Pod Insights Big Ideas Lab Drug Discovery About this episode For centuries, drug discovery was a slow, trial-and-error process—sometimes taking decades to develop life-saving treatments. But what if we could speed up that timeline? At LLNL, scientists are using supercomputing, machine learning, and AI to revolutionize how new medicines are found, tested, and developed.
Transcript 0:00 Speaker 00 Lawrence Livermore National Laboratory is hiring. If you're passionate about tackling real-world challenges in science, engineering, business, or skilled trades, there's a place for you at the lab. Right now, positions are open for a Senior Guidance, Navigation, and Controls Engineer, and a Senior Business Operations Specialist. These are just a few of the more than 100 exciting roles available. At Lawrence Livermore, you'll work on projects that matter, from national security to cutting-edge scientific advancements. Join a team that values innovation, collaboration, and professional growth.
0:40 Speaker 00 Explore opportunities at llnl.gov/careers, where your next career move could make history.
0:54 Speaker 00 When was the last time you had a headache? An ear-splitting, skull-pounding headache? What did you do? Probably massaged your temples, maybe dimmed the lights, and then you did what any normal, sane person would do: went outside, found the nearest willow tree, and chewed its bark. Right? If you'd asked a physician 4,000 years ago, that's the exact prescription they would have given you. All across ancient Greece, Sumeria, Egypt, and China, healers recommended willow leaves and bark to treat pain, fevers, and inflammation.
1:36 Speaker 00 But if that doesn't sound tasty, you could brew it into a tea. Or more like, hot bark water. This worked because willow bark contains a molecule called salicyte, which causes a reaction in your body that eases pain. In the mid-1700s, a cleric of the Church of England, Edward Stone, rediscovered the willow's properties. But instead of simply gnawing on leaves, he took things a step further, and dried the tree's bark for months, creating a powder. With it, he treated himself and several neighbors. To varying success. Isolating and extracting the active ingredient, salicyte, was the next great leap forward in 1826.
2:21 Speaker 00 After that, the development of the drug as we know it today began in earnest. But it wasn't until 1915, nearly 90 years later, that Bayer released over-the-counter aspirin, a small white pill that could banish a headache in 15 minutes. It almost seemed like magic, but it's science and centuries of experimentation, research, and testing. Nowadays, thankfully, creating new drugs doesn't take thousands of years, but it's still a long process, generally 15 years or more. The folks at Lawrence Livermore National Lab are hoping to change that.
3:06 Speaker 00 Using state-of-the-art supercomputing, machine learning, and AI technology, they hope to shorten the gap between research and discovery. It's a tall order, but they're already making huge strides in creating new medications, discovering cures, and saving lives. Welcome to the Big Ideas Lab, your weekly exploration inside Lawrence Livermore National Laboratory. Hear untold stories, meet boundary-pushing pioneers, and get unparalleled access inside the gates. From national security challenges to computing revolutions, discover the innovations that
3:47 Speaker 00 are shaping tomorrow today. Looking for a career that challenges and inspires? Lawrence Livermore National Laboratory is hiring for a power grid dynamics engineer and a transmission grid engineer, along with many other roles in science, technology, engineering, and beyond. At the lab, every role contributes to groundbreaking projects in national security, advanced computing, and scientific research, all within a collaborative, mission-driven environment. Discover open positions at llnl.gov/careers, where big ideas come to life.
4:37 Speaker 01 Drug discovery is the process of how we come up with a new medicine for a given purpose. So if we have a particular pathogen like SARS-CoV-2 virus, or a medical condition of some kind, what we want to do then is to find a molecule that could actually go into a person and restore their system to its normal operation.
5:01 Speaker 00 That's Jim Brossey.
5:03 Speaker 01 My job title is Deputy Associate Director for Computing. I'm the overall lead on our bioresilience initiative at Lawrence Livermore. I coordinate across different bio projects at the lab, particularly in the areas where we're bringing together high-performance computing and biology. That's the central theme of what we're doing, the integration of computing and biology to really enable better predictive models, better rapid countermeasure or drug development.
