Aurelia Butler

A growing repertoire of cell and molecule-based immunotherapies is offering patients with indomitable cancers new hope by mobilizing their immune systems against tumor cells. An emerging class of such immunotherapeutics, known as T cell bispecific antibodies (TCBs), are of growing importance with several TCBs that the U.S. Food and Drug Administration (FDA) approved for the treatment of leukemias, lymphomas, and myelomas. These antibody drugs label tumor cells with one of their ends, and attract immune cells with another end to coerce them into tumor cell killing.
One major challenge in the development of TCBs and other immunotherapy drugs is that the antigens targeted by TCBs can be present not only on tumor cells, but also healthy cells in the body. This can lead to “on-target, off-tumor” cell killing and unwanted injury of vital organs, such as the kidney, liver, and others, that can put patients participating in clinical trials at risk. Currently, there are no human in vitro models of the kidney that sufficiently recapitulate the 3D architecture, cell diversity, and functionality of organs needed to assess on-target, off-tumor effects at a preclinical stage.
Now, a new cross-disciplinary, cross-organizational study created an immune-infiltrated kidney tissue model for investigating on-target, off-tumor effects of TCBs and potentially other immunotherapy drugs. The team of bioengineers and immune-oncologists who performed the study at the Wyss Institute for Biologically Inspired Engineering at Harvard University, Harvard John A. Paulson School of Engineering and Applied Sciences (SEAS), Harvard Medical School (HMS), and the Roche Innovation Centers in Switzerland and Germany, developed an immune-infiltrated human kidney organoid-on-chip model composed of tiny kidney tissue segments that contain vasculatureand forming nephrons, which can be infiltrated by circulating immune cells. They used this model to understand the specific toxicity of a pre-clinical TCB tool compound that targets the well-characterized tumor antigen Wilms’ Tumor 1 (WT-1) in certain tumors. Importantly, WT-1 is also expressed at much lower levels in the kidney, making it an important organ to study its potential on-target, off-tumor effects in. Their findings are published in PNAS.
“Together with our collaborators at Roche, we extended our vascularized kidney organoid-on-chip model to include an immune cell population that contains cytotoxic T cells with the potential to kill not only tumor cells, but also other cells that present target antigens,” said Wyss Core Faculty member Jennifer Lewis, Sc.D., the study’s senior author. “Our pre-clinical human in vitro model provides important insights regarding which cells are targeted by a given TCB and what, if any, off-target damage arises.” Lewis is also the Hansjörg Wyss Professor of Biologically Inspired Engineering at SEAS and co-leader of the Wyss Institute’s 3D Organ Engineering Initiative.
Incorporating immunity into a kidney organoid-on-chip
In 2019, Lewis’ group, together with that of Joseph Bonventre, M.D., Ph.D. at Brigham and Women’s hospital along with co-author Ryuji Morizane, M.D., Ph.D., found that exposing kidney organoids created from human pluripotent stem cells to the constant flow of fluids during their differentiation enhanced their on-chip vascularization and maturation of glomeruli and tubular compartments, relative to static controls. The researchers’ observations were enabled by a 3D printed millifluidic chip, in which kidney organoids are subjected to nutrient and differentiation factor-laden media flowed at controlled rates during their differentiation. The chip device allows researchers to directly observe the tissue using confocal microscopy through a transparent window in real-time.
“Given that this in vitro model represents most of the cell types in the kidney and incorporates the immune system, itcould support the assessment of on and off-target effects from TCBs as well as complex cellular interactions,” said Kimberly Homan, Ph.D., a former postdoctoral researcher in Lewis’ lab, first author of the initial work, and a co-corresponding author of this new study. Homan has since left Lewis’ lab to join Genentech as Director of the Complex in vitro Systems lab where she continued to provide expertise to the project collaborators. More

