Deploy, Prototype, Bug: Interview with Dr. Xin Tang

Writing by Jonathan Dorogin from the Blakney lab, Michael Smith Laboratories

The Michael Smith Laboratories is a truly unique research unit, one that marries biology with many technological disciplines. In this interview, we speak with Dr. Xin Tang, an Assistant Professor jointly appointed in the Michael Smith Laboratories (MSL) and the Department of Computer Science, who exemplifies the transdisciplinary approach of the MSL. We ask Dr. Tang to reflect upon his early career success and provide some outlooks about the future in an exercise of Rose, Bud, Thorn – or more appropriately for his field, Deploy, Prototype, Bug. 

Jonathan Dorogin: Could you start by describing the central question your lab is trying to answer?

Xin Tang: The question I keep returning to is whether we can build computational models that do more than describe biology after the fact. We want models that can predict how cells and tissues change, explain why those changes happen, and help identify what experiment should be done next.

My lab studies living systems across scales, from single-cell molecular profiles, to spatial tissue organization and dynamic cellular behavior. We develop explainable and interpretable AI methods that relate molecular signals to changes in cell state, function, and fate, as well as animal behavior. The long-term goal is to place these models in a closed loop with experiments: the model makes predictions, experiments test them, and the new data improve the model. Much of our current work focuses on the brain, aging, and cancer.

Deploy

JD: In computing, “deployment” is the moment when an idea has to work in the real world. Is there a direction in your lab that feels like it has started to move from concept into practice?

XT: I think of deployment less as a single result and more as the point where a computational idea begins to shape biological questions and experimental design.

One example of this is self-driving discovery, particularly through our conversations and collaborations with Dr. Sabrina Leslie. Her expertise in live-cell imaging makes it possible to study dynamic cell behaviour, rather than static snapshots. In that setting, a model could observe how cells respond over time, update its understanding, and help determine next steps.

Another example is our work on dementia and aging with Drs. Freda Miller, Haakon Nygaard and Brian MacVicar. We are interested in why some individuals show more resilience than expected despite Alzheimer’s-related pathology. The goal is to model the expected decline and then study the individuals and their cells and tissues that deviate from that expectation.

Finally, we collaborate with Drs. Martin Hirst, Torsten Nielsen and Michael Underhill on cancer genomics and virtual cell modeling, as AI can help integrate complex genomic and single-cell data. All of these examples are biologically different, but they share a common principle: AI becomes most useful when it is grounded in experimental reality. We need our models to produce predictions that experimental biologists can interpret and test.

JD: Many people talk about bringing AI into biology. What makes this integration scientifically interesting?

XT: Biology is like blind men touching an elephant – it rarely gives us a complete picture of a system. We often observe one part of a process; a molecular snapshot, a spatial pattern, an environmental response, or a change over time. What makes AI interesting to me is the possibility of connecting these pieces together and asking whether they point to a coherent biological mechanism, though the model has to remain accountable to biology. It should tell us what it is relying on, where it is uncertain, and what kind of experiment could challenge the prediction. In my lab, that means building models that can point to the relevant cell type, molecular program, spatial niche, or temporal process behind a result. I am most interested in AI when it helps us move from data to mechanism, and guide us from mechanism to the next experiment. That is when it becomes scientifically useful.

Postdoctoral Fellow Jonathan Dorogin (left) sits down with Dr. Xin Tang (right) to discuss his research and career.

Prototype

JD: Your lab is still relatively new. What has the transition to being a principal investigator taught you?

XT: It has taught me that building a lab is itself a scientific project. As a trainee, you are often responsible for making one project work. As a PI, you are responsible for creating an environment where many people can develop their own scientific judgment. My students come from different backgrounds, and one of the most rewarding parts of the job is helping them develop scientific taste. That means learning to ask when a method is technically sound, when strong computational performance is meaningful, and when a biological question is worth pursuing. Our work depends on translation between fields: a biological question has to be framed in a way that computation can address, and a computational result has to be interpreted in a way that biology can test.

Bug

JD: Every fast-moving field has bottlenecks. What is a “bug” in computational biology right now that researchers still need to solve?

XT: One major bottleneck is biological grounding and interpretability. Models are getting larger, datasets are becoming more complex, and computation is becoming more powerful. The hard part is connecting a prediction to a biological explanation. If a model predicts disease progression or treatment response, we need to know what is driving that prediction, whether it’s a specific cell type, microenvironment, or something else entirely. Without that level of interpretation, even a high-performing model may not be very useful.

Another challenge is separating signal from noise. Biological systems are noisy, and modern datasets are often incomplete (missingness), have variable responses (batch effects), and have hidden computational structure. We need methods that can identify patterns that are reproducible and experimentally meaningful, without overinterpreting every correlation. Ultimately, the field needs tighter loops between modeling and validation.

JD: The field also moves very quickly. How do you plan research when AI methods may change by the time a paper is published?

XT: I think the key is to build around durable scientific questions, not around a single method. Methods, architectures, and benchmarks will change, but questions like how cells choose their fate, how tissues stay resilient, or how cancer cells adapt to therapy will remain important. We should adopt new methods quickly when they are useful, but the long-term direction should remain stable: predictive, interpretable, experimentally grounded models of living systems.

JD: Before we close, is there anything else you would like readers to know about your lab?

XT: I am excited to build this program at UBC. The questions we care about require collaboration across biology, computation, and engineering, and the MSL is a rare environment where those conversations can happen naturally. I also want prospective trainees to know that our lab welcomes people who are eager to work across boundaries, and no one needs to arrive fluent in every scientific language. What matters is the willingness to learn from other fields, to fearlessly ask questions, and to stay curious and rigorous when the science becomes difficult. My hope is to build a lab where people feel supported as they grow into researchers who can move confidently between data, AI models, experiments, and biological mechanisms.