AI in cancer research

Verena, you recently became Head of MedTech, Industrial AI. What lies behind that?
Our Industrial AI business unit is set up across sectors; in principle every manufacturing company is interesting to us, and that will stay that way. But we now increasingly have projects in the MedTech field, above all in the series production of medical devices, where high unit numbers meet zero tolerance for defects. And medicine is a subject close to my heart: I am a computer scientist, but I worked and did my doctorate in cancer research. So it made sense to build up MedTech as our first industry focus. There you get more specific topics on top of that – regulatory requirements, for example, play a far bigger role. If you build software for production in that environment, you are immediately dealing with GAMP 5 (Good Automated Manufacturing Practice) and questions of validation. 

The term MedTech sounds like high tech. What actually falls under it?
Essentially everything that counts as a medical product. That can be hardware products a company manufactures – an X-ray machine is of course a medical device. But the software that runs on it and supports diagnostics falls into the field as well. And in fact almost everything you find in a hospital. Catheters, for example – even the hospital bed is a medical device. The field is very broad. Medicines do not fall under it; legally they are a world of their own. 

So from the sticking plaster to the X-ray machine. And how does that fit into a business unit called Industrial AI?
By industrial we mean manufacturing above all, so companies that produce things. That includes a paper manufacturer just as much as a food producer, or pharma. Making medicines is production too. Anywhere you physically produce something. The problems in production are similar: unplanned downtime, scrap, quality fluctuations. Much of it ultimately comes down to overall equipment effectiveness, OEE. In a regulated environment there is the added point that efficiency must never come at the expense of quality and compliance. 

Making production more efficient is really nothing new. Why is it becoming pressing right now?
Exactly, industry has been working on that essentially since assembly lines have existed, and the need was always there. What is new is that agentic AI makes it possible to start pilots relatively quickly. You explore the data together with AI agents and generate hypotheses: what are the possible drivers? 

It becomes elegant through the combination: AI agents together with classical machine learning. The holy grail in production is predictive maintenance or predictive quality. That is, being able to predict when you actually need to service your machine before it fails, or when your process starts producing scrap. And that only works if you have the necessary sensors sending you the data, condition sensors together with machine states, for example. And that is classical machine learning; for millions of sensor readings, small specialised models are more accurate and cheaper than any language model. 

Two things stood in the way of that for a long time. On the one hand the data was simply missing; a lot has happened there with the smart factory, machines are connected and supply sensor data along with everything else. On the other hand it always took someone to analyse that data at length and in detail. That has become much faster with agentic systems. So today you can evaluate more quickly whether something like maintenance prediction is feasible at all given the data available. 

The pattern – too much data, too few people who can work through it – is one you know from your time in cancer research. What exactly did you do there?
Definitely! And I actually still work one day a week in cancer research, at the German Cancer Research Centre (DKFZ) in Heidelberg. I did my doctorate there; we analysed high-dimensional tumour data. 

You can digitise a tumour. If you are diagnosed with cancer, a biopsy is usually taken, or the tumour is removed after an operation. It is then cut into thin sections. One section you sequence – that gives you the genetic information for each position. A neighbouring one you put into a mass spectrometry imaging device, for example, which measures for each pixel which proteins are present at exactly that spot. 

Why are the proteins of all things so revealing?
For that you need to know the general principle. As a computer scientist, I hope the molecular biologists will forgive me the short version: DNA becomes RNA, and from these RNA transcripts a protein is ultimately produced, which then takes on a particular function in the cell. Every cell in your body has the same DNA, but a nerve cell in your eye is something completely different from a muscle cell. That works through this step: which proteins are produced, in which place, in which cell. That makes proteins the most immediate thing you can measure, and at the same time the point where many therapies take effect. 

And that is why people try to find out which proteins occur more frequently in cancer cells than in normal cells. Cancer arises because a cell keeps dividing uncontrollably and the immune system does not recognise it as defective and sort it out, as it normally would. If you know for every spot in the tumour section which proteins occur there, you can compare that with the protein profile healthy tissue would have at that spot. 

We research head and neck carcinomas in particular. As long as you can operate, that is the safe option, but in that region especially an operation has considerable consequences; speech and swallowing can be impaired. So the question is: how well does radiotherapy work in combination with chemotherapy? And that again depends on which proteins you find in that tumour. 

And that is where statistics no longer get you any further?
You simply get an extreme amount of data. Per pixel, 20,000 pieces of genetic information, and per pixel 50,000 protein fragments. That is a single pixel. From a single tumour. Classical statistics such as hypothesis tests depend on having as many samples as possible, ideally hundreds of tumours to compare. Here it is the other way round: few cases, but tens of thousands of measurements per case. What machine learning can do instead – dimensionality reduction with neural networks, for example – is filter the relevant drivers out of this huge pool. That would not be possible with simple statistics. 

There has been a real breakthrough in AI models in cancer research in recent years. What was it?
Much as in manufacturing, classical machine learning has been applied in cancer research for a very long time. And there, too, there was a revolution, very similar to what we are experiencing right now with language models. The problem with supervised learning is that someone has to sit down and produce annotations. If you think of a tumour section, that part is mostly done by pathologists; they select individual regions and say, "There I see cancer cells and there is normal tissue." That work is extremely laborious. 

