Bob Pirok originally planned to become a high school chemistry teacher, inspired by his own chemistry teacher when he was around 14. Research caught his attention while he was pursuing the academic degree required to teach chemistry in the Netherlands, eventually leading him through industry and back to academia. Along the way, he was drawn to the puzzle-solving side of analytical chemistry.
His interest in computational approaches grew during his PhD work in two-dimensional liquid chromatography, where method development could take months and represent a substantial barrier to using more sophisticated technology. That led Pirok and his colleagues to ask: could computational methods make method development easier and analytical technology more accessible?
Today, Pirok, Associate Professor at the University of Amsterdam, the Netherlands, is a leading expert in the use of artificial intelligence in analytical science, which he sees evolving into an “assistant” – one that could help develop methods, analyze data, and make sophisticated analytical technologies easier to use, while leaving scientists to interpret the outcome and recognize when something has gone wrong.
Here, Pirok discusses how a self-learning laboratory might work in practice, why he remains skeptical of the “dark laboratory,” and why the next generation of analytical scientists should learn to work with their data.
Did you always want to be a scientist?
No, I would say I always had a sense that I wanted to be a teacher. A high school chemistry teacher – that was the idea. I think that was the plan from when I was about 14. The main reason was because I had a great chemistry teacher. It’s the usual cliché story. That basically inspires you because you start to understand things.
Later on, as I got into research during my university career – because, in the Netherlands, you need an academic degree to be able to teach chemistry in high school at the higher levels – I realized that research was actually also kind of interesting and really fun.
Then I went into industry for a few years, but I figured out quite quickly that my heart was in academia – actually, mainly on the teaching side. I did miss teaching a lot. That’s the nice thing about working at a university. You can work with companies through your research, you can do research, and you can also teach. It’s a little bit of everything.
Is there a common thread running through your research interests?
It’s the puzzle aspect. It’s not just chemistry, which is of course very interesting in itself, but it is also the measuring – the science of measuring. What do you know, and what don’t you know?
Someone thinks they are making something or they’re trying to make something, and you can basically help them by showing them what is actually happening. Then, on the R&D side, there’s a lot of troubleshooting and puzzling when you’re developing a new technique: linking everything together and asking, “why is it working like this?” Unraveling the data is very intriguing. I like board games in real life.
I would say I’m an application-driven scientist. I also like the fundamental puzzles, and we do fundamental work, but what is always on my mind is: how will anyone be able to use this? Almost all of my PhD students are co-funded by industry.
When did you first become interested in the potential of AI for analytical science?
It started when I did my PhD and was doing method development for comprehensive two-dimensional liquid chromatography (LC×LC). This is one of those techniques where, already in the 1980s, people had theoretically shown that it was much more powerful than one-dimensional liquid chromatography. That’s not to say we should always use it, but it is certainly helpful in certain situations.
Our partners were funding us to develop LC×LC methods, but it wasn’t as though they were actually arriving at a point where they would use these routinely . I learned that a method needs to be robust. In industry and society, a large part of the cost are not just the instrument and consumables, but also the person who needs to develop a method – especially if this takes months.
That’s when we started thinking: what if we use computational methods, not just AI, to lower this barrier and actually make the technology more accessible, including to cost-limited environments, such as public institutes? As analytical chemists, we know that we cannot measure everything all the time, but that more sophisticated technology can help you get closer.
When we talk about AI in analytical science, what do we actually mean?
This is a really good question, especially at a time when the AI button is being pressed everywhere: “oh, we can solve this with AI.”
Many of us now have access to large language models (LLMs) that help us transcribe things, write things, and find things. Naturally, chemists start thinking about how these tools could help in chromatography. The question is which type of AI suits the task. An LLM can help us work with information, but experimental method development also requires measurements and models that learn efficiently from limited data.AI is an umbrella term for computational approaches to tasks associated with human intelligence, such as learning, reasoning, and recognizing patterns.
Early AI was very rule-based. Then followed the emergence of machine learning – probabilistic, data-driven AI. You give it data, it constructs a model, as it gains more data, it learns. That is crucial, because we need data for AI to work. For something like a large language model, there are zillions of pages that you can Google. Chemistry doesn’t work like that.
In chromatography, for example, we still need to inject a sample and measure what happens. The challenge is to use those experimental results to decide which conditions to try next. We were thinking about whether we could use computational methodologies – without using AI at the time – to help lower this barrier. The system would produce a method, test it, consume the data, learn from it, and then propose a new method.
Broadly, you have AI approaches such as large language models or artificial neural networks that consume large quantities of data, and you have models that don’t require that much data. In our context, we use the latter type of sample-efficient approaches, because we do not have a huge numbers of perfectly aligned and labelled chromatograms.
