Gastone Castellani is Full Professor of Physics and Biophysics at University of Bologna, where he combines his extensive background in physics of complex systems and biology. He is also a renowned figure in the field of Artificial Intelligence (AI) applied to medicine and healthcare, having spent 20 years at Brown University in the US during the 70s collaborating with other prominent AI researchers, including Leon Cooper and John Hopfield (both recipients of a Nobel Prize in Physics).
As part of our new series of interviews with GenoMed4All partners, Gastone reflects on the role he’s fulfilled during the project and his expectations for its impact on advancing precision medicine for hematology after the project ends in June 2025.
Work package leadership and interdisciplinary collaborations
In GenoMed4All, I’ve been leading Work Package 6, which is dedicated to developing artificial intelligence algorithms for the project. This work package has brought together a diverse and highly collaborative group, and our focus has been on designing AI models for both tabular data — like genomic information — and medical imaging.
We’ve worked across three use cases: sickle cell disease, multiple myeloma and myelodysplastic syndromes. Alongside our partners, we’ve developed a range of algorithms, and we’ve made our software openly available through a shared web repository for the benefit of GenoMed4All collaborators and the wider research community.
One of our most significant achievements has been the implementation of federated versions of these algorithms. Not all models are suitable for federated learning, but for those that are, we’ve successfully deployed and trained them within a federated network of GenoMed4All partners. This decentralized approach — where algorithms move instead of sensitive patient data — is one of the hallmarks of the project and a critical step forward in privacy-conscious AI development.
Impact on precision medicine for hematology
Precision medicine is an ambitious and evolving field, and while no single project can address all its challenges, I believe our contributions offer tangible progress.
For example, in the context of Sickle Cell Disease, which can be considered as a relatively underexplored condition, we developed a novel image analysis algorithm to detect “silent infarcts,” or white matter hyperintensities, using MRI data. These are very small lesions that are often hard to detect and easy to confuse with other brain regions. Our model, validated on over 500 MRI scans, has been well received by neuroradiologists, and we’re now submitting our results to a major journal.
In the area of Myelodysplastic syndromes, we’ve collaborated closely with other GenoMed4All partners to identify a genomic signature that can predict the risk of progression to acute myeloid leukemia, marking an important clinical milestone.
And in the case of Multiple Myeloma, we’ve combined genomic data with imaging techniques like PET and CT scans. This integration has led to more accurate predictions of overall and leukemia-free survival, and we’re currently preparing a publication on these findings.
It’s this kind of multimodal, integrative approach that I believe holds great promise for the future of personalized medicine.
The legacy of GenoMed4All going forward
One of the biggest challenges we faced was accessing data. In my experience, this is a recurring issue in many EU projects: it often takes two years or more — nearly half the project’s lifetime — to secure the data necessary for analysis.
That said, I’m hopeful that one of the lasting legacies of GenoMed4All will be the federated infrastructure we’ve established, which should significantly reduce these delays in future projects. This setup allows us to train models where the data resides, without needing to physically move sensitive information, something that aligns well with modern data governance principles.
Another important takeaway is that the approaches we’ve developed for hematological diseases can be extended to other areas of medicine, such as lung cancer, diabetes and neurodegenerative disorders. Hematology, much like neuroscience, is at the frontier of precision medicine, and the tools we’ve built here are widely applicable beyond the original scope of the project.
Looking ahead, I’m also excited about emerging fields such as synthetic data generation and digital twins. Through initiatives like the EU-funded project SYNTHEMA, we’re already exploring how to generate and validate synthetic datasets, which could provide an alternative when real patient data is hard to access; and digital twin technologies could enable us to model individual patient trajectories in-silico, not just in hematology, but across multiple disease areas. Thanks to GenoMed4All, we already have the federated software and infrastructure in place to support these next steps.