22 May 2026
The new Multimodal AI Across Scales Research Programme, within Human Technopole’s Computational Biology Research Centre, develops AI methods that connect biological data across modalities and scales, from molecules and cells through tissues and organs to patients and populations.
Led by Florian Jug, the Programme combines machine learning, generative and probabilistic modelling, computer vision, and bio-image analysis with a strong commitment to uncertainty-aware, open, and FAIR tools. Its ambition is straightforward: AI as a trusted partner for biomedical research, and, in time, for human health.
Florian, why is Human Technopole launching this new Programme now? What scientific need is it designed to address?
Biomedical research is changing on two fronts at the same time. The first is data: we can now generate cryo‑EM structures, spatial transcriptomes, single‑cell atlases, large clinical cohorts, and longitudinal electronic health records at a depth that, even a decade ago, would have looked unrealistic. The second is AI: modern methods have finally become powerful enough to start making real sense of that data.
The trouble is that these two transformations have been growing up in separate rooms. There are excellent AI tools for microscopy. Other excellent tools exist for single‑cell omics, for spatial transcriptomics, or for clinical risk prediction. Biology, however, does not respect those walls. A disease typically begins as a molecular event, propagates through cells and tissues, manifests in an organ, and is shaped by the rest of a person’s life. To understand and intervene along such a trajectory, we need AI that can move with the biology — across data modalities and across scales.
Closing that gap is exactly what the Multimodal AI Across Scales Programme is about. And this is the right moment to do it: the methods we need — uncertainty‑aware models, multimodal representation learning, physics‑informed and causal AI — are maturing fast and, frankly, they want to be combined. Our ambition is that AI becomes a genuinely trusted partner for biomedical scientists, at HT and beyond, and the conditions to build that partnership are finally in place.
What is the major biological question your Programme aims to help answer in the coming years?
We want to be able to predict, with calibrated uncertainty, how molecular and cellular events shape disease at the level of the patient, and to do this across many scales at once.
In concrete terms: given a cryo‑EM structure, a spatial transcriptome, and a patient’s clinical trajectory, can we identify which interventions are most likely to alter disease progression, and by how much? That is the umbrella question. Underneath it sit specific questions across our research portfolio: how the molecular architecture of cilia connects to ciliopathies; how breast tissue imaging and patient history together encode cancer risk; how histology, spatial transcriptomics, and genomics together predict disease progression. These are very different application areas, but they share the same underlying methodological challenge, and that is the challenge we are eager to tackle.
HT’s Strategic Plan focuses on multi‑scale research, from molecules to populations. Where does your Programme fit within this vision, and what bridges will it help build?
Our research sits at the methodological heart of HT’s multi‑scale vision. We do not own a single biological scale; we build the AI that lets the different scales talk to one another.
Concretely, we bridge structural biology and cryo‑EM to spatial omics and advanced light microscopy; we bridge those to single‑cell and population genomics; and we bridge all of that to the clinical and health‑data cohorts that anchor HT’s translational ambitions. In close synergy with HT’s Scientific AI Flagship, we provide the shared methodological foundation — uncertainty‑aware models, multimodal latent spaces, physics‑informed and causal AI — that the rest of HT can then plug into. The aim is to give all the biomedical research at HT a shared scaffolding.
HT is, frankly, the ideal home for this kind of agenda. It is small enough to still be flexible and agile — decisions can be quick, and people across Centres actually know each other well — and at the same time it spans the full sweep of biomedical expertise we need, from structural biology to whole populations, under one roof. That combination is rare, and in this case, golden!
What new expertise, technologies or ways of working will the Programme bring to HT’s research ecosystem?
Three things, mostly. First, deep expertise in the modern AI toolkit: Transformers, state‑space models, diffusion and flow‑based models, variational and causal methods, explainable AI and counterfactuals. All of this, though, is applied to biomedical problems, so that we can help speed up scientific discovery for our colleagues at HT, in Italy, and worldwide.
Second, a serious commitment to uncertainty. We treat “what the model does not know” as a first‑class output. In our experience that is what turns AI from an unreliable oracle into something a scientist, or a clinician, can actually work with: AI as a trusted partner!