A tailsitter is a fixed-wing aircraft that takes off and lands vertically (it sits on its tail on the landing pad), and then tilts horizontally for forward flight. Faster and more efficient than quadcopter drones, these versatile aircraft can fly over a large area like an airplane but also hover like a helicopter, making them well-suited for tasks like search-and-rescue or parcel delivery.
MIT researchers have developed new algorithms for trajectory planning and control of a tailsitter that take advantage of the maneuverability and versatility of this type of aircraft. Their algorithms can execute challenging maneuvers, like sideways or upside-down flight, and are so computationally efficient that they can plan complex trajectories in real-time.
Typically, other methods either simplify the system dynamics in their trajectory planning algorithm or use two different models, one for helicopter mode and one for airplane mode. Neither approach can plan and execute trajectories that are as aggressive as those demonstrated by the MIT team.
“We wanted to really exploit all the power the system has. These aircraft, even if they are very small, are quite powerful and capable of exciting acrobatic maneuvers. With our approach, using one model, we can cover the entire flight envelope — all the conditions in which the vehicle can fly,” says Ezra Tal, a research scientist in the Laboratory for Information and Decision Systems (LIDS) and lead author of a new paper describing the work.
Tal and his collaborators used their trajectory generation and control algorithms to demonstrate tailsitters that perform complex maneuvers like loops, rolls, and climbing turns, and they even showcased a drone race where three tailsitters sped through aerial gates and performed several synchronized, acrobatic maneuvers.
These algorithms could potentially enable tailsitters to autonomously perform complex moves in dynamic environments, such as flying into a collapsed building and avoiding obstacles while on a rapid search for survivors.
Joining Tal on the paper are Gilhyun Ryou, a graduate student in the Department of Electrical Engineering and Computer Science (EECS); and senior author Sertac Karaman, associate professor of aeronautics and astronautics and director of LIDS.The research appears in IEEE Transactions on Robotics.
Tackling tailsitter trajectories More

In the United States and other western countries, a form of esophageal and stomach cancer has risen dramatically over the last five decades. Rates of esophageal adenocarcinoma, or EAC, and gastric cardia adenocarcinoma, or GCA, are both highly fatal.
However, Joel Rubenstein, M.D., M.S., a research scientist at the Lieutenant Colonel Charles S. Kettles Veterans Affairs Center for Clinical Management Research and professor of internal medicine at Michigan Medicine, says that preventative measures can be a saving grace.
“Screening can identify pre-cancerous changes in patients, Barrett’s esophagus, which is sometimes diagnosed in individuals who have long-term gastroesophageal reflux disease, or GERD,” he said.
“When early detection occurs, patients can take additional steps to help prevent cancer.”
While current guidelines already consider screening in high-risk patients, Rubenstein notes that many providers are still unfamiliar with this recommendation.
“Many individuals who develop these types of cancer never had screening to begin with,” he said.
“But a new automated tool embedded in the electronic health record holds the potential to bridge the gap between provider awareness and patients who are at an increased risk of developing esophageal adenocarcinoma and gastric cardia adenocarcinoma.”
Rubenstein and a team of researchers used a type of artificial intelligence to examine data regarding EAC and GCA rates in over 10 million U.S. veterans. More

How is a donut similar to a coffee cup? This question often serves as an illustrative example to explain the concept of topology. Topology is a field of mathematics that examines the properties of objects that remain consistent even when they are stretched or deformed — provided they are not torn or stitched together. For instance, both a donut and a coffee cup have a single hole. This means, theoretically, if either were pliable enough, it could be reshaped into the other. This branch of mathematics provides a more flexible way to describe shapes in data, such as the connections between individuals in a social network or the atomic coordinates of materials. This understanding has led to the development of a novel technique: topological data analysis.
In a study published this month in The Journal of Chemical Physics, researchers from SANKEN (The Institute of Scientific and Industrial Research) at Osaka University and two other universities have used topological data analysis and machine learning to formulate a new method to predict the properties of amorphous materials.
A standout technique in the realm of topological data analysis is persistent homology. This method offers insights into topological features, specifically the “holes” and “cavities” within data. When applied to material structures, it allows us to identify and quantify their crucial structural characteristics.
Now, these researchers have employed a method that combines persistent homology and machine learning to predict the properties of amorphous materials. Amorphous materials, which include substances like glass, consist of disordered particles that lack repeating patterns.
A crucial aspect of using machine-learning models to predict the physical properties of amorphous substances lies in finding an appropriate method to convert atomic coordinates into a list of vectors. Merely utilizing coordinates as a list of vectors is inadequate because the energies of amorphous systems remain unchanged with rotation, translation, and permutation of the same type of atoms. Consequently, the representation of atomic configurations should embody these symmetry constraints. Topological methods are inherently well-suited for such challenges. “Using conventional methods to extract information about the connections between numerous atoms characterizing amorphous structures was challenging. However, the task has become more straightforward with the application of persistent homology,” explains Emi Minamitani, the lead author of the study.
The researchers discovered that by integrating persistent homology with basic machine-learning models, they could accurately predict the energies of disordered structures composed of carbon atoms at varying densities. This strategy demands significantly less computational time compared to quantum mechanics-based simulations of these amorphous materials.
The techniques showcased in this study hold potential for facilitating more efficient and rapid calculations of material properties in other disordered systems, such as amorphous glasses or metal alloys. More

