BenevolentAI founder Ken Mulvany on using artificial intelligence to find new drugs to tackle disease
PUBLISHED: 19:58 10 March 2018 | UPDATED: 20:15 10 March 2018
Iliffe Media Ltd
Exclusive interview with man who built company valued at £1.3bn as it moves into Babraham Research Campus
Sometimes it’s not what you know that’s important, but what you don’t.
For Ken Mulvany, it’s what we should know, based on what we already know, that’s particularly interesting.
He established BenevolentAI to take the notoriously challenging task of developing new drugs beyond the limitations of human intelligence and into the world of artificial intelligence (AI).
It’s now valued at £1.3billion, making it the one of the world’s five biggest private AI companies – and the biggest in Europe.
BenevolentAI is setting up at Babraham Research Campus after acquiring the UK operation of Proximagen, a successful biotech firm that Ken founded and sold in 2012 to Upsher-Smith Laboratories for $553million.
The move creates an AI-enabled research centre designed to reduce dramatically the time it takes for potential new drugs to enter clinical testing and significantly improve the number that are successful in treating or curing disease.
And it will be a global first for an AI company to work end-to-end on the entire drug development process.
“There are 13,600 diseases, of which only 5,000 have a treatment,” Ken tells the Cambridge Independent. “There are 8,000 diseases with no treatment whatsoever and we’ve got to start tackling that.”
BenevolentAI’s approach is to suck up all of humankind’s bioscience knowledge, held in scientific papers, patents, clinical trial information and data sets, and use AI to uncover connections and concepts that we haven’t yet considered.
“We have the largest knowledge graph related to bioscience on this planet,” says Ken. “To get that we have just about every database – public ones and some proprietary databases as well. We call those structured databases: all the chemistry databases, genomic databases, the proteomic databases – all the omics, everything you can think to put into the system.
“Then we extract knowledge from what is written. You may have a sentence that says: ‘This gene upregulates this process’. The ‘X upregulates Y’ would be extracted. We do that for every sentence of every paper that’s been published – and also every patent.
“There are 10,000 new updates to PubMed alone every day and there’s a new paper published every 30 seconds. We ingest all of that.”
But compiling a vast knowledge bank is not enough.
“We also have to understand what is accurate and what is not,” explains Ken. “Unfortunately, just because something is published in a scientific journal doesn’t mean that it’s true.
“So we need to provide a basis of confidence of each one of those facts that we’ve extracted and relate in that a large knowledge graph that allows us to understand the biological network around each and every fact.
“The output is an enormous knowledge graph that forms what is known, but more importantly we employ reasoning algorithms on top of that to determine what should be known.”
The graph features more than one billion relationships between genes, targets, diseases, proteins and drugs, making it the largest of its kind.
“It’s also the largest proprietary knowledge graph because of all of these inferred things that humanity has not yet identified,” says Ken. “As our hypotheses are formed, they tend to use anywhere between 300,000 or 400,000 of distinct pieces of information to support them.”
The process has generated many promising leads since the formation of BenevolentAI in 2013, with 19 validated programmes in various stages of development in its biotech subsidiary, BenevolentBio. The move to Babraham will enable a rapid scaling up of activity for the company, which has raised £109.5million since its inception.
“We have over 100 projects that we want to take forward but we haven’t had the physical capabilities to do that,” says Ken.
“BenevolentBio has chemists, pharmacologists, toxicologists, clinical affairs and regulatory affairs. It looks like a biotech in many respects, but the ideas are coming from a machine brain rather than a human brain.
“Now we have a chance to scale that beyond our current capabilities into new areas of disease.”
The process of end-to-end drug discovery begins with selecting the right target.
“Even in late stage clinical development – even in phase III – there is 50 per cent attrition because it’s the wrong biology. We need to do something about that,” says Ken.
“The way we think of the process is that it starts with an understanding of the disease. Existing ontologies, or existing understandings, are not sufficient. There are better ways of determining the features or fingerprints of disease.
“Then it’s about having a way to do drug design where you can optimise many parameters at the same time.
“What happens in molecular design when you are looking to impart certain selectivity to hit a receptor is you end up making choices – I might lose a little bit of my ability to go into the brain when I optimise this, or I want a half-life of this length…
“What machine learning is very good at doing is understanding all of those parameters and optimising simultaneously.
“Our system is looking at 220 million variables simultaneously. I can barely hold four. It’s a purpose-built technology for scientific innovation.”
Having created a molecule that can influence a mechanism involved in disease, the next step is understanding what subset of the patient population will respond best to it.