The idea is – and this is exactly the same methodology that is also applied to LLMs – that you have a huge amount of data and let the model generate labels itself, on the basis of what it sees. In cancer research that could be a huge amount of image data, CT scans or MRI images for example. That was precisely the breakthrough with ChatGPT. Nobody sat down and said: this is a verb and this is an object and this is what a sentence has to look like. Instead the technology behind it, neural networks, was set up so that the model works out for itself, from the sheer volume of sentences, how a German sentence is usually constructed. 

And the breakthrough in cancer research was analogous: you first train a model purely on huge amounts of image data, without anyone having annotated anything – and in doing so it learns for itself what tissue looks like and what deviates from it. Foundation models like these can answer entirely new questions: do you see something in this image that predicts survival in cancer? And through that, new markers really were found, patterns nobody had noticed before. That is what made a great many new AI models possible in the first place over the past five years. 

And what does that mean in concrete terms for treatment?
Until now it was the case that all patients with the same type of cancer were essentially treated the same way: if the tumour is locally confined and easy to operate on, you operate. If it is locally advanced or hard to operate on – put very simply – there is usually a combination of radiotherapy and chemotherapy. The radiation is delivered over several weeks in individual fractions. 

But we are seeing things move more and more towards personalisation. Because it is now possible, in principle, to sequence anyone’s tumour. You analyse markers and then know: chemotherapy adds no value for you, so we leave it out. Some types of cancer are also triggered by viruses; HPV is one of the drivers for a large group of head and neck carcinomas. And we know: if the tumour is HPV-positive, the chances of recovery are considerably better. Studies are therefore currently investigating whether a lower dose of radiation is then sufficient – so that those affected have correspondingly fewer side effects. So: behind the big buzzword cancer, we already distinguish an enormous number of dimensions in treatment today: molecular markers such as HPV, sex-specific differences, the age of the patient. The single disease called cancer no longer really exists. 

If I go to the doctor with cancer today and it is analysed, does that take a day, a week, a month?
If you mean a full molecular analysis with a personalised treatment recommendation: we are not there yet, that is still a way off. But there are already first steps in that direction. 

At the German Cancer Research Centre, together with the National Centre for Tumour Diseases (NCT), there is a programme called MASTER. It mainly enrols young patients for whom the standard therapies have been exhausted, and patients with rarer types of tumour. Various molecular markers are determined from their tumours, and then a molecular tumour board advises on which personalised therapy might still be an option. 

The sequencing run itself is a matter of hours to days. Then sample preparation and the bioinformatic analysis come on top. In practice, with the MASTER programme it may take several weeks before a case is discussed in the molecular tumour board. 

Sounds as though it only needs to scale. Where is the catch?
The catch is that with increasing personalisation you run into a measurability problem. With new therapies you compare two treatment groups with each other, and statistically that only works so well because everyone in the group was treated the same way. If you now treat every person differently, you can no longer form those groups and you need different study designs. There is no longer standard therapy versus new therapy, because every individual person tends to be treated differently. 

This gold standard by which new therapies are validated in medicine – the randomised controlled trial, in which you compare two treatment groups – therefore reaches its limit: the methodology with which medicine proves its own progress no longer works this way for personalisation. 

A look into the crystal ball: will we be able to treat cancer considerably better with the help of AI in future?
I think the biggest leap is happening one stage earlier: what we see from the pharmaceutical field is that new active substances can be developed much faster. It is about finding molecules that can dock onto the right spot on the cancer cell, and whether a molecule can do that was above all laboratory work. AI is now being used there in a supporting role, and that means you can find good candidates for the next drug much earlier. 

The question is rather: will care keep up that quickly? You also need the necessary structures for it. And it comes with additional costs that would have to be affordable across the board. 

And that is why prevention would actually be the more effective lever?
At the Cancer Research Centre that has become a central topic, yes, because it would be far more effective to prevent cancer. Probably both strands are needed. But really it would be much cooler if we did not have to treat in the first place. 

The sad thing is just: prevention is incredibly unsexy, because it is really always about going without. You are not supposed to smoke, not to drink alcohol, you are supposed to do sport five times a week instead of sitting in front of the television eating crisps. At least that is how it always sounds. 

Can’t AI help there?
Depends where. On the basic question of what actually prevents cancer, the problem lies deeper: how you get hold of reliable data in the first place. With smoking, the connection was essentially discovered in hindsight, because there were strikingly many smokers among lung cancer patients. The effect is so large that you could not miss it, even without a clean study. But if you want to know, for example, whether a vegetarian diet or a diet with meat is the better one, then it is very hard to analyse which data you need for that and over what period. 

Clinical trials normally work by having clearly defined endpoints. People in both groups have a tumour, and you look at how the disease develops over time. The most important endpoint is usually survival: how long do patients live after treatment? Then you get two curves and you see: after five years, 80 per cent are still alive with the new therapy and 60 per cent of all patients with the standard therapy. 

With prevention, though, fortunately none of us has a tumour yet, and all you can do is try to observe the effect of prevention over a very long period. You define a group of people A who live vegetarian and a group B who do not – and now you actually have to follow them for 20 years to see who gets cancer. And even then you do not know whether it was down to the diet: people who live vegetarian usually differ in other respects too; they smoke less often, they move more. Perhaps some vegetarians also become vegans. There are too many influencing factors and they are hard to isolate. That is why prevention is a hugely important topic, but hard to research. We have long known the important levers, though: smoking, alcohol, obesity. We do not have a knowledge problem, we have an implementation problem. Prevention is just unspectacular. Not a breakthrough, but everyday life.