A Conversation with Dr. Adeel Razi (OHBM 2025 Keynote Interview Series)
Author: Simon R. Steinkamp
Editors: Kevin Sitek, Joon-Young Moon, Alfie Wearn
Next up in our Keynote Lecture Interview series is Dr. Adeel Razi, who is a Professor of Computational Neuroscience at the School of Psychological Sciences at Monash University in Australia. His laboratory is affiliated with the Turner Institute for Brain and Mental Health and Monash Data Futures Institute.
Dr. Razi is a computational neuroscientist with a background in engineering. He leads a highly cross-disciplinary laboratory, using methods from physics, engineering, and computer science to understand the functioning of the brain. A main theme of his group’s research is computational modeling of the brain using methods such as dynamic causal modeling to understand how different parts of the brain work together and how diseases or psychedelics alter and disrupt these processes. Their research further focuses on developing neuroscience-inspired machine learning methods to further our understanding of how the brain performs reasoning, learning, and planning.
In 2002, Dr. Razi received a B.E. degree in Electrical Engineering (with a University Medal), from the N.E.D. University of Engineering & Technology in Pakistan. He then obtained a degree in Communications Engineering from the RWTH Aachen University in Germany, and a PhD in Electrical Engineering from the University of New South Wales, Australia in 2012. After finishing his postdoctoral studies (2012-2018) at the Wellcome Centre for Human Neuroimaging, University College London, UK — where he worked together with Karl Friston and Geraint Rees – he returned to Australia, first as a Senior Research Fellow at Monash University, where he became a Professor in 2024.
We are greatly honored that Dr. Razi took the time to talk with us about his career path and the motivation behind his work in psychedelics and computational modeling.
Simon R. Steinkamp [SRS]: Could you tell us a little bit about your scientific career so far?
Adeel Razi [AR]: I'm an electrical engineer by background. Initially, I was interested in integrated circuit design, so I started in micro electronics. Then I got interested in systems engineering, and did my master's and PhD in wireless systems – I was developing ways to transfer information from one place to another wirelessly. This is an area in communication and information theory, combined with signal processing. After I finished my PhD in Sydney, I did an industrial job at Broadcom Corporation, developing Wi-Fi chips. As I tell people, whenever you're sending an email or browsing the internet using your iPhone, you're using some part of the code that I wrote. I only worked in industry for six months, but I realized that this is not what I wanted to do for the rest of my life. I wanted to use my skill set, for something which has a more direct impact — not that what I was doing was not impactful — but I wanted to do something where I could work on more pressing causes, like understanding the brain. So that's when I moved into human brain mapping. My work focuses on building generative models of brain function, primarily using dynamic causal modeling. These are used to understand how neural systems interact, adapt and malfunction, particularly across neuropsychiatric and neurodegenerative diseases, and in altered states of consciousness.
SRS: How was the experience coming from a probably more ordered world of engineering, into the messiness of the human brain?
AR: I started in engineering and was drawn to problems that required mathematical rigor. I think I already had some real-world impact, but I wanted to do more and I thought that understanding brain function would be that. It wasn't an accidental move, as it is for some people. I left my industrial job and took a teaching job at a university, in my hometown Karachi, in Pakistan and stayed there for a year. I applied for positions that I thought would help me realize what I wanted to do for the rest of my life. And I got the job to work with Karl Friston at UCL. But as I landed in London, I soon realized that I had entered a mess (as you said). The first thing was the jargon. It's very different how you describe the same thing in engineering compared to medical sciences. The first two years were extremely tough. I had the requisite background in maths and in engineering, but it wasn't an easy transition. I took quite some time to understand what problems are important and how I can contribute meaningfully to further the existing state of knowledge. I was very fortunate to be part of a team that had people like myself before, who are coming from very different areas wanting to make an impact. But once you get a hang of it, then things become much easier: You start to see the connections to what you were doing before, and the things that you can do.
There's a tendency I see at different institutes and organizations of getting someone who has an engineering, computer science, AI, or physics background and putting them together with a biologist or neuroscientist, thinking that they will do something exciting together. My experience is that it often does not work just like that. I think one person needs to know both sides: they do not have to master both ends, but they have to be comfortable. It takes time.
SRS: Let’s go into your work. You do a lot of computational and dynamic causal modeling. Could you just give a quick rundown?
AR: The Bayesian framework is extremely powerful because it treats inference as a fundamental computation, whether it's a neuron interpreting sensory data or scientists interpreting experimental results. So my work has focused on building hierarchical models of brain connectivity where we can infer hidden causes of observed dynamics using priors and observed data. These are the components of Bayesian modeling. What's exciting is that this offers a principal way to go beyond description to explanation. The challenge lies in balancing the model complexity with its interpretability, especially when we have to generalize across individuals or modalities. Bayesian approaches are very powerful because they give you a tool to explain your findings, to provide mechanistic insights. The methods are providing the experimenter with a lot they can do. For example, they are given a possible outcome with a certain probability and it's up to the experimenter to assess the risk that comes with making a decision, including other exogenous factors that they would consider.
SRS: Your work has been relying on Friston’s Free Energy Principle, e.g., via dynamic causal modeling. It is highly influential, but is not without criticism. What is your take on this?
AR: The free energy principle is highly influential, as you say, but also controversial, and it can rub people the wrong way. I see the free energy principle as a generative lens, rather than a testable hypothesis, per se. It can produce ideas, which allow you to explain your hypotheses in a certain way. It provides a unifying theoretical framework for understanding perception, action, and learning as inference problems. Yes, it’s abstract and ambitious, but that’s also its strength. Criticisms often come from misunderstanding its scope or overextending its application. I think as we refine its mathematical formulations and link it to empirical data, for example through tools like dynamic causal modeling, it will continue to be a central guiding framework for neuroAI and theoretical neuroscience.
