Interview with Florian Jug, FocalPlane scientific advisory board member

Florian Jug

What first inspired you to become a scientist?

For my undergraduate studies, I did computer science with a minor in philosophy of science. During this time, I also worked as a software developer for a company that worked with BMW, where I led a small project team. I learned a lot about how to be a computer scientist and how to create software, but I realised that the main goal of this work was to do the same things but more efficiently, so that everyone makes more money. I found this massively unattractive. I started to become most interested in philosophy of science, my minor field of study. We were asking, what is science and how is it conducted. We were approaching this question as a social phenomenon and as a theoretical phenomenon. This work, as well as a project with Angelika Steger in Munich, encouraged me to move away from software development. Angelika moved to ETH while I was applying for PhD positions, and I was offered a place there. This meant that I ended up doing a rather weird PhD in computational neuroscience in a group that otherwise did discrete mathematics.

Computer science combined with philosophy of science seems like an unusual combination, how did you first become interested in philosophy of science?

Yes, it was not a combination that you could just choose, I had to get permission to study it as a minor. Looking back now, I think that one of the reasons that I became interested in the field was that I really liked Paul Watzlawick. He is a radical constructivist who doesn't believe in the world being in a certain way but believes that everybody creates their own view of the world, their own reality. As scientists create their views of the world, many of us are positivists, believing that we understand more and more of the real reality that is out there. I think that this was the initial hook, while another benefit was having the opportunity to spend some hours each week hanging out with ‘weirdo’ philosophers and thinking about the meaning of life or science or understanding something about the world. It was my hobby!

Can you tell us about your PhD research and your route into bioimage analysis?

In my PhD, I started doing computational neuroscience and I was very much unguided. Angelika is a fantastic scientist, but her focus was on extremal properties of random graphs or randomized algorithms, disciplines that are very mathematical and abstract. However, she always had scientific hobbies. While I was in Munich, I was part of a team attempting to reconstruct 16,000 bags of hand torn documents from the Stasi era. Our system never made it into production, but we demonstrated how a full book we tore to pieces was automatically reconstructed. When Angelika moved to ETH, she decided to have another hobby, computational neuroscience, asking how our brains work. For 2 or 3 years, I had a lot of freedom to roam, think and try things. It was a fantastic time in my life and, at the same time, the least productive in terms of tangible outputs. But I think that these 3 years are probably the most influential for how I think about research today. I believe it gave me the ability to judge for myself whether a direction has any value and worth, or whether it is a waste of time. I would say that I learnt it the hard way, but I would not want to have missed a single day! My remaining career path is not a classical science story either. Since I was doing computational neuroscience, I had to learn some biology. This was hard as I didn't even study much biology in high school. While I was learning more, I met Gaia Pigino, who later became my wife. She was a visiting scientist in Zurich between her PhD and her first postdoc, and was already an exceptional electron microscopist and biologist. We randomly met in the student dorm where we lived, and we stayed in contact after she left Zurich. While I was finishing my PhD in the years that followed, Gaia had already been offered a group leader position at the Max-Planck Institute in Dresden. This meant that I really had to think what I was going to do next. My output wasn't terrible but wasn't great either and I was seriously contemplating looking for jobs in industry. The one thing that convinced me to continue in academia was that Gene Meyers moved to the Max-Planck Institute, and he was looking for postdocs to do image analysis. Gene's lab was also very hands off, and if there's one thing that I had learned in Zurich, it was how to deal with little guidance and lots of freedom! During this time, I think that at least half, if not more, of the papers from Gene's group involved me somehow. I had become reasonably knowledgeable about biological image analysis as it was conducted back then, and I also brought a disproportionate amount of machine learning spirit to the work. I was in the right place at the right time and, in contrast to my PhD, I produced and published lots of results. Things seemed to just fall into place, it was good fun.

After that, I was lucky to secure a group leader position. I had multiple offers, but ended up with the one job I didn't actually apply for! This was at the new Center for Systems Biology in Dresden. I didn't initially apply because the Max-Planck in Dresden had a rather strict non-internal hiring policy. When looking for positions, Gaia and I had decided that we wanted to be able to wake up in the same bed and when the directors heard we might both be leaving because of other offers we had received, they asked me to become a group leader there. This happened during the institute Christmas party and was a nice surprise!