“Again, precision medicine is enabled by machine learning – we use semi-supervised learning to find the molecular signature based on endotypes, based on how patients respond. It’s not a single biomarker at this stage, it is multiple biomarkers, which can provide that fingerprint of who is the best responder,” explains Ken.
“What we’re trying to do is get the right medicine in the right people, to bring the cost of this down and the success rate of this up, as well as the obvious one of curing things that have no treatment.”
BenevolentAI’s most advanced programme is a drug for the treatment of motor neurone disease, also known as ALS. The company entered into significant license agreements with world-leading pharmaceutical companies and last year began its first phase IIb clinical study – these typically define the optimum dosage that will deliver biological activity with minimal side-effects.
“We’ve shaved at this point four years off the process and it costs us 40 per cent of what it costs the industry to do that, because we have to make fewer molecules and have better designed clinical studies,” says Ken. “Our ambition is to shave six years off the development process.”
This goes to the heart of why BenevolentAI is a unicorn – a company valued at more than £1billion.
With drugs often taking decades from discovery to clinical use, the process of bringing one to market is extremely expensive, not to mention risky, meaning it usually takes one of the big pharmaceutical companies to see the process through.
Patents generally expire after 20 years – although drug firms can get an extension in some circumstances – meaning that the window of time to bring a drug to market and earn money from it before its formula becomes available to others is a short one.
Most importantly of all, patients in desperate need of new treatments are having to wait many years to get their hands on them.
BenevolentAI uses business logic but also the availability of patient tissue to validate a hypothesis in determining what to take forward.
“In a disease state, that’s often the best way to determine whether something is going to work – whether there is real disease tissue assays, rather than other assays, to test a hypothesis,” says Ken.
“That hypothesis is completely unbiased. It’s a machine which is generating this: it doesn’t have a prejudice in its understanding of a disease the way a scientist would – such as ‘I particularly like this class of compounds or these targets because I studied these and that’s where my bias leads me’.”
The rise of the machines took another step forward this year at BenevolentAI, when the company’s human bosses decided to trust the AI at its heart still further.
“When a hypothesis is generated by the system it provides a narrative of why it believes it to be the case – the existing and inferred relationships it has used to deduce that this particular mechanism is involved. Our scientists would read that and interrogate it to triage our hypothesis.
“But as of January 1, we switched from a ‘triage-first’ methodology and transitioned to a ‘validation-first’ methodology. The hypothesis has proven to be sufficiently robust in its translation to effect that we don’t have that human triage as part of the process in the beginning anymore.
“We use it more as an introspection after our results come back. We allow our system to get smarter based on that knowledge. It’s a self-learning system – it learns from its own successes and failures,” says Ken.
The company is also looking to apply its AI technology in other valuable science-based markets, such as advanced materials, agriculture, nutraceuticals and animal health.
It has begun by looking at the multi-billion dollar energy storage market, establishing BenevolentAI Energy, which is looking at what materials can provide good energy density.
Right now, it seems there is no end to its capabilities.
Babraham proves a natural choice for BenevolentAI
Having previously founded and then sold Proximagen, Ken Mulvany will have a good idea what he’s getting after acquiring its UK operations for an undisclosed sum.
The serial entrepreneur – also a two-time world sailing champion who competed in the 1992 America’s Cup – told the Cambridge Independent that coming to Babraham, at the heart of “one of the biggest biotech clusters in Europe” made perfect sense for BenevolentAI.
“There are great drug development capabilities there and it’s well kitted-out exactly for our needs. There is a broader opportunity for collaborations because of all the companies there and all the research conducted there. The Institute itself has a large staff. So we’re looking at being an important part of the ecosystem there,” he said.
With the acquisition, BenevolentAI adds a large scientific team with deep expertise in assay development and screening, medicinal and synthetic chemistry, drug metabolism and pharmacokinetics, pharmacology and clinical development.
“We’re forever recruiting quite heavily. We have about 150 in London, 25 in New York, an office in Belgium and a small office in San Diego,” adds Ken.
“We hope to double the size of the entire group this year. That’s everything from mathematicians and computer scientists, to biologists and pharmacologists.”
In a statement announcing the deal, he said: “We have built an artificial intelligence technology that can read and understand the world’s biomedical information, then perform complex reasoning, to tell us things about disease never understood before. This enables us to discover new medicines like no other organisation and drive cures for diseases that were previously untreatable.
“The acquisition has expanded our scale and capability overnight and created something that previously did not exist – an AI company truly integrated across every stage of the drug development process.”