At CPhI North America earlier this month, Brandon Allgood, CTO and co-founder of Numerate, delivered a keynote on the role artificial intelligence (AI) currently has and will have in the pharmaceutical industry.
In his keynote, Allgood stated that AI is both a buzzword and a potential industry changer. It has a greater capacity for running systems than the human mind and that’s an opportunity, not a threat.
Allgood said in his keynote, “I would never tell you that AI will replace you, but what AI will do is make every one of you superhuman. It's about augmenting your abilities.”
His talk continued on the idea of AI lending a hand in processing – something the industry needs as science continues to accelerate and data becomes a bigger source. The brain can process around 1,000 instructions a second, but an artificial intelligence program can process more in the range of 10 billion a second, according to Allgood.
This idea of faster processing is what can make machine learning and AI seem like the possible next industrial revolution. Much like the machines in factories, AI systems do not need to sleep – a human’s recharge time is greater than that of a computer, according to Allgood.
To get an even deeper picture of AI’s potential, in-PharmaTechnologist (IPT) spoke with Allgood (BA) on the subject.
IPT: What is your definition of artificial intelligence?
BA: It’s such a great term because not even we as experts can define it.
Intelligence is the ability to achieve complex goals, period. It’s broad but it covers all the bases. So, what is artificial intelligence? It is an intelligence designed by either a human or another artificial intelligence that can achieve complex goals. So, an AI can design an AI and a human can design an AI.
One thing I didn’t put in my definition is ‘computer’ because the vast majority of AIs are running on computers, but there are already papers out there of people having developed machine learning models and putting them on wetware. Artificial intelligence can run on a cell-based system, an artificial cell-based system – it’s not just hardware, it’s not just computers.
IPT: What are some different types of AI?
BA: What we’re talking about, when we’re generally talking about AI, is narrow AI. Narrow AIs can achieve one task, often better than humans.
A lot of researchers believe that AI refers to something with humans and believe that humans are the only measure of intelligence, but I don’t agree.
When most people talk about AI today, they’re talking about machine learning and machine learning is the ability to take in a set of examples and identify patterns in those examples to then encode those patterns in an equation, such that you can evaluate new data based on the patterns you found in that training set.
In reality, AI is a set of equations that encodes knowledge – that’s the simplest definition of AI. Machine learning acquires those equations from those examples.
If you think about that, you’ve probably used Excel, and you may have even done a linear fit to a set of data; you have then done machine learning. All that deep learning and other techniques that are coming online today are more complex versions of linear regressions, that’s all it is – it’s just mathematics.
Machine learning and AI aren’t magic boxes, they are just a set of equations that are arrived at through different means that encode knowledge.
IPT: How can the industry take on AI as something that will change technical operations?
BA: What you see today is the places where AI has really exploded, which are places where you have a lot of data, and the data is generally cheap or free to collect.
AI has been very successful at image recognition and has become very successful in human language translation.
If you look at our industry, I’d say that the vast majority of our data sets are not publicly available, they are within the walls of a pharmaceutical company, [contract research organization] CROs or manufacturing companies. So, there’s not tons of publicly available data to aggregate or build large models. The data in our space is expensive to generate generally, synthesizing a compound can be a thousand dollars and testing it can be a few thousand more and on the manufacturing side, it's expensive.
I think what we need to do as an industry is look at where AI has been successful so far, but we will not be able to take those methods and blindly apply them to our industry. What we need are people that are trained that are subject matter experts in various topics needed in our space, as well as have some AI literacy.
The data is not ready, in most industry’s the data is not where it should be.
Right now, we’re working really hard in pharma to get the data in shape but it’s going to be about having those domain experts that also understand AI and also having the data in the right form, such that we can generate those models that can then start propelling the industry.
IPT: In what ways can AI be applied to pharma manufacturing or drug development?
BA: Target identification is one. There are a lot of low-throughput, high -content types of biology going on, such as cell-based assays. There are companies that are using pretty standard convolutional neural networks of images of cells to identify diseases and targets identified to diseases.
Drug discovery, which is what Numerate does, is another. The problem with drug discovery is the lack of data, so what we’ve done is develop techniques that are very good at making predictions with small amounts of data.
If there is an industry, or area within an industry, where there is an expert, the person that understands things at some other level, AI can work. Within discovery there’s the medicinal chemist that can somehow think about molecules in this ten to the 60 space of all small molecules but can still come up with solutions, machine learning can also do that.
There’s also clinical trial design. To identify patients, you can use it in adaptive trials, you can look at biomarkers and; predictive biomarkers. Then there are places of pharmacovigilance. One place where AI is having a huge effect is pharmacovigilance (PV), where you can take the data that’s coming in, as long as its in good form, then this data can be applied to AI can flag the anomalies and categorize and file and run statistics over 95% of the data that comes in, leaving 5% for the humans to look over the anomalies.
I think PV is the place where AI is, bar- none, going to be the only way to go, especially with IoT and all the data we have coming in.
On the manufacturing side, you have quality controls that are happening within your facilities, and individual instruments, and you need to be able to react in real time. Recurrent neural networks and machine learning can be listening to those sensors 24 hours a day – if there is a slight temperature variation or quality control issue,s the AI can take actions in real time, whereas you may not notice it until the next morning. In manufacturing, there’s a huge place for using AI, in terms of automation.
That’s just what I know and that’s a small fraction.
IPT: How will AI be utilized in the future of the pharmaceutical industry?
BA: If you look at the internal return on revenue right now, we as an industry are below the cost of capital in terms of pharma R&D, and we are going to be at zero by next year.
Long gone are the days of ‘inhibit an enzyme and lower blood pressure and make a million dollars.’ What we’re left with today are very complicated diseases, neurodegeneration, oncology, diseases are most definitely poly-pharmacological or involve something newer, like protein therapies and other things.
On top of that, the development of those disease therapies is getting more and more expensive. These diseases have smaller and smaller patient populations, so this is a real problem – as complexity goes up, the cost goes up and the population you’re going to cure goes down.
We need machine learning, we need AI, we need something, because the industrialized process we have developed as an industry is not ready for that kind of non-blockbuster environment.
My vision is that AI must be part of that solution in order to get us to the point where it can become financially viable to go after these diseases and to go after these small patient population diseases.
I think if you’re a pharmaceutical executive and you’re not thinking about AI, you should be thinking about getting a new job.
Brandon Allgood currently holds the position of CTO at Numerate and, previous to this role, Allgood was co-founder and vice chair of The Alliance for Artificial Intelligence in Healthcare. He also holds a place as a member of the Forbes Technology Council.