SRS: Great, I want to pivot a bit to the other really fascinating research domain of yours: psychedelics. What made you go in this direction?
AR: My reason for going into psychedelics is that I wanted to understand what it means for the brain to generate its own reality and what happens when that model breaks down or becomes more fluid? Psychedelics offer the rare opportunity to perturb this generative process in a controlled setting. It allows me to break or change consciousness, and I can then image people's brains while they are having these altered states and subjective experiences. From a bird’s eye perspective, it challenges the assumptions about priors, precision, and self-representations, making psychedelics ideal for testing and refining theories like predictive coding and the free energy principle. With one of my first PhD students – Devon Stoliker, we started that kind of work, resulting in realising PsiConnect, which is currently the world's largest single-site dataset with 60 people who were given psilocybin. It's a multi-modal dataset with fMRI, diffusion imaging, EEG, acquired over six years. We have now published our first paper and the dataset.
SRS: I didn’t think about that research in this way before: Taking the brain in a natural state and using psychedelics as perturbation, to investigate the process of it breaking down. It makes me appreciate that research even more.
AR: Exactly. My personal focus is on discovery science using psychedelics as a tool which perturbs the brain and helps us to understand the brain dynamics — rather than its immediate clinical translation. Of course, there's a lot of potential that psychedelics can be used to treat various sorts of mental illnesses. But my personal interest is using them to perturb consciousness in real time, having observations of people’s brains while they're hallucinating, having ego dissolved or out-of-body experiences. And I think that has helped us to understand more fundamental questions about neurobiology. But the area is changing, I now have many collaborators and colleagues who are interested in clinical applications. This was not my main motivation, but we're going a bit more in that direction. In the past years there have been more people like me, not thinking of psychedelics as something magical, but as a tool to perturb perception and to try to understand how the brain works. We have made our data open access for everyone and invite others to join us in mining the data, and do things that we didn't think about.
SRS: Can you share how you navigate and plan these studies in terms of ethical and legal aspects?
AR: These compounds often have high barriers to access and use them in clinical or scientific settings. We have been a bit lucky in Australia as there's a bit more openness: It is the first country to legalize psilocybin as treatment for depression. But when we started in 2019 it was extremely hard even getting the compound. Our ethics application was the first in Australia to administer these compounds in healthy people. But even now that we have done these trials, starting a new study still requires very careful planning and constant dialog with ethics committees, regulators, and participants. We have worked very hard to build trust and transparency into our protocols, from vigorous safety assessments to informed consent, collaborating with experienced clinicians and adhering to the highest standards of trial design has been absolutely crucial. It requires a lot of effort and time, but in the end, I think it's all worth it.
SRS: What would you say has been one of the most significant obstacles in your career and how did you work around that?
AR: One ongoing challenge has been reconciling interdisciplinary expectations: What counts as a good model in neuroscience, may not satisfy an engineer, and vice versa. This dialogue between different disciplines, I think, has been the main challenge. I could develop a very good model, which may be very robust, but it may not capture the underlying biological effect that we are trying to investigate. For example, we have done some work in spectral dynamic causal modeling, which is a very standard state-space model from an engineering perspective, not very cutting edge. But it works for the problem at hand. You do not always go after the best possible technology. You start with a biological question, and then find out what's the best way to address that question. And it might turn out that it is just a linear regression in the end. That's fine.
SRS: I think it might be great to hear from you, what advice would you offer early career researchers to help navigate their academic lives?
AR: I don't know if I am qualified to give much advice. But I think the main thing is: Work on questions that truly fascinate you. Everyone tells you to do that, but I think that it's really, really important that you love what you are working on. And I think that we have to find the right balance between mentoring and over-mentoring. Using where I was in my career before moving into neuroscience as an example: if I had a mentor at that time, they probably would not have allowed me to take that jump. I was reasonably established in my career, but then went into a totally different direction. Sometimes there's a high risk, but you also get much bigger benefits. I think it's also easy to chase trends, but long-term progress comes from intellectual commitment, you have to commit yourself intellectually to a problem that fascinates you. Science is social, and finding the right people to work with can be just as important as the right idea. Lastly, I think one should be very comfortable with failures. It's not just a learning opportunity, it's where innovation often begins. If you have failed at something, in many cases, it teaches you more than when you have a very good outcome in the first shot. So learn to be comfortable with failure.
SRS:You said you have been a member of OHBM for quite a while — can you tell us about your experience with OHBM?
AR: I have been volunteering for a while now with OHBM, and I love it. I have been a beneficiary of the community and I really feel like I have to give back. It's a cliché, but that's what every scientist should be doing. I was involved in the education program committee for several years and when the pandemic hit and everything had to go online, I was part of the technical committee. I've been part of other initiatives as well, like trying to build consensus across nomenclature for brain networks, the WHATNET group (led by Lucina Uddin). I have had a great time in OHBM, volunteering, making new friends, and meeting old friends and colleagues. I would highly, highly recommend all the ECRs and senior colleagues to get involved. It's a great community.
SRS: Speaking about OHBM could you also give us a little sneak peek into your keynote lecture this year?
AR: I will trace a theme I call inference engines, from Dynamic causal modeling to our recent work on psychedelics. I will show how perturbation based approaches, including psychedelics, can help us understand generative brain function. I will show multiple ways in which we have been extending dynamic causal modeling as a generative modeling framework. And I will also talk about more topical stuff: Foundational models for neuroimaging, which are large scale, pretrained models that can capture shared representations across data sets, from fMRI time series to structural and effective connectivity.
