At the Barcelona Supercomputing Center (BSC) – home to one of Europe’s most powerful supercomputers, MareNostrum 5 – Davide Cirillo, head of the Machine Learning for Biomedical Research Unit, applies artificial intelligence to research on cancer, rare diseases, cardiovascular, and neurodegenerative disorders. We talked with him about data-driven discovery in biomedicine and the biggest bottlenecks in AI-powered scientific research.
What’s your role in the BSC?
My research group uses artificial intelligence to address various problems in biomedicine, for example, to accelerate research on cancer, rare diseases, cardiovascular diseases, and neurodegenerative diseases. We cover the full spectrum of biomedical applications using AI, working with diverse data types, including omics data, textual and imaging data, and contributing to many European and national AI-focused projects.
What recent real-world breakthroughs have you seen in which AI has accelerated discovery in biomedical research?
We are witnessing a transformation of our research domain and the way we work. This is truly a paradigm shift.
Let me explain why. Historically, much of science has been hypothesis-driven. Researchers formulated a hypothesis and then collected data to evaluate whether the evidence supported or refuted it. This is the way a lot of scientific research has progressed in the past. In the biomedical domain, especially after the first release of the human genome in 2001, we discovered a new level of complexity, with plenty of data we could generate using high-throughput experimental techniques. This provided us with much more information than we could have anticipated through a purely hypothesis-driven approach.
This revolution started a new era of data-driven scientific research. Nowadays, it is arguably the dominant approach, not only in biomedicine and healthcare but in many other fields. For example, think about large language models and all the progress we are seeing. Those models are based on heavy data consumption and rely on high computational power. The paradigm shift consists of moving from testing predefined hypotheses toward navigating massive amounts of information to generate new insights.
We are discovering a lot, especially at the mechanistic level, about how diseases work and what underlies their manifestation and progression. This is very useful for developing new treatments. However, such technological breakthroughs come with serious implications.
First, if we produce all this data, we need bigger models and adequate infrastructure to handle it. This raises concerns about environmental sustainability and potential dependencies on private companies with the financial resources to provide these infrastructures. In this context, the Barcelona Supercomputing Center is an exemplary case: it is a public research institute with first-class computational capacity, and MareNostrum 5 is ranked among the greenest supercomputers worldwide. Unfortunately, such high standards are not always met elsewhere.
Second, when dealing with so much information, it becomes easier to lose sight of important issues, such as biases in the data, in the modeling process, or in the way models are delivered and used in society. Massive AI models are often black boxes that require deep scrutiny throughout their development lifecycle, in line with the ethical guidelines for responsible AI that the BSC promotes and applies across all its projects. So yes, the paradigm shift is powerful and impactful, but it requires careful consideration of these aspects.
After DeepMind’s AlphaFold got a Nobel Prize in 2024, have you seen more visibility and funding for AI-driven biomedical research?
Definitely yes; however, the Nobel Prize and media attention are not the only drivers. There are many channels that allow scientific developments to reach a broad audience. In the past, you might read about a new discovery in a newspaper. Now you can see scientific breakthroughs communicated directly and frequently across many platforms, even social networks. This increased visibility is accompanied by economic considerations.
Drug development has long been shaped by economic forces, with pharmaceutical companies serving as pillars of the industry. What is changing now is the entry of technology companies into the biomedical space. For example, GPUs, originally developed for graphics processing in the 1990s, have become essential tools in biomedical research. Although these technologies were not originally designed specifically for biomedical applications, pharmaceutical companies have rapidly adopted them to support AI-driven drug discovery and development.
This fosters innovation by allowing companies to explore opportunities in healthcare.
However, it is useful to distinguish three levels. The first is academic and industry research, where AI is developed and described in peer-reviewed papers and patents. The second is the market and regulatory level, where AI-based software has passed approval pathways, such as FDA clearance in the United States or compliance with the MDR in Europe. The third is actual adoption, meaning what commercial AI software hospitals are actually purchasing and using in practice.