Third, a way of working that takes open, FAIR, professional‑grade software seriously. Through our involvement in, for example, AI4Life and the Bioimage Model Zoo, we already know how to turn research code into tools that other scientists use day to day. We will measure ourselves not only by publications, but by the methods that get adopted by the wider community.
Collaboration across disciplines is central to HT’s mission. Which Research Centres, National Facilities or Flagship Research Programmes do you see as the most natural partners?
Almost all of HT, honestly — which is the whole point. But to name the most natural daily partners: the Structural Biology Research Centre and the National Facility for Structural Biology, where we are integrating cryo‑ET with spatial omics in the context of ciliary disease; the Genomics and Neurogenomics Research Centres, with whom we plan to build multimodal atlases combining histology, spatial transcriptomics, and whole‑genome sequencing; the Health Data Science Research Centre, which connects us to large clinical cohorts and multimodal approaches to disease risk prediction; the National Facility for Genome Engineering and Disease Modelling and the National Facility for Light Imaging, where image‑based readouts feed directly into our models. Historically, I have built popular methods and tools for microscopy image data analysis, and in my own Group we will very much continue this line of research.
On the Flagship Research Programme side, the Scientific AI Flagship is our closest partner by design. The Ciliopathies, Cardiovascular and Metabolic Diseases, and Immunogenomics, Cancer and Infections programmes are where our methods, and the methods of other AI‑heavy groups at HT, will be tested against the real biomedical questions that matter.
For people outside your field, it can be difficult to see why basic research matters for human health. Can you give one concrete example of how your Programme’s work could open new opportunities to understand, prevent or treat disease?
Take breast cancer screening. Today, mammography gives one snapshot of the breast; an electronic health record gives a long list of risk factors; and, when available, genomic and lifestyle data sit in yet another silo. A clinician has to integrate all of that in their head, often under time pressure.
We are working on AI models that integrate these different data streams natively and, very important to us, can report how confident they are about their predictions. Instead of a single number, the system can say: “Based on this woman’s imaging, her family history, and her trajectory over the past ten years, her five‑year risk is X, with uncertainty Y, and the features in the image that drive that prediction are these.” That kind of output is what lets a radiologist trust AI as a partner, rather than as a black box.
The same logic — multimodal integration, calibrated uncertainty, explainability — applies to neurodegeneration, to ciliopathies, and to many other diseases where, today, we underuse the data we have or could decide to acquire.
What kind of people do you want to attract to your Programme? Which scientific skills, but also which attitudes, will be key to building the team?
On the skills side, we are looking for strong AI and machine learning foundations — modern deep learning, probabilistic modelling, optimisation, causal inference — paired with real biomedical curiosity. We also need outstanding research software engineers: people for whom building robust, open, reusable tools is a craft, not an afterthought.
On attitude, three things matter to me. First, intellectual generosity: the willingness to spend time understanding a collaborator’s data before reaching for a model. Second, patience: meaningful biomedical AI does not always produce a paper every six months, and we are working on a longer horizon. Third, a builder’s mindset: people who care that their work is not just published but actually used by others. If those three things describe you, you will fit in here, regardless of which discipline you come from.
I have recently put the “mindset” we live by into 20 ISMs. If you are interested, check them out!
Looking ahead, if we met again in five years’ time, what result, change or new capability would you hope this Programme had made possible, for HT and for the wider scientific community?
I would like to be able to point to two kinds of impact.
First, on the science. In at least one disease area — most plausibly breast cancer risk, or maybe a ciliopathy — our multimodal, uncertainty‑aware models should be part of a working pipeline that has changed how the question is approached, both inside HT and beyond. Not as a demo, but as part of how scientists and clinicians do their work.
Second, on the ecosystem. I hope to see a set of open, well‑engineered AI tools used routinely by labs across Europe and beyond. Used because they are good, and because they enable real research. If, five years from now, a researcher in Milan, Munich, Madrid, or Mumbai is solving a biological problem with a method or tool built by us, then we have done our job.