5:34 Speaker 00 To make it into your medicine cabinet, a drug has to travel a long, complex road. There are five stages of development. The first phase is discovery and development. Here, chemists and biologists decide what malady or disease they want to target, and then design and formulate molecules to treat it. Phase two and three are preclinical and clinical research. This is where countless tests are run and safety for human consumption is determined. The next is FDA review. At this stage, hopefully, the drug is effective and passes final reviews to be cleared for market. Once each of these steps has been
6:14 Speaker 00 completed, the medicine finally goes to patients, and the fifth stage begins: post-market safety monitoring. If it sounds like a complicated process, that's because it is. From start to finish, creating a new drug takes between 10 to 15 years. One of the biggest bottlenecks is at the first stage, discovery and development. Forming a new drug can take up to five years. Lawrence Livermore wants to change that. But how?
6:47 Speaker 02 As a Department of Energy laboratory, we are interested in advancing and creating new technologies. The application of drug discovery and development is one of many areas that we can apply high-performance computing.
7:01 Speaker 00 That was Felice Lightstone, a leader in the Biochemical and Biophysical Systems Group at the lab.
7:07 Speaker 02 When you say drug discovery or drug design, you are designing a small molecule, which is a chemical entity, designed to target a protein in exactly the same way each and every time it works. These would be drugs like you'd take in a pill form. You've got a headache, you take acetaminophen or Tylenol. It's a small molecule. It goes into your body and it finds its target, meaning where it's going to change some aspect of how your body functions. Usually, it's a protein. You hope to change the function of that protein to show improvement of whatever disease you might have. It could be as simple as a headache. It could be complicated like diabetes
7:51 Speaker 02 or metabolic syndrome or cancer.
7:53 Speaker 00 Felice and her team are doing some really exciting stuff with high-performance computing in biology. But to appreciate their work, first, we need to understand the traditional method of drug development.
8:05 Speaker 01 Historically, drug discovery has really sort of used a random search with experiments. You build up big libraries of chemical compounds of the types that you think might be applicable to a certain drug target.
8:19 Speaker 00 Think of the chemical compounds like puzzle pieces and the targets as empty spaces in the puzzle.
8:27 Speaker 02 Once you have a target, you need to have what they call a hit molecule. That's any molecule that will bind to your protein target.
8:33 Speaker 01 Then you test them against that particular molecular target and you see what works or what looks like it's close to working. Then you modify your set of chemicals. Then you do it again. Through this process of making and testing molecules, you eventually narrow in on molecules that could be effective.
8:55 Speaker 01 In binding and neutralizing this target.
8:58 Speaker 00 They have to see if each puzzle piece or molecule will fit.
9:02 Speaker 01 The problem with that approach is that it can take months for every one of those cycles to go through. You have to make molecules successfully. That's complicated. You need chemists to figure out the recipes. Then you have to actually do the testing and interpret the data and so on. You have to do quality control. It's a very time consuming process. That's how drug discovery has traditionally been done, and that's why it generally takes many years.
9:31 Speaker 00 At its core, the process now is still the same. Discover, make and test molecules.
9:39 Speaker 03 Our institution is more in the technology development and probably more on the discovery side of it. So our goal is to de-risk the development of these molecules.
9:50 Speaker 00 For the actual creation of molecules, Livermore has partnered with the National Cancer Institute's Frederick National Laboratory for Cancer Research and Bridge Bio Oncology Therapeutics, a company looking for novel treatments for cancer. But how do they run 10,000 simulations a week and narrow it down to only 20 molecules for Bridge Bio to synthesize? This is where Livermore's supercomputers and machine learning come in.
10:19 Speaker 01 There's really two trends in this that are important. One is our understanding of biology is getting better. Our approaches to simulating molecular interactions are getting better. Our computational power has grown, so we can actually do physics-based simulations. The other one is this data-driven machine learning AI, bringing those two together.
10:42 Speaker 02 You've heard about ChatGPT, that's a language model, but we're exploring different kinds of algorithms for machine learning. We get to use the biggest computers in the world. We're using high-performance computing to try to solve biology problems. We start from the basic science where we're looking at new capabilities, actually trying to use the computers to make methods go faster. So how do we design drugs in a faster way?
11:09 Speaker 01 The AI can predict new molecules in new regimes that we haven't seen, but they generally don't have enough data to drive them. We really get a lot of increased power in the predictive models by bringing those two streams of work together.
11:26 Speaker 00 In our analogy, the AI would be searching for puzzle pieces while computers run tests to see if they fit the missing space.