The most suitable methods so far are probabilistic, data-driven forms of AI. However, it’s a little like reversed-phaseand hydrophilic interaction liquid chromatography. They’re both modes of liquid chromatography, but you separate different analytes with different selectivities. It’s like that with machine-learning methods as well. Some are very useful for one application, whereas other techniques are more appropriate for different tasks.
You’ve discussed autonomous, self-learning analytical laboratories as a way to structure and expand scientific knowledge. What might such a laboratory look like in practice?
It’s actually not that far away. I think many of the essential building blocks are there.
There are several groups already working on self-driving analytical systems. We have them in our lab. Basically, our systems can be controlled by an AI system.
Say, for example, a pharmaceutical company is using an AI-based drug discovery tool to continuously produce candidate molecules – as many companies now are. You’ll need something to separate all the resulting products. I know of many companies that suddenly face such an overload of samples. I can imagine that you feed these samples to a system and the system performs the method-development exercises for you – and I do mean for you, not instead of you.
Analytical scientists will still be needed to interpret and validate the outcomes, but they become more effective. In a forensic context or in water analysis, for example, we could imagine advanced technology – maybe a two-dimensional liquid chromatography system or a liquid chromatography-tandem mass spectrometry system – where interpreting the data or developing the method is no longer a huge bottleneck, because you have an AI assistant helping you.
Why can’t AI interpret data in the same way a human can? Could we eventually see the analytical equivalent – a “dark laboratory”?
What is a good chromatogram? What is a good analysis? Imagine you have a chromatogram with 20 peaks, all equally spaced and perfectly separated. What if time is of the essence? What if I take less time and five peaks coelute? Is that worse than having perfect separation if it takes 10 more minutes? If two peaks coelute and then I change the method to solve it but then two other peaks coelute, is that better? Depending on what you actually want, these solutions can be equally good or bad.
If you cannot tell the AI – if you cannot give it a number like a compass that says, “Go in that direction” – it will go in all sorts of directions. It will become very inefficient and less likely to find a fully satisfactory result.
We have spent considerable time pursuing results that later turned out to be affected by instrumental problems rather than the programmed method conditions. These experiences underline why independent checks and human oversight remain essential. To underline this last point, imagine peaks are suddenly being split. That doesn’t mean I now have double the number of separated peaks. More likely, something is happening with my column or system. This requires expertise. Sure, I can imagine AI getting better and better at taking expertise into account – also because of our input. However, in the end, I still think it remains largely a task for the analyst.
I do think that AI can alleviate some burdens. It could assist in method development or data analysis, making it easier for analysts to use sophisticated technologies. It’s more an extension of the analysts’ capabilities than a threat to replace them.
Until now, I unfortunately have plenty of anecdotes and Friday-evening-drinks stories about how one of our AI systems set us back, because something unexpected happened.
What would need to happen to make more autonomous analytical laboratories a reality?
This is maybe the part where I get less excited because it is very difficult.
We have such a lab in operation here, but getting it to work elsewhere takes many more steps. All processes need to be robust and validated. A company would need to take on the task to commercialize the concept, so that users obtain both accountability and support. I can’t provide a help desk for everyone who uses it. To make that happen, we need industry commitment.
We have formed a consortium of partners who are thinking in the same direction, and we are trying to attract partners for commercialization.
For a younger researcher deciding which area of analytical science to pursue, what advice would you offer? Where do you think the field’s cutting edge will be in five to ten years?
It’s difficult to give universal advice, but there is one common thread to any tailored advice I would give: whatever you do, do make sure that you can process and use the data you generate.
For a few years now, all of our students have been required to learn programming – a programming language and how to work with data – rather than relying on the instrument software. Not because the instrument software is bad, but because it is so important to understand what you can distil from the data and which assumptions are being made when the instrument does it for you.
You need to make that connection toward computer-assisted work, such as machine learning.
To make it even simpler: I used to be a 100 percent lab scientist, and now I’m maybe only 30 percent that. I realized that, at least as I see the field, computationally assisted methodologiesare basically the next big thing for mass spectrometry, chromatography, and other areas.
I don’t think a liquid chromatograph, mass spectrometer, or another analytical system will look hugely different in 20 years. They won’t be innovated to the same degree as they were between the 1970s and today. But what I think will be completely different are the data-analysis and control aspects.
It wouldn’t surprise me if there were some kind of central AI hub somewhere that would be able to assist the analysts in any way. It then becomes immensely important to understand what it going on, what the limitations are and what can be improved. are.
For that, you need to educate yourself in the computational sciences within analytical chemistry.
Newsletters
Receive the latest analytical science news, personalities, education, and career development – weekly to your inbox.
