# From Signal magazine: Meet the team at Land O’Lakes who are seeking to use AI to help farmers protect their crops, optimize their yield potentials and respond quickly to threats > Source: > Published: 2026-07-23 18:56:22+00:00 On the office wall behind Ava Amini, a whiteboard is covered with a dense sprawl of graphs, equations and mathematical symbols. It’s the kind of background detail a film director might use to establish, at a glance, that the scene belongs to a scientist working at the far edges of human understanding. The impression is not entirely misleading. As principal researcher at Microsoft Research in Cambridge, MA, and co-lead of Project Ex Vivo, Amini works at the meeting point of artificial intelligence, machine learning, cancer biology and biophysics. But if the science behind this role is complex, her manner is the opposite: patient, friendly and generous with attempts to translate the unfamiliar. That approach matters, because Project Ex Vivo starts with a simple question: what if cancer cannot be understood through mutations alone? For decades, precision cancer medicine has often centered on identifying mutations and other molecular changes that drive a tumor, then matching patients with targeted drugs. That approach has produced important advances and changed outcomes for many patients. But it only captures part of what makes cancer so difficult to treat. Project Ex Vivo, a collaboration between Microsoft and MIT-Harvard biomedical research center, the Broad Institute – with support from the Dana-Farber Cancer Institute – grew from the idea that cancer should be understood as a dynamic system. It is a disease shaped not only by DNA mutations, but by how cancer cells behave, change and respond to their environment. Amini describes the project as a shift from binary signals to more continuous forms of measurement. A mutation is often treated as present or absent, on or off. But cancer cells are constantly changing based on input from many biological signals. “If we expand toward more continuous measurements, we can look at how different genes or programs in the cancer are changing,” she says. “AI then helps us learn from this data at very high scale and complexity.” When Amini joined Microsoft five years ago, Project Ex Vivo was still a “very tiny seed” of an idea around modeling cancer outside the body. Together with her colleagues, Microsoft’s Lorin Crawford, Peter Winter from the Broad Institute and Srivatsan Raghavan at the Dana-Farber Cancer Institute and Broad Institute, the team began asking where they could challenge the traditional approach to mutation-based therapy and use their combined expertise to imagine something different. ## Expanding the model The project’s name points to the practical challenge at the heart of that idea. Ex vivo comes from the Latin for “out of the living” and refers, in biomedical research, to studying cells or tissues outside the body they came from. But if cancer lives in the body, why look at it somewhere else? The answer is both practical and ethical: researchers cannot test thousands of possible drugs directly in patients, and cancer studies have therefore long depended on lab models. Part of the quest for Amini and her colleagues has been to make those lab models more representative of the tumor biology they are meant to represent. Traditionally, cancer cells are often grown in a 2D layer on a plastic dish. Project Ex Vivo builds on approaches that grow cancer cells in groups or clusters, creating more complex 3D models that better preserve aspects of the biology they would have in the body. The result is a fuller picture of cancer where mutations matter, but do not explain everything. “We know that there’s so much more at play,” Amini says. “The cancer cells interact with immune cells and other cells in the tissue. They interact with each other, they respond to signals from their environment. There are all these dynamic changes that are not reflected in DNA changes.” Scientists use the term “cell state” to describe this more fluid picture, and Amini uses a simple analogy to illustrate the point. “Think of a cancer cell system like a vending machine,” she says. “There is a state that exists, and when you put money in, you choose which state you transition to next.” In this analogy, the money can be understood as the input, and the selected item as the output. In cancer, the inputs might be drugs, signals from neighboring cells, immune activity or the physical conditions around the tumor. What follows is the cancer cell’s response. “When we think of cancer cell state, it’s really capturing how the cancer is changing in response to things around it, whether that’s drugs, other cells, signals or shape in the body,” Amini says. “You can think about this input-output response, and state captures both of those.” The difference is not academic. A lab model can preserve the mutations found in a patient’s tumor, while losing the behaviors that shape how the tumor responds to treatment. In pancreatic cancer, for example, standard lab models can carry the same mutations as patient tumors, making them appear accurate. But when Broad researchers and collaborators looked at how those cells were Researchers at work at the Broad Institute behaving, in a 2021 study published in *Cell*, a different picture emerged: important cell states seen in patient tumors could be missing from some 2D lab models. Those missing states can affect how tumors respond to treatment. In some cases, the drug response seen in lab models can be very different from what happens in patients. Project Ex Vivo is working to build more reliable and scalable models by preserving, or restoring, the cell states that matter, giving researchers a better testing ground for potential therapies. Put another way: a model can look genetically right but behave biologically wrongly. ## Connecting the dots Once researchers have a model that behaves more like cancer in a patient, they can use it to test possible treatments. And this is where the scale of the problem starts to look very different. A lab might be able to check tens, or perhaps hundreds, of possible treatments. But AI allows researchers to computationally screen thousands of possibilities before deciding what to validate experimentally. “We can give information to our AI model about the pancreatic cancer, use the AI itself to screen thousands of different compounds in silico, and then it nominates a shortlist to test in the lab,” Amini says. “To do that experiment in the lab [without AI] would take years and millions of dollars. But we can do this in a matter of