How Turbine is bringing the winds of change to drug discovery
Turbine is developing a new drug discovery paradigm.
The company, which has headquarters in Hungary but strong links to Cambridge, has built a computational model of the human cell that can be used to simulate hundreds of millions of experiments.
This can dramatically narrow down the list of targets, dramatically accelerating the discovery process.
Turbine is collaborating with pharma companies to aid their programmes and is now working on its own oncology pipeline.
On his latest visit to Cambridge, CEO and co-founder Szabolcs Nagy told the Cambridge Independent: “The original vision was to create a computational system that can help us extend the functional human health span.
“To do that, we need to help diagnose and treat disease and we need to get the drugs that will really help patients to the market. They need to be very targeted and personalised for a very specific patient population.
“But that also means these drugs need to be developed in a different way to the current drug discovery paradigm. We know that it costs $2.3billion-$2.7billion per drug approval and drug discovery returns are coming down.
“If we want to get more drugs to smaller patient populations, we need to rethink the way that drug discovery is done. This is why we started using the Simulated Cell that was developed during PhD research, originally in 2010.”
CTO and founder Kristóf Szalay invented Turbine’s method of simulating human cells.
He was joined by CMO Dr Dániel Veres and through a friend they met Szabolcs, who had been building secure cloud technology start-up Tresorit in Budapest. Turbine was spun out of the medical university in Budapest in 2015, promising to offer simulation-first drug discovery.
“The Simulated Cell has a number of layers,” explained Szabolcs. “The core idea was to create a mathematical abstraction or representation of how human cells work at a molecular level.
“That means describing with mathematical functions how various proteins interact and how other important molecular components of cellular decision-making interact and drive the system’s behaviour.
“All human cells could behave like this. This core map is built partly based on published literature and the methodology we created for translating a known biological fact into a mathematical equation, and partly based on proprietary data we generated with partners.
“But this is general. If you want to understand how a specific cancer cell would behave in a patient, you would need to turn this into a specific model by adding in genomic data, which is turned into predictions about how active each protein in our model is.
“Are they over-active because they are mutated? Or are they not present at all because they have a damaging mutation?
“The other layer is transcriptomic data, which gives us relative concentration levels of the proteins.
“If you add this together you have a simulated model of a type of cell measured in real life when it was sequenced. It could be a cancer cell line or a patient’s sequenced data.”
Described as the first ‘high definition virtual tumour cell’, the Simulated Cell features 2,000 proteins and their 5,000 interactions. Using it, the team can simulate the behaviour of the cell.
“For example, you can observe it proliferating aggressively if it has those molecular characteristics, or you can observe it going through programmed cell death if you administer a drug that pushes it to that,” said Szabolcs
“The disadvantage is that this is an abstraction of reality. It’s not as accurate as cells in a dish. But the advantage is it has the scale of a computational simulation, which means instead of a couple of hundred or a couple of thousand experiments, you can run tens of millions, or billions. We use various clouds and can scale as needed.
“The other advantage is that you can get a very mechanistic read-out from the cells.”
While typical cell viability readout experiments in oncology labs might tell you whether a cell lives or dies, or whether a particular protein is present, Turbine can reveal what a cell looks like at the molecular level before and after a simulation.
“We can see exactly how active each protein was, and how much protein there was.
“This can create several terabytes for one experiment if you read out all the data,” said Szabolcs.
Machine learning is then applied to derive insights, such as the most effective target to hit.
“Or if you administered a drug, for example, to decrease the activity of a protein, you can find out what types of cells were responsive or resistant, and find a particular molecular characteristic that marked them. You can use that as a biomarker hypothesis that you can bring to a wet lab and confirm,” said Szabolcs.
“The obvious advantage is you don’t need to do large screens or have a high throughout capability. You can be super targeted in the experiments that you do.”
The effectiveness of this approach has been borne out in 10 collaborations with pharmaceutical companies over the past couple of years.
“In 2017 to early 2019, we were focusing with Bayer and a couple of other big pharma companies to support their oncology pipeline and answer distinct questions like: What are the best patient populations to bring my drug to? How can I find those patients in the clinic? What are the biomarkers I should use? Can I combine it with other drugs?” explained Szabolcs.
Many of these in silico predictions for pharma companies are now in the clinical validation stage.
Turbine has used the partnerships to bootstrap the company, growing to 50 employees.
Then last November the firm secured a €3million seed funding round, led by Delin Ventures, which has a Cambridge-based partner in Jonathan Hay, with follow-on angel investment from Esther Dyson, Vishal Gulati and Atlantic Labs. Hauxton-based o2H Ventures was a new investor in the round, which is helping the company to develop its “rational, simulation-basis of drug discovery” and transition into a biotech company in its own right.
“We have a number of targets that we’re following up on, but one is already validated and we are pushing it into the initial stages of medicinal chemistry,” said Szabolcs.
“Usually you would have to run large screens to uncover hypothesis, but we’re running a very small number because we’ve done large screening in silico in the simulations.”
While it will continue to help others on their drug developments, potentially on a value-sharing basis, the company is keenly focused on building out its own pipeline.
“We are going to be very open to partnering around those assets because we are not the experts on bringing them to the clinic and beyond,” Szabolcs acknowledged.
It is here that Turbine’s connection to Cambridge’s life science ecosystem could prove particularly valuable. Already it has engaged with some University of Cambridge scientists.
“We have to learn things that we have never done before – that we have only seen pharma companies do it, so we are looking for partners: great medicinal chemists for example, who can help us run wet lab experiments at scale. We have found a lot of those in Cambridge,” said Szabolcs. “There are some really great oncology labs in Cambridge we want to work with.”
Turbine’s seed funding is a springboard for a Series A funding round, expected this year, which will lead to more recruitment.
Its focus, as a simulation-first biotech, will be around DNA damage repair, which it worked on with pharma companies already.
“There is a thriving ecosystem here around DNA damage repair and we’ve seen that we can develop novelty even in areas where a lot of academic research been done,” noted Szabolcs. “We’ve been able to support some very important assets.
Early signs are encouraging.
“We ran 20 million experiments in silico and we were able to filter down to three target hypotheses, using algorithm post-processing that suggests, for example, if one is a weaker signal or one is a model artefact.
“We had a team of translational scientists look at the data, but they only had to work with a couple of dozen hypotheses.
“In a couple of months, we have come up with four or five novel targets. We are setting up the experimental pipeline with a partner who can hopefully help us be very fast.”
Wet lab experiments are due to begin imminently, and Turbine will track the speed of the approaching compared to existing benchmarks.
“We think we can have a significant time advantage,” said Szabolcs. “In all the validation experiments based on predictions we made, we had about 50 per cent hit rate.
“From a search base of tens or hundreds of millions, we submitted one or two dozen hypotheses. Half were validated and some of those were novel. We included positive controls if they came up in the simulation. So the hit rate is very good. Rethinking drug discovery is fun. Now we have to prove it…”