Andrea Townsend-Nicholson: From Building Virtual Humans to Simulating the Simulator

Andrea Townsend-Nicholson’s career has spanned the globe as she explores turning biology into code. An experimental molecular biologist and Vice Dean for Health in University College London’s Faculty of Life Sciences (Division of Biosciences; Research Department of Structural and Molecular Biology), she uses high performance computing (HPC) and simulations to narrow hundreds of experimental possibilities to the few that matter. This trajectory now points toward a full-fidelity virtual human, a safe space for clinicians to safely “test-drive” high-stakes decisions before a single intervention.

At SC25, Andrea co-leads the Digital Twins Workshop, a half-day event that first surveys digital twins across HPC, and pivots to focus on twins of HPC data centers, exploring power and cooling strategies, network behavior, and other routes to efficiency and reliability for next-generation infrastructure.

Igniting HPC

Ahead of the workshop, the SC25 Communications Team caught up with Andrea to talk about digital twins, what HPC is igniting, and why “simulating the simulator” is an important frontier.

Andrea Townsend-Nicholson

Professor of Biochemistry & Molecular Biology, University College London, Structural and Molecular Biology

Q: Hi Andrea! To start, it’s always interesting to learn how people got into HPC. Can you please introduce yourself and your path into computation?

Andrea Townsend-Nicholson: I’m an experimental molecular biologist by training, and my career has been happily itinerant; Canada for my bachelor’s at the University of Toronto, France for a PhD, Australia for post-doctoral work, and now the UK at University College London, where I’m Vice Dean for Health in the Faculty of Life Sciences (Division of Biosciences). My research centers on proteins that enable cells to communicate. These proteins are dynamic. They shift conformation depending on partners and context, which historically made it very hard to capture their structure in a native-like environment.

Over the last decade, enough structural snapshots accumulated that computation became a practical guide for experimentation. I can now model how large complexes might assemble, how membrane proteins and their complexes change conformation, or how a candidate drug might bind, and then take the most plausible options to the bench. That shift has changed my lab’s tempo: simulations give us sharper hypotheses, save time and materials, and open doors to questions we simply couldn’t afford to explore experimentally first.

Q: When did a full-fidelity digital twin of the human body go from sci-fi to achievable?

Townsend-Nicholson: The realization came as colleagues began coupling models that used to live in isolation. You see a digital heart beating, digital blood moving through a digital vasculature. These components, once linked, started to behave like a system. That’s when it clicked for me: build validated parts, then integrate them into higher-order structures. About five to eight years ago I felt the feasibility shift. The computing looked daunting then, but with exascale machines emerging and AI/ML methods accelerating, simulations that were once impossible or prohibitively expensive are now within practical reach.

We’re not at an end-to-end virtual human yet. Today, we can build credible organ and process-level twins. The next stage is integration: making those parts talk to each other reliably enough to guide clinical decisions.

Q: Define a “digital twin” in one sentence, and why is it now pivotal?

Townsend-Nicholson: A digital twin is a computational model of something in the real world that behaves as the real thing does. It’s pivotal now because three curves have crossed: far more computing power is available, methods are broadly accessible, and awareness is spreading. Success in one domain propagates quickly. Once people see a twin work for aircraft or a heart, they look to adapt the approach to their own problems. Crucially, fidelity should be fit-for-purpose; the model’s detail and scope should match the decision risk and timelines.

Q: What changes when clinicians can safely “test-drive” treatments on a patient’s twin?

Townsend-Nicholson: You replace guesswork with rehearsal. A powerful example is the conjoined-twin surgeries at Great Ormond Street Hospital in London. Surgeons build detailed digital models to map anatomy, explore “what if we alter this vessel?” and sequence the procedure. They can run multiple scenarios, discard the ones that lead to trouble, and converge on a plan that maximizes safety for both patients. If something fails in simulation, nobody is harmed. You reset and try another approach. That ability to practice, stress-test, and learn in silico translates to better real-world outcomes.

Q: What does modern supercomputing make possible for digital twins?

Townsend-Nicholson: Early in model development, we rarely know which parameters truly drive behavior, so we have to sample widely; ensembles, sensitivity analyses, uncertainty quantification. And then iteratively refine until the model mirrors reality, whether that’s a protein, a heart, or a coupled multi-organ process. Doing this credibly and quickly takes scale. You won’t run every patient’s personal twin on an exascale system, but you do need that class of machine to discover governing parameters, validate model fidelity, and establish boundaries. Once those are known, you can prune the model to something that runs on more modest infrastructure for day-to-day clinical decision support. We pair that work with verification and validation against experimental and clinical benchmarks, and we quantify uncertainty so decision-makers understand confidence bounds.

Q: Do you have a favorite success story from the field?

Townsend-Nicholson: Personalized medicine for ultra-rare diseases. Traditional trials need large cohorts; a person with a one-in-the-world condition will never meet that bar. In silico clinical trials change the equation. You can build a virtual cohort and test likely treatments computationally. Colleagues in places like Oxford and Barcelona are already using this approach, including screening for potential cardiac toxicities before anyone takes a dose of any treatment. It reduces reliance on human and animal testing, accelerates timelines, and still demands rigorous verification, so the confidence isn’t a shortcut; it’s earned.

Q: Beyond medicine, which digital twin applications excite you most?

Townsend-Nicholson: Transportation and weather rise to the top for me. Imagine a transportation digital twin that ingests real-time conditions and your personal context to prevent traffic jams, optimize routing, and make door-to-door journeys reliably predictable. On the climate side, better severe-weather twins could inform everything from hurricane response to where we should (and shouldn’t) build. Aviation operations can benefit for similar reasons, including optimizing airspace use and smoothing traffic flows. Ultimately, we all want to get where we’re going quickly and stay safe and dry.

