Biosensing, Artificial Intelligence, and Olfaction – The SBT Episode 17 with Rosa Biotech

Our ‘Science Behind the Technology’ series gives us the chance to speak to our members about what they are working on, their plans for the future, and why you should care about their work. 

Rosa Biotech is a biosensing and diagnostic company that span out of the University of Bristol in 2019. Their platform technology combines computationally-designed peptides with machine learning to mimic the logic of the olfactory system – or sense of smell – of mammals. Rosa Biotech are addressing challenges previously thought too complex, time-consuming, or expensive for biosensing. 

This month, our Communications Intern Henry Stennett sat down with Andy Boyce, CEO, to discuss how their technology works, how it could make earlier disease diagnosis possible, and why it’s surprisingly difficult to tell the difference between cats and dogs.


Rosa Biotech was incorporated in March 2019, and soon after the company secured one of the last available laboratories at Unit DX. The company are targeting their biosensor to the industrial biotechnology and diagnostics sectors. 

As Andy explains:

“We do biosensing in its broadest sense. A biosensor is a device that combines a biological component – a cell, protein, or nucleic acid – with a detector. The biological element does the sensing by interacting with an analyte – a substance you want to detect – and producing a signal. This signal is transformed into a more easily measured form – a change in colour, charge, or mechanical pressure – so that an analyte can be measured. 

“Our expertise isn’t in looking for individual compounds, but their sum in complex mixtures. This is particularly relevant to biotech because complex mixtures go into most biological processes.

Arne Scott, Senior Scientist at Rosa Biotech, working in their lab at Unit DX

“We’re exploring the manufacture of biologics, a class of drugs from biological sources with large, intricate structures, such as vaccines and antibody therapies. Biologics are expensive: a year of monoclonal antibody treatment costs $100,000 on average. These life-saving drugs are inaccessible to many individuals, and even in highly developed countries, they’re often not cost-effective.

“Our pharmaceutical partners see 20% swings in the yield of these drugs during manufacturing, which raises their costs considerably. The feedstocks for the cells that produce biologics are a major cause of this variability.” 

So you’re suggesting better mixtures to feed the cells that make biologic drugs?

“Exactly. Our partner uses a chemically-defined medium – a soup of compounds dissolved in water, each at a known concentration. However, they have to add a protein source: usually a lysed yeast extract from a third party. It’s essentially marmite, and as you can imagine, its composition is poorly understood. The differences between yeast extract batches are the biggest cause of variation in the yields of biologics. Our biosensor can help us to understand the differences between yeast batches that perform well and poorly so that our partners can make drugs more economically. 

“The second potential application of our technology is in early diagnostics. For example, in Alzheimer’s disease, diagnosis relies on documenting mental decline, by which point the patient has already sustained brain damage. The earlier Alzheimer’s is caught, the earlier it can be treated, and the better the prognosis.

“For many diseases, subtle metabolic changes occur in the body before symptoms begin. Our technology could measure all of the biological interactions in a complex patient sample, enabling earlier diagnosis.”

How does your sensing technology work?

“For a long time, researchers have been using biosensors optimised to sense one analyte, such as glucose detectors for diabetes, or the antibodies in pregnancy tests. If you want to sense something else, or if there’s a change in the system you’re analysing, you have to completely redesign your sensor. 

“This isn’t how nature does biosensing. Our olfactory system works by differential sensing. Your nose has around 400 nonspecific receptors that are activated to different degrees by the molecules you smell. The brain interprets these complex signals by pattern recognition – it learns to associate patterns of receptor activation with certain smells. Importantly, it’s a trainable system: while I can tell the difference between garlic and red wine, I’m not so good at Bordeaux versus Claret, but with more training, I could be.

Rosa’s barrel structure as seen from the top (left image) and side (right image).

“Instead of the G-protein-coupled receptors found in the nose, we’re using alpha-helical barrel proteins, which are entirely new to nature. They’re made of short stretches of amino acids that coil to form a helix. These helices are designed to stick to their neighbours, forming barrel-like structures.  

“Because our barrels are made from chemically-synthesised peptides, we can vary their sequences. For a 30 amino acid peptide, there are 2030 different permutations – not all form barrels, but it’s still a huge sequence space to explore. Eight residues contribute to the barrel’s central cavity, and by substituting these we can change the size, shape, and hydrophobicity of the pore, which determines how well it binds different compounds. We’ve designed a panel of 96 different barrels and will select the best for each new application.”

“The key transduction event in our biosensor involves a dye which is blue when bound inside the barrel, but colourless in solution. We load our panel of barrels with this dye and then add our analyte, which competes for binding to the barrels. Depending on the chemistry of the pores, the analyte will displace the dye to different degrees in each barrel. We measure this with a fluorescence plate reader, and interpret the results with machine learning.”

So the machine learning then has an image processing task?

“A different type of barrel is added to each well on our plate, giving rise to a unique fingerprint for a given analyte. Our algorithms recognise different fingerprints and relate them to valuable information, such as, ‘Sample A is healthy, and sample B unhealthy’. 

“The classic machine learning problem is training an algorithm to tell the difference between a cat and a dog. The human brain finds this easy, but computers struggle. And it is a difficult task – both are furry animals of similar sizes. You need an algorithm that can cut out all of the noise – eyes, whiskers, noses – and pull out the discriminable information.

“We’re training our system to examine complex mixtures and find the interactions that are meaningful for the questions our clients are trying to answer.”

What’s unique about Rosa Biotech’s system?

“We have a different approach to sensing. Artificial intelligence is changing the way we solve problems: computers can analyse far more data than humans, and make better decisions. It’s not important to know why AI has made a decision, just that it’s a good one. You want to know whether you have a disease, not how many parts per million of lysine are in your blood, for example. Biological systems are so complex that it’s rare that a single component is important – the interplay of many factors results in the outputs we care about. At Rosa, we’re on the cutting edge of the machine learning revolution, and we have a platform that is robust, scalable, and trainable.”

Rosa’s liquid handling robot helps to automate their lab work.

How did Rosa Biotech get its start?

“Our foundation is twenty years of world-class computational biology research in Professor Dek Woolfson’s lab. Dek is a synthetic biologist, and he enjoys playing around with proteins, one of the building blocks of biology. Our alpha-helical barrels were discovered serendipitously – Dek’s group were trying to understand how to make four-helix bundles when they accidentally made one with six. It was fundamental curiosity-driven science, which is where I think the best innovation comes from. After the discovery of the barrels, further research established the design rules so that we could make barrels with different properties and pore sizes.

“One of our pharmaceutical partners asked Dek if his work in protein design could be used to sense the difference between manufacturing inputs. He made the connection between our panels of barrels and olfaction and put his team into action. Within three months they had a proof-of-concept, and two more years of hard work turned an academic project into a robust biosensing pipeline.”  

Next steps

“In the next few months, we’ll be converting an empty white shell full of cardboard boxes into a fully functional, automated lab. We’re going to turn those boxes into a chemistry and data platform for biosensing. 

“In the next year, we’ll carry out pilot studies to understand where our biosensor is best deployed. Ultimately, our goal is to completely revolutionise the way biosensing and diagnostics are done.”

To keep apprised of Rosa Biotech’s research, follow them on LinkedIn or Twitter, or visit their website.

By | 2019-10-31T10:25:25+00:00 October 9th, 2019|The Science Behind the Technology|