At the market level, about 80 percent of AI products in healthcare are focused on medical imaging. Around 10 percent focus on cardiovascular diseases, and the remaining share is distributed across other medical areas. But if you look at health systems globally, hospital use remains limited, though it is increasing.
For example, in the United States in 2025, around 22 percent of hospitals already had some form of AI running. That is not yet universal, but it has been increasing year after year. The applications most frequently adopted are very practical. Ambient AI systems that record doctor-patient conversations and generate summaries are among the most widely used. Hospitals are adopting AI for efficiency and for very specific tasks.
How have large language models and foundation models changed biomedical research since 2022? Do they reshape how we generate hypotheses?
They have been very revolutionary. It is interesting that the types of data driving AI right now are mainly images and text. This is also true in biomedicine.
Large language models are very useful for processing medical notes and other textual information collected in clinical settings. This includes free-text in electronic health records and audio, which falls under the umbrella of natural language processing.
Why images and text? These data types are more widely available and easier to access. Large language models are very complex, with billions of parameters, and require large volumes of data. Developing comparable models for niche data types with limited availability is considerably more challenging.
In imaging, foundation models are becoming one-stop solutions for many tasks. You can fine-tune them for image segmentation. You can use their embeddings to generate synthetic images. A foundation model is malleable. You can reuse it for different tasks in image processing. This aligns with the broader Big Data approach, where we aim to learn as much as possible from data and then apply that knowledge in different ways. The same principle that started with genomics now applies to other high-throughput data types.
What is currently delaying progress in AI-driven life sciences? Regulation, lack of resources, lack of data?
More than regulation, which I consider necessary, the main limitation is the lack of resources, especially funding. Research is complex and requires funding. Exploration is inherent in research. When we talk about research, market, and adoption, research is the phase where you explore what is possible. This leads to innovation and new products, but it requires investment at the country and global levels. Science needs better economic support.
Regulation may delay processes, but it rightly requires safety and efficacy. Without regulation, it would be an uncontrolled environment where developments could pose serious risks. History in drug development shows this clearly. Strict regulatory frameworks were established to prevent harm. In AI for healthcare, we are seeing examples that underline the need for safeguards. The need for regulation is not a barrier to progress. It is something that enables responsible progress.
Data is another issue. We talk about Big Data, but we do not always have enough data, especially in rare diseases. Even in common diseases, as we move toward precision medicine and define subgroups, we work with smaller datasets. Producing high-throughput data is expensive. For example, technologies such as spatial transcriptomics enable high-resolution analysis of tissue and gene expression at a near-single-cell level. This is extremely powerful but also very costly.
Education and talent are also critical. We need to foster education and retain talent in research. If resources are limited, it becomes difficult to hire and train the next generation of researchers. AI and engineering have historically been male-dominated fields. More inclusive teams bring new ideas and help mitigate bias. Investing in equitable education and diversity is essential.
Is AI already speeding up drug development?
AI is already part of the toolkit in pharmaceutical research. It is used in preclinical stages and in clinical trials. For example, AI supports drug discovery by predicting pharmacokinetic properties and optimizing molecular structures. Also, it can aid patient recruitment for clinical trials, making them more targeted and precise. This is especially important for precision medicine. At the same time, these systems are not fully autonomous; humans provide ethical review, complex judgment, and regulatory adherence, as required by frameworks such as the AI Act.
Also, there is growing momentum around New Approach Methodologies (NAMs) that aim to reduce the use of animal models, including in vitro approaches, such as organoids, and in silico methods, such as the AI-based methods that we develop at the Barcelona Supercomputing Center.
What are you looking forward to in the next three to five years?
The most promising are new computational paradigms, such as quantum computing. At the BSC, we have quantum computers already installed and active work in this area. Quantum machine learning is emerging as a new field. However, real-world applications remain limited because quantum computing itself is still in its early stages of development. Much work is still needed to discover, design, and operationalize quantum algorithms.
Technological progress is driving major changes. We are seeing new developments almost daily. The acceleration is real, coming from multiple directions. But it must be matched with responsibility, resources, regulation, and education in order to realize its full potential in life sciences.