When I was applying for positions, I proposed really nice work, but ended up conducting rather different projects in the end. I was doing a lot of tracking work at the time, but nobody knows me for tracking today. I'm sure it would have been great fun to do what I initially intended, but at the same time that I started the lab, Martin Weigert, a PhD student with Gene who had sat next to me in the lab, had the glorious idea of throwing U-Nets at image restoration tasks. When I saw the first preliminary results, I was immediately thinking: “Wow, this is freaking awesome”. I thought we needed to work fast because it's such an obvious thing to do, and I was sure that there must be other people seeing the potential as well. The final result was our content-aware image restoration (CARE) paper (Weigert et al., 2018). From that point on, it has been difficult to find people to hire who don't want to use deep learning, and I believe that is similar for Martin and Loïc, another key person for the success of CARE. All my attempts to establish any tracking-related projects in the lab have failed because lab members wanted to move on to image analysis questions involving deep learning.

Related to that, how would you describe your mentoring style?

I'm not exactly sure – I would invite you to contact people in the lab to get their opinions! But I would say that we have a really good mood in the lab, which I'm very grateful for. I don't think I follow rules I could easily write down on a piece of paper and I'm not sure how to assemble a team reproducibily – I just happen to be lucky enough that it seems to work out great for us! Twice, actually, once in Dresden, now in Milan! One thing that seems to work well is to allow people to be themselves. And since we are a very international team, no-one feels like the odd one out.

In terms of day-to-day supervision, I try to have contact with everybody at least once a week. We discuss ongoing projects, have an in-depth discussion about technical challenges and anything that is a roadblock to moving forward. This is a lot of fun, and I really love these meetings. In recent months, it has become a bit more difficult because I have a lot of extra things on my plate. These things are part of growing older in academia, I believe, but the last thing I want to sacrifice is the time I spend with everybody in the lab, talking science and being creative together.

As you mentioned, your group is now at the Human Technopole in Milan, can you tell us about the institute and your position there?

Human Technopole is a fantastic, and very young, institute. We are striving to build a research institute in Milan that is spoken about in the same way as people talk about EMBL, the Max-Planck or other world class institutions that conduct life-science research. We are too young to have achieved this quite yet but are moving in big strides towards that goal. Personally, I am part of the Computational Biology Research Centre. There are a number of other centres from Structural Biology to Neurogenomics to Genomics and Health Data Science (with Biophysical Modelling and Simulation and Molecular Cell Biology currently being added). We cover a broad range of different techniques and ways of looking at human health and pathology and our goal (and tagline) is ‘Improving human life and technology by investing in human health and prevention research’. One key part of the project is that slightly more than half of our annual budget is dedicated to creating national facilities. These five national facilities are for genomics, genome engineering and disease modelling, structural biology, light imaging and data handling and analysis. The data handling and analysis facility has two equal sections – one for bioinformatics (i.e. sequence-based analysis), and the other for scientific image analysis. I helped setting those facilities up, with a particular focus on the image analysis part, which is intended to grow to a size of 17 full-time employees, composed of eight bioimage analysts, seven research software engineers and two image labellers and data experts.

What is the current focus of your research group?

In the research group, we are still working on some details of image restoration and denoising, but this is becoming less and less of a focus because of diminishing returns. These days, we are thinking a lot about how to semantically split images. For example, you can acquire multiple fluorescent signals in a single fluorescent channel and then, using adequately trained networks, you can split or decompose this one channel into multiple individual image channels as if they were acquired individually. Our preprint describes how we combined unmixing with denoising, bringing together our expertise in both arenas (Ashesh and Jug, 2024; arXiv). I think that splitting or unmixing will be a theme in the lab for years to come. We have some ideas for combining unmixing with spectral imaging and spatial transcriptomics, but these directions are not ready for prime time quite yet.

When you are deciding what to work on, does it start from thinking about what is possible or do you see a biological problem and think that is something you could help with?

Great question. The reason why we group the facility staff and the methods research team on a single floor at Human Technopole is precisely to combine ideas and insights that draw motivation from what new AI technologies can enable but ground those ideas into what is practically relevant and needed. People sit next to each other, and go for lunch, dinner and drinks together, so that even the most method-heavy employee knows about the daily sorrows of the bioimage analyst. However, while this is a good primer on what problems to look at, in practice, what is important for the methods developers is to appreciate the possibilities of the latest and greatest methods in machine learning and AI. We closely follow the machine learning and computer vision research conducted worldwide, and whenever we see a new and promising approach, we discuss it together and ask ourselves if this new idea has potential applications in an area we care about.

