BIO4 Campus AI (Life Sciences) Lead Stefan Milosevic spoke on the opening panel of the Belgrade Bioinformatics Conference, BelBi 2026, on one of the field's most contested questions: Human Digital Twins: Are We Ready?

His answer reframed the question. A faithful virtual replica of an entire patient is the wrong goal  and chasing it is much of the reason digital twins disappoint clinicians. The twin that earns the name is narrow and falsifiable. It tells a doctor one thing the scan cannot, and it can be proven wrong.

Grow-twins versus go-twins

Stefan introduced a distinction that ran through his entire contribution. A grow-twin mirrors visible disease and tracks how it changes. It is easy to validate, which is also why it is of limited clinical value. A go-twin predicts the unobserved — and that is where the real diagnostic and therapeutic potential lies.

He grounded this in glioblastoma research: the tumor that kills is not the visible mass. It is the infiltrating cell population that has already moved past the surgical margin, invisible on imaging, and that is where the disease returns. The tumor we see is not the tumor we fight.

Ground truth, not more data

The field is rarely data-limited — it is ground-truth-limited. The hard part is not gathering more data but integrating what already exists and agreeing on the labels against which predictions are checked.

On trust, his position was equally direct: trust is not earned by a higher accuracy number, it is earned by falsifiability. A model has to commit to predictions that can be proven wrong. A prediction that does not change a clinical action is a vanity metric.

On the concern about black-box models, he was blunt: explainability is the wrong bar. What matters is how well a prediction is validated and how reversible the action it triggers is.

Finding a research edge

Later that day, Stefan delivered an educational session — Finding Your Research Edge: From Cambridge to the Future of Life Sciences and AI — speaking to students about finding a genuine research fit rather than a fashionable one. Drawing on his own path from Serbia to Cambridge, he used the digital twin as a thread: where the concept comes from, what it is beginning to make possible, and neuro-oncology as the use case he knows best — where the gap between the disease we can image and the disease that drives the outcome is at its widest.

That gap between the disease we can see and the disease that decides the outcome is exactly the kind of problem the work at BIO4 Campus is built to take on.