Pharmaceutical companies expend a lot of effort and millions of dollars developing new drugs -- and not the least of their problems is the amount of time spent researching potential advances that turn out to be a blind alley.
A UK-based company believes that by the creative use of artificial intelligence, it can speed up that process by a factor of 10. BenevolentAI is using artificial intelligence tools to help it wade through millions of pages of drug research, analyse them, and come up with the most promising areas for research.
The company recently appointed Jérôme Pesenti as CEO of its technology division, BenevolentTech. Previously he was vice president of IBM's Watson Platform, and is now leading a team that is using AI to dramatically accelerate the process of developing new pharmaceuticals.
ZDNet recently spoke to him about his new role and the company's keys to success.
ZDNet: You came to BenevolentAI from IBM, where were you before IBM?
Pesenti: I went to the US to study AI and then I started a company outside of Carnegie Mellon. The company was in text mining and search, and I sold it to IBM in 2012. So that company and myself were part of the Watson Unit. I had a big chunk of the R&D [at IBM], a group of 350 researchers to develop the foundational technologies for AI. We were in Astral Place in Manhattan, so I was based there but my team was all over the world.
We were developing what is now the Watson Platform. What I was working on was basically ways to create a lot of technologies that would allow other people to build on top of that: speech recognition, virtual agents, machine translation, and so on.
What does your role at BenevolentTech involve?
Here we are focused on the drug discovery process. If you ask the experts here they will tell you that the drug discovery process in certain companies is broken. It's very expensive. People generate a certain number of candidates, they attach themselves to the candidates very quickly, and they try to push their candidates through the pipeline.
Drugs and drug discovery costs a lot of money. You have to do clinical trials and go through them. The strategy is to get many drugs and try to develop them very quickly.
What we do here is to generate a lot of ideas, look at a lot of patterns, look at a lot of literature, and use computers to look at all these patterns and literature and figure out what they see as the good candidates.
We propose to our scientists a tremendous number of candidates, and then we can help them to eliminate them and to do this very, very drastically.
Whereas tradition companies generate a small number of candidates very quickly that they get attached too, and then stick to them for too long.
They can spend a lot of money and at the end of it find out that they can't handle that drug for some reason.
So [at Benevolent] we eliminate the drugs very quickly and also, we improve the tools with which you do that.
How long has Benevolent been in business?
Two years. I have just joined. The way it was formed was to put together some specialists and some drug discovery specialists. What I am here for is to get more AI into the pipeline, and to make our products more focused and more productised.
Already today, just with the set of tools we have developed, we are seeing some results in making drug discover specialists more efficient.
Is it a case of bringing the two sides together: the drug specialists and the AI specialists?
This is the key. The challenge in AI is to get the people who understand the domain -- the drug discovery specialists here -- and the people who can understand what the technology can do together, because the key is to find the low-hanging fruit.
Take any process and most likely it's not going to solve it end-to-end, right? If you look at all the processes that a human does, there are some tasks that are repetitive that require a huge amount of reading or parsing, and if you can get a computer to do that instead of a human, you can optimise that process.
What is really interesting is you get the drug discovery people sitting right next to, and talking with, the AI specialists and then we can find these things.
The computer can't replace everything. The key is, what can it replace?
So the key is the process?
That is why I was brought in. The aim is to be more systematic. Before I arrived, the company really had two stages. The first was to have the developers and AI specialists figure out what should be done. And, you know, they don't really work too well together.
The second stage, which is where we have had good results, is to really put them together. Get the drug discovery people and the AI people in the same room and try and get them to develop something together.
What I am bringing for the third stage is to be much more systematic about it. The way to be systematic is to actually measure the impact.
What you want to do is organise internal competitions, if you like, which is to have the machine produce the most likely candidates, get the humans to evaluate that, and then change algorithms to make it better, to keep refining it to make your predictions better.
How is that going?
What the company has done in the past year is prove that with good AI tools that can read through the literature and find the connections and can be used by the drug specialists, you can speed up the process.
We have already been able to identify new potential drugs with our technology. It's already happening but what we are trying to do now is make them more robust and more rigorous.
But we can now go through the whole process. What the company is doing here is not just the research but putting the drugs through the clinical trials ourselves -- and that process will be improved using computer-aided technology.
Presumably you are doing this with the drug companies?
Now that is really interesting because it is possible today to have very small companies that can do this. So we have 24 people on the bio side of this office and these 24 people are able, thanks to outsourcing of all of the testing, the clinical testing and so on, to drive the process to the point where they can put the drug to market. We don't know if they are going to put the drug to market because we may license that out.
We are not at that point yet, but we can really grow and then we might go right through to commercialisation. We don't know yet.
In a sense, are you leading the drug companies?
We can license back to them. But we have to get to the stage where we have something out that is proven and then we can license it to them.
We also, sometimes, in-license, because drug companies develop compounds and they might find that it doesn't do what they want to do with it and we find another use. It's called re-purposing.
Do you have a mainframe churning through everything?
No, we do everything in the cloud. We have just announced the new Nvidia [The DGX-1 AI processor], and we do have a few more machines for deep learning. Some things are hard to do in the cloud but we try to do most things there.
What areas are you focused on?
We are focused on different disease areas. We are focused on neuro diseases like Parkinson's and Alzheimer, as well as on rare cancers.
But the proof for us in AI will be having new drug programs. Our goal over the next three years is to have a very wide set of drug programs.
We want to show that in this area, AI actually works.
You say that you read through all the literature but surely you don't mean literally read it?
We scan it all and analyse the content. What are the drugs, the genes, the targets? We do process here the maximum amount of public and semi public information from licensed sources.
Then we get a drug discovery person who has an interest and can query the system, and our system will generate potential ideas.
Read more about artificial intelligence
- Google's Brain Team: 'AIs can be racist and sexist but we can change that'
- Strata HadoopWorld Fall 2016 postmortem: Maybe AI's the future, but can we make the data science work?
- Beware the Midas touch: How to stop AI ruining the world
- Google's next big step for AI: Getting robots to teach each other new skills