You're also involved in the AI4Life project, can you tell us about that?

AI4Life (https://ai4life.eurobioimaging.eu/) is an infrastructure project funded by the European Commission, which exists under the umbrella of Euro-BioImaging, which aims to bridge the gap between life science and computer science communities. The project involves ten partners who have teamed up for this 3-year project. I believe AI4Life is a great idea and holds lots of potential, but it feels at times a bit like plumbing – in the sense that it requires a lot of infrastructure and standards and ‘boring’ work to enable something potentially really powerful and beautiful. In a recent call, someone who's name I've forgotten said that one needs solid plumbing to build shiny fountains, and this certainly holds true for AI4Life as well! In more general terms, the idea of AI4Life is to bring the computer science and life science communities closer together. The way that we can do this is by taking the latest AI methods that we have just been discussing, and then enriching them with standards and infrastructure that allow us to store trained models in a format that is ‘findable, accessible, interoperable and reusable’ (FAIR). At BioImage.IO, we have a platform that takes care of the findable, so you can find reusable trained models, and bring these models into existing open-source software tools, for example, ilastik, ImJoy, deepImageJ, Fiji and others. These are tools that are made by the community for the community. However, to run deep learning models trained by another researcher, the developers of each tool would need to invest a lot of time to ensure everything is compatible. Then, the individual solutions that the tool developers come up with would likely be incompatible, and everything would be a mess. It is our hope that by having the standards of how models can be stored (i.e. by having a unifying layer), this can be prevented. Within AI4Life, we are developing a standard library that allows people to fetch these models from the model zoo and run them on anyone's data. Users should also be enabled to train a model, or fine tune any model they want on their own data, and re-upload their models back into the model zoo for others to benefit from. At this point, everything becomes FAIR. Today, the project is about two-thirds over, and I would very much hope that by the end of our 3-year funding cycle, most of what I've said has become reality. If the community rejects the idea, we will have no chance to see it through, but if we find people who buy into our vision, we might end up with a FAIR resource that makes use of the European Open Science Cloud (EOSC) and will offer AI services to life scientists who will in turn be more efficient and effective to conduct their own research.

You’ve spoken a lot about ‘FAIR’ and making research open access and open source also come into this, why is this so important?

For me, the most important and most motivating goal is to elevate the rate of scientific discovery. Additionally, conducting science in an open and FAIR way also addresses the reproducibility of results in a fundamental way. Currently, you see papers which describes the training of a model but if you don't have access to the model itself, you can't see what it really does or benefit from its capabilities yourself. If, on the other hand, the model is made publicly available, then everything becomes fully reproducible and reusable. All this also comes with an improved attitude to data hygiene because, ideally, you also upload some of your data to enable others to also reproduce your model training results. And it's not only important because of the reproduction of your results, but also to reproduce complex analysis pipelines. I think it should be the norm to have a mindset where we aim to avoid useless duplication of efforts, especially given that many of us are using taxpayer money.

As a member of the FocalPlane Scientific Advisory Board, how do you think the site and the FocalPlane Network can benefit the microscopy community?

FocalPlane (https://focalplane.biologists.com/) is benefitting the microscopy community in many ways already! There are countless posts on the website that are very popular and informative. FocalPlane also helps organising all kinds of valuable events for people at all career stages and creates a community we can all feel part of. I think that is great and I want to thank you all very much for doing all this!

Could you tell us an interesting fact about yourself that people wouldn't know by looking at your CV?

Hmmm… not sure. Some might know that I love to do pottery and even have a wheel and a kiln at home. In our new house, my wife and I have a whole pottery studio dedicated to this hobby. Unfortunately, we lack the time to do it as much as we would like to.

Another thing that keeps me sane is running. To date, I have done somewhere around 40 marathons or longer events, including multiple 50 milers, some 100-km races, and once, 140 km during a 24 h event. I've never done a 100-mile race, which is an outstanding item on my bucket list – let's see if it will ever happen!

Florian Jug was interviewed by Helen Zenner, Online Editor at Journal of Cell Science and Community Manager of FocalPlane. This piece has been edited and condensed with approval from the interviewee.