Q: AI vs. physics-based simulation, how do you choose the tool?

Townsend-Nicholson: It depends entirely on the question. AI excels at pattern recognition over vast datasets; image triage for skin lesions is a good example because there’s so much data. But if a neurosurgeon is placing a stent in your specific blood vessel, you want precise calculations tailored to your specific anatomy. There’s a zone of overlap, but the more unique the case, the more detailed, mechanistic modeling matters. Use the tool that matches the accuracy and specificity the decision demands. Often, the best results are hybrids, using machine learning for parameter estimation or surrogate models wrapped around a mechanistic core.

Q: Bidirectional data raises hard questions. What ethical concerns keep you up at night?

Townsend-Nicholson: First, it’s essential to involve people, especially patients, in shaping twins that will affect their care. The model has to do what they need. It should be explained in plain language (including its limits and appropriate use), and they need to feel that it’s usable and trustworthy. Ethical handling of sensitive medical data is obvious. There is also sustainability to consider. These systems are energy-hungry. We may need governance models like those used for major scientific facilities, with proposal-driven access where people justify the compute they’re requesting and how they’ll use it. Scarce resources should be prioritized transparently and equitably, with clear criteria. Trustworthy twins also depend on clean, well-annotated longitudinal data with clear provenance and adherence to standards

Presenters from the SC24 Digital Twins workshop gather after a successful session.

Q: The workshop dedicates an afternoon to twins of HPC data centers. Why is “simulate the simulator” an important frontier?

Townsend-Nicholson: Because data centers are complex ecosystems, much like the body. A trusted twin lets you explore the “what if” questions with power and cooling strategies, understand network behavior under stress, and, as the workshop will explore, even consider AR/VR for operations. When I first heard the idea, I smiled. It’s the same logic we apply elsewhere: if a twin gives you actionable predictions and smoother operations, why wouldn’t you build one for the very infrastructure running our science?

Q: SC25’s theme is “HPC Ignites.” How do digital twins ignite collaboration across academia, industry, and government?

Townsend-Nicholson: Think Lego blocks. We’re building validated components, hearts, vessels, skeletal systems, that can be composed into larger twins. Academia contributes rigorously tested parts; industry repurposes them for practical systems like national logistics or utilities; government brings scale and mission problems. The catalyst is combinability across sectors.
Once you think in those terms, there’s no reason why digital twins initially built for one application, say, academic biomedical research, can’t be repurposed and integrated into industry-level solutions, like optimizing national transportation or logistics networks. Imagine having something akin to a “pick-and-mix” library of digital twins. This concept is exciting because it prevents redundancy. You don’t need to reinvent components that already exist, are tuned, and have been validated.

Q: For early-career readers: where should they start?

Townsend-Nicholson: Decide whether you’re primarily drawn to the domain or the technique. Go deep in one, but keep sight of the other. My domain expertise lets me ask better questions of computational collaborators, down to details like, “Why are we simulating at 300 Kelvin when the human body isn’t?” and “What changes if we reflect actual physiological conditions?” The most rewarding work lives at those interfaces; experimentalists working with modelers, engineers with clinicians. Pick a starting pillar, then build bridges.

Q: It sounds like cross-disciplinary collaboration is the norm. What does that look like in practice?

Townsend-Nicholson: It’s diverse teams where no one holds all the pieces: computer scientists and code builders, HPC facility experts, experimentalists who can validate predictions. The key is listening and understanding how each person approaches problems, what they’ve tried, what failed, and where the gaps are. Increasingly, people span more than one area, which makes those conversations richer. That’s how complex twins get built.

I’d especially love to see the arts, social sciences, and humanities embrace digital twins alongside the sciences. New kinds of users bring fresh questions and constraints that can, and should, reshape future machine architectures and software stacks to match their needs.

Q: Could twins also inform computing architectures, and how are you bringing new users into HPC?

Townsend-Nicholson: I’m not a computer scientist, but I care deeply about using the machine well and teaching others to do the same. This year, we put 120 undergraduate students (around 17 years old) on a UK supercomputer and taught them replica-based ensembles that fully utilize nodes. On 128-core nodes, we organized groups of 16 students × 8 replicas each, no wasted cores, and explained why that mapping matters. The facility staff were excited to see it run at scale; the students were thrilled (a memorable machine name and photo online helps!). We received overwhelmingly positive feedback and are planning to expand to accommodate the demand. The point is to demystify large-scale systems and make good practice repeatable.

Q: What should attendees expect at the SC25 workshop, and what equals success?

Townsend-Nicholson: Expect both familiar and novel. We’re planning “Bring Your Own Digital Twin” lightning talks. Last year included a wonderful art-history twin that modeled how light filled an ancient palace, which is the sort of question you can’t answer by simply “turning on the lights.” If you have a quirky or cross-domain twin, bring it! We’re actively soliciting short lightning-talk proposals from across disciplines; reach out to the session chairs. You’ll see examples close enough to engage you, and different enough to spark new ideas. Success is people leaving energized, following up, and returning to present their own twins at future editions of the workshop.

Simulation of a potential anti-cancer drug binding to its target (HDAC8)

Fluid-electro-mechanic model of the human heart (simulated with Alya)

Digital Twins at SC25

Join the Digital Twins Workshop. Bring your questions, your datasets, and your “what‑ifs.”