a consequence of the technology rather than a central object of research,” says Hardt.

Social Foundations of Computation has four focus research areas: Applying machine learning in social and economic contexts, formulating social and dynamic actions as mathematical models, as well as examining the validity and reliability of statistical methods and the construction of datasets within scientific communities. Another focus is the pursuit of normative goals and, in particular, how to formulate values and norms mathematically. “My goal as a director at the institute is to make it easier for other researchers to work on forward-looking questions that are currently underappreciated,” Hardt says.

Celestine Mendler-Dünner, a group leader who is helping to build up the new department, has already started her work at the Tübingen site of MPI-IS. Her research focuses on modeling social dynamics and how to incorporate them into the design of machine learning algorithms. She has previously worked with Moritz Hardt in Berkeley and contributed to shaping the department’s research direction. Additional appointments will follow in a timely manner. The department is expected to grow to a total of about two dozen scientists.

Prior to his appointment as director at MPI-IS, Moritz Hardt was an Assistant Professor in the Department of Electrical Engineering and Computer Sciences at the University of California, Berkeley. He received his Ph.D. from the Department of Computer Science at Princeton University with a dissertation on privacy and fairness in the classification of sensitive data. He subsequently worked at IBM Research and Google. Hardt is co-founder of the conference “Fairness, Accountability, and Transparency in Machine Learning.” He is co-author of “Fairness and Machine Learning: Limitations and Opportunities” (MIT Press, 2022) and “Patterns, Predictions, and Actions: A Story About Machine Learning” (Princeton University Press, 2022).

For more than a hundred years, neuroscientists have been studying the properties of neurons in the brain. A key question here is how the different types of nerve cells differ from one another and how this affects brain activity. The BRAIN Initiative Cell Census Network (BICCN), a large international research collaboration including scientists from the University of Tübingen has now published groundbreaking results: They have created a cell atlas providing an overview of the different neuron types and their respective properties in the motor cortex, the brain region that controls movement, in the brains of mice, monkeys and humans. The atlas is described in a special package of 17 articles published today in the journal Nature, including a single flagship paper that describes the entire atlas.

The BICCN collaboration sought to describe the neural types in the motor cortex as fully as possible. Its goal was to provide a catalogue based on a “Cell Census”, that similar to a population census would define the neural types, their distribution in the brain and their defining characteristics.

Using novel experimental and data analysis methods the scientists were able to collect information from over a million cells. For a part of the cells, they recorded their spatial location, their shape, their genetic activity, their electrical properties, and traced their connections to other neurons in other brain areas. Since this was done in mice, in marmosets, and in humans, it even allowed them to reconstruct the evolutionary development of neural types. The result is a cell catalogue that gives a unique overview of the brain cortex and its evolution.

Scientists from the University of Tübingen contributed to this joint effort with a paper on a multimodal characterization of the different cell types in the mouse motor cortex, integrating different types of data. The work was led by Prof. Philipp Berens, Professor at the Institute for Ophthalmic Research at the University of Tübingen and speaker of the Cluster of Excellence “Machine Learning”, Prof. Andreas Tolias from Baylor College of Medicine in Houston, Texas (USA), and Prof. Rickard Sandberg from Karolinska Institutet in Stockholm (Sweden). It provides one of the most complete characterizations of the diversity of neural types in the mouse brain so far. The paper, which had already been published online, is part of Nature’s special issue.

Neuroscientists mostly use three fundamental features to describe neurons: their anatomy or how they look under a microscope, their physiology or how they respond when stimulated, and their transcriptome, which are the genes they express. The team used an experimentally challenging technique called “Patch-seq” to collect a large database including anatomical, physiological, and genetic information from cells in the mouse motor cortex.

Gathering these three fundamental features from more than 1000 cells at the same time was the key that enabled them to get a deep understanding of how neurons in the motor cortex are related to each other. Using machine learning approaches, they integrated the anatomical, physiological, and genetic information and revealed relations between the neurons that had not been known before. “While the broad genetic families of neurons had distinct anatomical and physiological properties, within each family the neurons exhibited continuous anatomical and physiological diversity”, Dr. Dmitry Kobak from University of Tübingen explains. He and Federico Scala from Baylor College of Medicine are the lead authors of the study.

By drawing an analogy with the evolutionary tree which describes the relationships between the various species, the researchers came to the following conclusions: Neurons follow a hierarchy consisting of distinct, non-overlapping branches at the level of families. Within each family, however, they show continuous changes in their genetic, anatomical, and physiological features. This view suggests a fundamentally different model to describe the diversity of neurons we find in the brain.

“The data from the new cell atlas of our collaboration will be a tremendous resource for neuroscience,” says Philipp Berens. “Bringing our machine learning expertise to the table, we provide the links from genetics to the neurons’ physiology and anatomy, which may be crucial to understand brain-related diseases at the cell type level.”