Proteins consist of amino acids, which are linked to form long amino acid chains as specified by our genetic material. In our cells, these chains are not simply rolled up like strings of pearls, but fold into complex, three-dimensional structures. How a protein is folded decisively influences its function: It determines, for example, which other molecules a protein can interact with in the cell. Knowledge of the three-dimensional structure of proteins is therefore of great interest to the life sciences and plays a role in drug development, among other things.
“Unfortunately, elucidating the structure of a protein is anything but trivial, and focusing on a single state does not always provide meaningful information, especially if the protein is highly flexible in terms of its structure,” says Tobias Schneider, a member of Michael Kovermann’s lab team in the Department of Chemistry at the University of Konstanz. The reason: complex proteins often fold into several compact subunits, called domains, which in turn may be connected by flexible linkers. The more flexibly connected subunits are present, the more different three-dimensional structures a protein can theoretically adopt. “This means that a protein in solution, for example inside our cells, has several equal states and constantly switches between them,” Schneider explains.
Tracking down the structural ensemble
A simple snapshot is not sufficient to fully describe the structural features of such multi-domain proteins, as it would capture only one of many states at a time. To get a detailed picture of the possible structures of such proteins, a smart combination of different methods is needed. In an article published in the journal Structure, Konstanz biophysicists led by Michael Kovermann and Christine Peter (also Department of Chemistry) present a corresponding approach using complementary methods.
“Through NMR spectroscopy, for example, we get information about the dynamic properties of such proteins. Complex computer simulations, on the other hand, provide a good overview of the range of possible folds,” explains Kovermann. “So far, no general approach that comprehensively maps the dynamic and structural properties of multi-domain proteins had existed.” The researchers from Konstanz therefore devised a workflow that combines NMR spectroscopy and computer simulations, allowing them to obtain information on both properties with high temporal and spatial resolution.
Proof of feasibility included
The researchers also provided evidence that the method works: They examined various ubiquitin dimers. These consist of two units of the protein ubiquitin that are linked by a flexible bond, just like the situation in cells. It is thus a prime example of a multi-domain protein for which different structural models have been suggested so far and which is of great scientific interest.
The researchers were able to show that the ubiquitin dimers they studied exhibit a high structural variability and that this can be described in detail using the developed combination of methods. The results also explain the different structural models that currently exist of ubiquitin dimers. “We are convinced that our approach — combining complementary methods — will work not only for ubiquitin dimers but also for other multi-domain proteins,” Schneider says. “Our study opens new avenues to better understand the high structural diversity of these complex proteins that plays a crucial role in their biological functions.” More

Structural formulae show how chemical compounds are constructed, i.e., which atoms they consist of, how these are arranged spatially and how they are connected. Chemists can deduce from a structural formula, among other things, which molecules can react with each other and which cannot, how complex compounds can be synthesised or which natural substances could have a therapeutic effect because they fit together with target molecules in cells.
Developed in the 19th century, the representation of molecules as structural formulae has stood the test of time and is still used in every chemistry textbook. But what makes the chemical world intuitively comprehensible for humans is just a collection of black and white pixels for software. “To make the information from structural formulae usable in databases that can be searched automatically, they have to be translated into a machine-readable code,” explains Christoph Steinbeck, Professor for Analytical Chemistry, Cheminformatics and Chemometrics at the University of Jena.
An image becomes a code
And that is precisely what can be done using the Artificial Intelligence tool “DECIMER,” developed by the team led by Prof. Steinbeck and his colleague Prof. Achim Zielesny from the Westphalian University of Applied Sciences. DECIMER stands for “Deep Learning for Chemical Image Recognition.” It is an open-source platform that is freely available to everyone on the Internet and can be used in a standard web browser. Scientific articles containing chemical structural formulae can be uploaded there simply by dragging and dropping, and the AI tool will immediately get to work.
“First, the entire document is searched for images,” explains Steinbeck. The algorithm then identifies the image information contained and classifies it according to whether it is a chemical structural formula or some other image. Finally, the structural formulae recognised are translated into the chemical structure code or displayed in a structure editor, so that they can be further processed. “This step is the core of the project and the real achievement,” adds Steinbeck.
In this way, the chemical structural formula for the caffeine molecule becomes the machine-readable structure code CN1C=NC2=C1C(=O)N(C(=O)N2C)C. This can then be uploaded directly into a database and linked to further information on the molecule.
To develop DECIMER, the researchers used modern AI methods that have only recently become established and are also used, for example, in the Large Language Models (such as ChatGPT) that are currently the subject of much discussion. To train its AI tool, the team generated structural formulas from the existing machine-readable databases and used them as training data — some 450 million structural formulas to date. In addition to researchers, companies are also already using the AI tool, for example to transfer structural formulae from patent specifications into databases. More