days.” She is quick to clarify, however, that AI does not replace experiments, but rather helps researchers decide which ones are worth doing first. This narrows a vast field of possible treatments into a smaller set that scientists can test in the lab. It has also reinforced another important lesson: that in biology, scale alone is not enough. AI models need the right biological signals, not just more information. Amini says Project Ex Vivo has produced learnings about “what data is most important” and how to generate “the right type of data to give good signal to the models” about cancer biology. Recent work from the Ex Vivo team has explored the same question from the AI side, looking at how the composition, size, diversity and handling of underrepresented cell states in training data can shape how well models perform. The emphasis on useful biological signals extends beyond Project Ex Vivo. Across Microsoft’s healthcare and life sciences research, AI is being applied to cancer from pathology images to chemistry and drug discovery. One strand of that work focuses on finding richer ways to see and interpret tumors. The Virchow family of pathology foundation models, developed by Microsoft Research in collaboration with Paige, enable accurate detection of both common and rare cancers as well as a variety of downstream tasks, such as determining a cancer’s subtype. GigaTIME approaches pathology from another angle, using AI to translate routinely available pathology slides into virtual multiplex immunofluorescence images – essentially richer virtual images – that can help researchers study the tumor immune microenvironment, including how tumors and immune cells interact inside tissue. This connects with Amini’s broader view of cancer modeling. “Our vision is to think about how all these different kinds of data connect,” she says. Pathology and microscopy, she adds, are all “different ways to view cancer.” In foundational AI models the team is developing, “we can connect those imaging modalities, like pathology, to measurements of cancer cell state.” The goal, she says, is to “learn a landscape” of the tissue, including “what the spatial organization of a tumor looks like.” If Virchow and GigaTIME point to how AI can help researchers see tumors more clearly, other Microsoft research explores how biological insight might be turned toward treatment. Amini points to TamGen, a Microsoft model focused on target-aware molecule generation, as one example of how different strands of AI research could eventually connect. “At Microsoft, we have research in chemistry, and we have research in biology, like in Ex Vivo, but the ultimate goal is to bridge the two,” she says. For Amini, that link between chemistry and biology is central: how a potential therapy might affect a particular cancer cell state, or how researchers might work backwards from a desired change in cancer behavior toward a possible treatment. She says new AI models the Microsoft team is developing are aimed at exploring these relationships, connecting questions about cancer cell state with questions about potential therapies. RetroChimera, another Microsoft initiative, explores a different part of the same broad challenge by helping chemists think through how a promising drug-like molecule might actually be made. Taken together, it’s clear that the path forward in cancer research cannot depend on any single view of the disease. These projects reflect a broader Microsoft approach, also outlined in a recent *Cell* perspective paper by Microsoft researchers, co-authored by Crawford, Amini, and Microsoft colleagues: connecting cell state, pathology, chemistry and drug discovery to build a more integrated understanding of how cancer behaves and how it might be treated. ## Pathways to patients For all the promise of connecting these strands of research, the route from discovery to patient care is still hard. Promising findings must pass through a maze of preclinical validation, clinical trials, regulatory review, manufacturing, cost constraints and the realities of health care delivery. The use of AI adds its own complications, particularly around data access, provenance and privacy, as progress relies on connecting biological, chemical and clinical data without weakening the protections around them. For Amini, the challenge is knowing which boundaries should be respected, and which silos may be slowing progress unnecessarily. “Where do barriers exist for real reasons that should be respected,” she asks, “versus how can we potentially break down some of those silos to enable effective translation?” Nevertheless, she is excited by AI’s potential to accelerate discovery. “AI can be a tool for experimental scientists,” she says, helping them move toward possible treatments “more effectively and with higher confidence.” Too many drugs fail in what the industry calls the “valley of death,” she says, and “if we can build AI models that help us traverse that valley, that’s really powerful.” Amini’s excitement about the work goes back to an early fascination with biology and nature. She recalls being “super obsessed” by the living world from a young age, “trying to understand why nature operates in the way it does.” Watching wildlife presenters such as Jack Hanna and Steve Irwin as a girl, she started asking deeper questions about biology and came to see math and physics as powerful ways to understand the world. Later in her education, computer science, machine learning and AI opened up another way to interrogate biology. “These are powerful tools that we can leverage to discover the fundamental underpinnings of life, design new therapies and engineer biology in an intelligent way.” By the time she was a PhD student, she says, one of her main goals was clear: to translate research into real-world impact for human health. “I worked in a very translational lab where it was all about preclinical development, looking at how this could make a difference in the life of a patient. So that aspect is super important to me.” That focus on impact makes her cautious about overpromising. AI models will not become the doctor or replace human-led lab experiments. “We’re developing these methods thoughtfully and collaboratively, and through partnerships where we can complement other forms of research,” she says. “It’s really about enablement more than anything.” Asked whether she is both amazed by the possibilities of her work and overwhelmed by the deluge of data and parameters at play, she laughs in agreement. “Yes, but I’m also motivated 100 percent of the time by the possibilities of what we’re working toward.”