11:33 Speaker 02 The key here is that we moved to a physics-based simulation model, and we can run about 10,000 simulations a week. We have a three-dimensional model of every atom in that protein that we're targeting. We use computers to look at small molecules in that virtual space, and then we design it so that that small molecule will bind tightly with the protein target.
11:58 Speaker 00 So instead of sifting through a box of millions of puzzle pieces by hand, they're using models to narrow down what pieces needed. A corner? Edge? Middle?
12:10 Speaker 02 But this way, they can hone and target their experiments, and there's hopefully an acceleration in the process and/or a decrease in the cost.
12:19 Speaker 00 They saw huge success in a recent study.
12:22 Speaker 02 We only had to synthesize about 500 or 600 compounds, and it sounds still like a big number, but a traditional drug program runs anywhere from 2,000 to 5,000 compounds.
12:34 Speaker 00 That's a 75% decrease, which means drugs that might have taken years to develop can now be finished in a matter of months.
12:43 Speaker 03 My name is Jonathan Allen. I'm an informatics scientist at Lawrence Livermore National Laboratory.
12:51 Speaker 00 That means he works with data and computational problems to aid in biology research. His main focus now is small molecule drug discovery, but he's used high-powered computing for other work as well.
13:04 Speaker 03 Thinking back to a previous project I worked on, where we published a study that was a reanalysis of human microbiome samples. An instance where we were able to take advantage of unique computing facilities. We had access to this large cluster that allowed us to search every publicly available human microbiome sample against a very large database of microbial genomes to try to assess what actually was in these human microbiome samples.
13:39 Speaker 00 Every living person has microbes living in their body. These can be bacteria, viruses, and cells that contain DNA. Analyzing them gets scientists closer to developing new treatments for diseases, both genetic and otherwise.
13:56 Speaker 03 What we actually ended up finding was there was a lot more human DNA in these samples than was previously recognized because we were able to pick up on more of the natural human genetic variation in the samples that hadn't been previously detected because it hadn't been searched for. It was on the order of like weeks that we were able to do thousands of samples, which would otherwise take maybe months or a year to do.
14:21 Speaker 00 This acceleration is obviously an exciting prospect in drug development. But protecting public health doesn't stop at headaches or genetic diseases. Some of the greatest dangers to humanity lay in other threats, like bioterrorism or viral outbreaks. Will we be ready for those?
14:45 Speaker 00 Lawrence Livermore National Laboratory is hiring. If you're passionate about tackling real-world challenges in science, engineering, business, or skilled trades, there's a place for you at the lab. Right now, positions are open for a senior guidance, navigation, and controls engineer, and a senior business operations specialist. These are just a few of the more than 100 exciting roles available. At Lawrence Livermore, you'll work on projects that matter, from national security to cutting-edge scientific advancements. Join a team that values innovation, collaboration, and professional growth.
15:26 Speaker 00 Explore opportunities at llnl.gov/careers, where your next career move could make history.
15:45 Speaker 01 Livermore's had a biology program since way back in the 1960s. Understanding genomics, bringing new technologies into biology, building detectors for biological defense against bio threats.
15:57 Speaker 03 A big one was this program called BioWatch. Lawrence Livermore was responsible for contributing to a lot of the early technology.
16:04 Speaker 00 BioWatch was introduced in 2003. The system, managed by the Department of Homeland Security, monitors the air to detect possible bioterrorist attacks.
16:15 Speaker 03 The lab has this two-pronged focus in the space, which is around countermeasures, which is developing therapeutics for something that would be a biological threat. The flip side of that is also understanding, what do we need to be developing countermeasures for? How can we be prepositioning ourselves so that we can rapidly respond to something that's an outbreak?
16:41 Speaker 04 From the Oval Office tonight, the president announced his boldest steps yet to combat the nationwide spread of the novel coronavirus.
16:49 Speaker 05 To keep new cases from entering our shores, we will be suspending all travel from Europe to the United States for the next 30 days.
16:58 Speaker 01 The COVID-19 pandemic started. We quickly pivoted to focus on antibodies for viral infections, and that was extremely successful. We demonstrated that we could actually develop or redesign an antibody for a viral variant in just a few weeks, using that, rather than taking a couple of years to do this.
