AI is reengineering drug discovery by speeding up testing and scanning petabytes of data for connections between diseases

In December, The Conversation hosted a webinar on AI’s revolutionary role in drug discovery and development.

Science and technology editor Eric Smalley^[1] interviewed Jeffrey Skolnick^[2], eminent scholar in computational systems biology at Georgia Institute of Technology, and Benjamin P. Brown^[3], assistant professor of pharmacology at Vanderbilt University.

Skolnick has developed AI-based approaches to predict protein structure and function that may help with drug discovery and finding off-label uses of existing drugs. Brown’s lab works on creating new computer models that make drug discovery faster and more reliable. Below is a condensed and edited version of the interview.

Let’s start with the big picture. How is AI changing biomedical research and drug discovery, and what is the potential we are talking about?

Skolnick: The upside, potentially, is very large. One of the frustrating things about drug discovery is that, in spite of the fact that the people doing it are extraordinarily intelligent and have done an extraordinarily good job, the success rate is very low^[4]. About 1 in 5^[5] drugs will have negative health effects that outweigh its benefits. Of the ones that pass, roughly half don’t work^[6].

In drug development, there are several key issues: Can you predict which target is driving a particular disease? Once this target is identified, how can you guarantee the drug is going to work and isn’t simultaneously going to kill you?

These are outstanding problems in drug discovery in which AI can play an important, though not 100% guaranteed, role. Unlike us, AI can look at basically all available knowledge^[7]. On a good day it makes strong and true connections called “insights^[8],” and on a bad day it does what is called “hallucinating^[9]” and sees things that are weak and probably false.

At the end of the day, many diseases do not have a cure. Most diseases are maintained, such as high cholesterol or autoimmune conditions. A treatment for cancer might buy you five years, and now you’re in Stage 4 and you’ve exhausted all the standard care drugs. AI can play a role^[10] to suggest alternatives where there are none.

Let’s give some basic definitions here. When we use the word drug, we’re talking about a wide range of therapies. Can you explain the range – we’ve got small molecule drugs, biologics, gene therapies, cell therapies.

Brown: We have fairly large molecules in our bodies called proteins. They are like machines that carry out specific functions^[11] and interact with one another. Oftentimes, when we’re trying to treat disease, we’re trying to alter functions of specific proteins^[12]. Many drugs, like aspirin^[13] and Tylenol^[14], are small molecules that can fit into a protein and change its function. Fundamentally, drugs don’t have to just interact with proteins, but this is a major way in which our current repertoire of medications work.

There are also proteins that act like drugs, such as antibodies^[15]. When you receive a vaccine for a virus, your body is basically given instructions on how to develop antibodies^[16]. These antibodies will target some part of that virus. Your body is creating these big molecules, much bigger than aspirin, to go and interact with foreign proteins in a different way. Gene therapy^[17] is a larger step beyond that.

So these modalities – molecule, protein, antibody or gene – are very different types of molecules. They have different scales and rules, so the way you approach designing and discovering them various widely.

Can you briefly explain artificial neural networks, and what the “deep” in deep learning means?

Skolnick: AlphaFold, developed by DeepMind, involved understanding how neural networks worked. They built a network with a lot of inputs, which are stimuli, and outputs with different weights^[18], similar to how your brain actually works. These simple connections, or neurons, have reinforcement learning^[19].

They also created sophisticated neural networks, such as transformers, which do specific things^[21] like a special-purpose tool that can learn, and they added a mechanism called “attention,” which amplifies critical details^[22]. Super neural networks with transformers is what we call deep learning. These now have literally billions, if not trillions, of parameters.

Essentially, these machines can learn higher order correlations between events^[23], meaning the patterns of conditional interactions that depend on the properties of multiple things simultaneously. In these higher order correlations, AI has the potential to see previously unknown things that are embedded in petabytes (a unit of data equivalent to half of the contents of all U.S. academic research libraries^[24] of biological data.

AlphaFold, which predicts three-dimensional, bioactive forms of a protein^[25], has millions of sequences and a couple of hundred thousand structures. It can tell you, based on a particular pattern, what small molecule to design^[26] that sticks to a protein to induce some kind of structural shift.

How is this technology being used in biomedical research to understand molecular dynamics or, essentially, the biological processes involved in health and disease?

