Many computational models work on the principle of “homophily”—assuming that similarly acting drugs will yield the best results. However, this assumption may overlook the complexity of biological systems.
Sriram Chandrasekaran, Associate Professor, Biomedical Engineering, and Danai Koutra, Associate Professor, Electrical Engineering and Computer Science, are 2025 recipients of a Propelling Original Data Science (PODS) award from the Michigan Institute for Data Science (MIDAS). Dr. Chandrasekaran will partner with Dr. Koutra on a pioneering project called, “Enhancing Drug Combination Therapies Through Heterophilic Link Prediction with Graph Neural Networks,” selected under the highly competitive Track 3C: Data science and AI for health science and healthcare research.
A Shift in Drug Synergy Prediction
Drug resistance in cancer and infections has prompted increasing reliance on combination therapies, but current methods for identifying powerful drug cocktails are often slow, costly, and imprecise. Many computational models work on the principle of “homophily”—assuming that similarly acting drugs will yield the best results. However, this assumption may overlook the complexity of biological systems.
Dr. Chandrasekaran and Dr. Koutra’s proposal marks a significant departure from convention. “The key idea from this project is that maybe we should not only look for similarity —sometimes, very different things, when combined, work much better than just similar things,” Dr. Chandrasekaran said. “In biology, there are examples where when you combine very different drugs together, the results are synergistic. That’s the hypothesis we’re testing.”
The project will jointly model diverse biochemical networks, focusing on “heterophily”—the tendency for dissimilar nodes (such as drugs or proteins with distinct properties) to interact—using a novel heterophily-aware graph neural network (GNN). This fresh perspective could significantly improve the accuracy and interpretability of synergy prediction for drug combinations treating infections and cancer.
Leveraging U-M Strengths in AI and Data Science
Dr. Chandrasekaran brings years of experience in biomedical applications of AI, while Dr. Koutra contributes expertise in large-scale network analysis and graph-based AI, previously honed in industry settings such as Amazon. “Over the past five years, my group has made major strides in understanding how heterophily affects the performance of graph learning models, particularly GNNs,” said Dr. Koutra. “Beyond theory, we’ve also designed new models that are significantly more powerful under these conditions. So far, our work has been applied to large social and collaboration networks, but this project offers an exciting opportunity to bring that theoretical insight into the biomedical domain, where heterophily naturally shapes how entities like drugs and proteins interact. This domain brings new challenges, but also a chance to push the boundaries of graph-based AI and build models that could have real impact on health and medicine.”
“Danai’s work is about looking at large networks of data—treating different pieces of information as nodes, and understanding the interrelationship between them,” Dr. Chandrasekaran noted. “When my lab works on AI and drug discovery, we use off-the-shelf computational tools and adapt them to our applications. Danai is focused on developing new theoretical frameworks, treating data as graphs and modeling how information flows and connects.”
This latest collaboration aims to apply these innovative concepts to vast datasets—potentially involving 100,000 drug combinations each for infection and cancer—and rigorously test whether this paradigm shift can increase predictive accuracy from today’s typical 80% to as high as 95%. “If we can get it to 95%, that would be amazing,” Dr. Chandrasekaran said. “Then we can have very high confidence whenever the AI model makes a prediction.”
A Platform for Innovation and Translational Impact
The one-year pilot grant from MIDAS will allow the team to test these hypotheses using rich existing datasets and accelerate the pathway to larger external grants. “We already have lots of data and trained AI models, so we can hit the ground running right away,” Dr. Chandrasekaran explained. “If this works, it’s not just about drug combinations. Maybe this approach could help us understand how proteins interact, or how drugs bind to their targets—there are many directions to expand. It’s not there yet, but maybe this new idea will help get us there.”
Future Directions
“This is a great opportunity to improve the technology itself, and to bring new ideas to solve big problems,” Dr. Chandrasekaran said. “This funding encourages us to be creative and really push the boundaries.”
About the MIDAS PODS Program
Since 2016, the MIDAS PODS program has empowered U-M faculty to break new ground in disciplinary and interdisciplinary research driven by data science and AI. The grants support the formation of novel collaborations, the pursuit of high-risk, high-return research, and the acquisition of external funding to expand successful projects. As of 2024, MIDAS has already supported 94 research teams with more than $13 million in funding—sparking over $130 million in subsequent external grants.
“MIDAS is committed to supporting U-M faculty and their cutting-edge research,” said MIDAS Director H.V. Jagadish. “Through the PODS grants, we aim to enable innovative uses of data science and AI, foster new collaborations, and help researchers secure major external funding.”