A new study led by investigators from Mass General Brigham has found that ChatGPT was about 72 percent accurate in overall clinical decision making, from coming up with possible diagnoses to making final diagnoses and care management decisions. The large-language model (LLM) artificial intelligence chatbot performed equally well in both primary care and emergency settings across all medical specialties. The research team’s results are published in the Journal of Medical Internet Research.
“Our paper comprehensively assesses decision support via ChatGPT from the very beginning of working with a patient through the entire care scenario, from differential diagnosis all the way through testing, diagnosis, and management,” said corresponding author Marc Succi, MD, associate chair of innovation and commercialization and strategic innovation leader at Mass General Brigham and executive director of the MESH Incubator. “No real benchmarks exists, but we estimate this performance to be at the level of someone who has just graduated from medical school, such as an intern or resident. This tells us that LLMs in general have the potential to be an augmenting tool for the practice of medicine and support clinical decision making with impressive accuracy.”
Changes in artificial intelligence technology are occurring at a fast pace and transforming many industries, including health care. But the capacity of LLMs to assist in the full scope of clinical care has not yet been studied. In this comprehensive, cross-specialty study of how LLMs could be used in clinical advisement and decision making, Succi and his team tested the hypothesis that ChatGPT would be able to work through an entire clinical encounter with a patient and recommend a diagnostic workup, decide the clinical management course, and ultimately make the final diagnosis.
The study was done by pasting successive portions of 36 standardized, published clinical vignettes into ChatGPT. The tool first was asked to come up with a set of possible, or differential, diagnoses based on the patient’s initial information, which included age, gender, symptoms, and whether the case was an emergency. ChatGPT was then given additional pieces of information and asked to make management decisions as well as give a final diagnosis — simulating the entire process of seeing a real patient. The team compared ChatGPT’s accuracy on differential diagnosis, diagnostic testing, final diagnosis, and management in a structured blinded process, awarding points for correct answers and using linear regressions to assess the relationship between ChatGPT’s performance and the vignette’s demographic information.
The researchers found that overall, ChatGPT was about 72 percent accurate and that it was best in making a final diagnosis, where it was 77 percent accurate. It was lowest-performing in making differential diagnoses, where it was only 60 percent accurate. And it was only 68 percent accurate in clinical management decisions, such as figuring out what medications to treat the patient with after arriving at the correct diagnosis. Other notable findings from the study included that ChatGPT’s answers did not show gender bias and that its overall performance was steady across both primary and emergency care.
“ChatGPT struggled with differential diagnosis, which is the meat and potatoes of medicine when a physician has to figure out what to do,” said Succi. “That is important because it tells us where physicians are truly experts and adding the most value — in the early stages of patient care with little presenting information, when a list of possible diagnoses is needed.”
The authors note that before tools like ChatGPT can be considered for integration into clinical care, more benchmark research and regulatory guidance is needed. Next, Succi’s team is looking at whether AI tools can improve patient care and outcomes in hospitals’ resource-constrained areas.
The emergence of artificial intelligence tools in health has been groundbreaking and has the potential to positively reshape the continuum of care. Mass General Brigham, as one of the nation’s top integrated academic health systems and largest innovation enterprises, is leading the way in conducting rigorous research on new and emerging technologies to inform the responsible incorporation of AI into care delivery, workforce support, and administrative processes.
“Mass General Brigham sees great promise for LLMs to help improve care delivery and clinician experience,” said co-author Adam Landman, MD, MS, MIS, MHS, chief information officer and senior vice president of digital at Mass General Brigham. “We are currently evaluating LLM solutions that assist with clinical documentation and draft responses to patient messages with focus on understanding their accuracy, reliability, safety, and equity. Rigorous studies like this one are needed before we integrate LLM tools into clinical care.” More