Brown: In 2013, there was a Nobel Prize for molecular dynamics simulations^[27], computational tools that help you understand the motions of molecules as they move according to physics. There’s a huge body of scientific research built around those ideas.

AI and deep learning are large right now, but it’s worth mentioning that for the last decade and a half, people have been using much smaller machine learning algorithms^[28] to help design drugs. A lot of the ideas, such as [using machine learning for virtual screening], are not new and have been in practice for a while.

With AlphaFold’s technologies to help people design proteins and predict their structure, we’ve changed how we think about a lot of these problems. We have this new repertoire of approaches^[29] to build ideas around and to start thinking about drug discovery.

From 20 years ago to now, what has today’s AI technology done in terms of scale of change in this process?

Skolnick: A lot of diseases, like cancers, are caused by a collection of malfunctioning proteins^[30]. AI now allows us to start to think conceptually about how these diseases are organized and related to each other.

Diseases tend to co-occur. For example, if you have hyperthyroidism, you’re very likely to develop Alzheimer’s^[31]. Kind of weird, right? We can look at pieces, but AI can look at all the information, integrate the collective behavior and then identify common drivers. This allows you to construct disease interrelationships which offer the possibility of broad spectrum treatments^[32] that could treat whole collections of diseases^[33] rather than narrow-spectrum treatments.

Relatedly, AI also can help us understand disease trajectories^[35]. Diseases that tend to co-occur often present themselves consecutively^[36]. You have disease 1, it gives you disease 2, then gives you disease 3. This suggests that if you go back to the root with disease 1, you may be able to stop a whole bunch of stuff. You can’t analyze millions of trajectories and millions of data without a tool, so you couldn’t do this before.

This holds a lot of promise, but one also must be careful not to overpromise. It will help, it will accelerate, but it is not a substitute yet for real experiments^[37], real clinical validation and trials.

References

1. ^{^} Eric Smalley (theconversation.com)
2. ^{^} Jeffrey Skolnick (biosciences.gatech.edu)
3. ^{^} Benjamin P. Brown (medschool.vanderbilt.edu)
4. ^{^} the success rate is very low (doi.org)
5. ^{^} 1 in 5 (doi.org)
6. ^{^} roughly half don’t work (doi.org)
7. ^{^} all available knowledge (academic.oup.com)
8. ^{^} insights (doi.org)
9. ^{^} hallucinating (theconversation.com)
10. ^{^} AI can play a role (doi.org)
11. ^{^} carry out specific functions (www.ncbi.nlm.nih.gov)
12. ^{^} alter functions of specific proteins (doi.org)
13. ^{^} aspirin (doi.org)
14. ^{^} Tylenol (doi.org)
15. ^{^} antibodies (doi.org)
16. ^{^} instructions on how to develop antibodies (doi.org)
17. ^{^} Gene therapy (doi.org)
18. ^{^} inputs, which are stimuli, and outputs with different weights (doi.org)
19. ^{^} reinforcement learning (theconversation.com)
20. ^{^} Creative Images Lab/Getty Images (www.gettyimages.com)
21. ^{^} transformers, which do specific things (doi.org)
22. ^{^} amplifies critical details (doi.org)
23. ^{^} can learn higher order correlations between events (doi.org)
24. ^{^} half of the contents of all U.S. academic research libraries (www.eecis.udel.edu)
25. ^{^} predicts three-dimensional, bioactive forms of a protein (doi.org)
26. ^{^} small molecule to design (doi.org)
27. ^{^} molecular dynamics simulations (doi.org)
28. ^{^} using much smaller machine learning algorithms (doi.org)
29. ^{^} new repertoire of approaches (doi.org)
30. ^{^} caused by a collection of malfunctioning proteins (doi.org)
31. ^{^} hyperthyroidism, you’re very likely to develop Alzheimer’s (doi.org)
32. ^{^} possibility of broad spectrum treatments (doi.org)
33. ^{^} could treat whole collections of diseases (www.nih.gov)
34. ^{^} Grace Cary/Getty Images (www.gettyimages.com)
35. ^{^} understand disease trajectories (doi.org)
36. ^{^} co-occur often present themselves consecutively (doi.org)
37. ^{^} it is not a substitute yet for real experiments (www.scienceopen.com)