17:19 Speaker 03 RNA viruses are just a naturally concerning biological threat in general, and no better example of that than SARS-CoV-2. There's been a lot of activity at the DOE labs, Lawrence Livermore, as well as the others, in terms of bringing to bear all of these computational tools to develop various responses and countermeasures. We were able to engage in developing some small molecule drug discovery on some of the protein targets for SARS-CoV-2 and develop some potential candidates.
17:55 Speaker 02 We need to be ready for the next infectious disease that comes around. And I think that this is a national mission now that Lawrence Livermore can fulfill.
18:05 Speaker 00 Thankfully, the urgency of the pandemic has subsided. But the lab is using that same technology in other ways to protect Americans.
18:14 Speaker 01 What we're doing now is designing antibodies to support warfighter protection. So if the DOD sends soldiers into a particular area of the world, they may be exposed to a virus. They want to have antibody therapies that they can give to those soldiers going in to protect them. They want to be able to develop these things rapidly. That's going very well.
18:36 Speaker 00 Lawrence Livermore is using computational power to speed up biology research. But almost everyone already has an extremely powerful computer in their pocket.
18:46 Speaker 02 Hello?
18:47 Speaker 00 A modern smartphone runs over 5,000 times faster than one of the original 1980s supercomputers known as Cray II.
18:57 Speaker 02 High-performance computing is one of the cornerstones of Lawrence Livermore National Lab. Every three to five years, we'll get a new number one machine. Next is El Capitan, bigger and better than any other machine that's on the planet Earth.
19:12 Speaker 06 El Capitan, the fastest supercomputer in the world. About 1.7 exaflops. That means El Capitan can perform 1.7 quintillion math problems per second. Everything happens in a very short time span, but very complex physics going on. El Capitan is about 20 times faster than the lab's previous supercomputer, Sierra, dedicated in 2018. What took days or weeks on Sierra, now just hours on El Capitan.
19:41 Speaker 00 El Capitan's unclassified companion system, Tuolumne, a 288 petaphlop system using the same components as El Capitan, is currently the world's 10th most powerful supercomputer. It will be used for open science, including some of the team's drug discovery work. So how does toolamine compare to a smartphone? The device in your pocket operates at 2 to 2.5 teraflops, a measure of computing power.
20:12 Speaker 02 We're going to get 100 petaflops of dedicated time just for biology. And this probably will be the largest computing resource for biology in the world, just for biology problems.
20:24 Speaker 00 A petaflop is 1,000 times the power of one teraflop. And with a quick little math, that means Felice and her team will have over 50,000 times the computing power that your cellphone has.
20:40 Speaker 02 For dedication to biology, that's never been achieved before.
20:44 Speaker 00 But they can't simply plug in the computer and put their feet up. These operations don't just happen automatically.
20:51 Speaker 03 Part of the challenge of these types of problems is that it's not always easy to formulate the problem to be solved on a large computing system. It can take time to think about how to structure the problem in such a way that it could be solved with a large computing resources.
21:08 Speaker 00 In other words, you have to know exactly which puzzle piece you're looking for and give the computer the parameters to find it. And once you find it, you still have to see if it fits.
21:19 Speaker 01 One of the current bottlenecks is that it still takes quite a long time once we actually do the computational modeling. We still have to go out and make and test molecules with this. So what we're working on now is directly integrating our laboratories with these computational models. So the computational model can actually directly specify what experiment it needs to have done. Sometimes that's to validate what we're doing. Or sometimes it's because the model is uncertain on what the prediction should be. In that case, what it can do is it can say, I really need more experimental data with this set of molecules. So can you go run those experiments for me? Then we can actually have automated systems
22:04 Speaker 01 in the laboratory which actually run those experiments, bring that data back to the computational. Modeling systems, that system can update its models to improve their predictions. And then we keep going and we can have this iteration then between computational modeling and experimental make test cycles and have this positive feedback between those two that allows us to actually converge to solutions faster yet. We call that whole process active learning.
22:34 Speaker 00 It's finding, making and testing puzzle pieces.
22:38 Speaker 01 We still have lots of humans in that loop. We don't have an AI system going out and manufacturing molecules. There are humans in the loop doing that, but the system is really specifying exactly what it should be done.
22:48 Speaker 02 It also does take expertise. I don't want to minimize my team. They are a great team and they have the knowledge that it takes to recognize that certain parts of the drug molecule need interactions with certain parts of the protein molecule. And so by having that intuition or education really can focus the simulations and the screening that we do as well.
23:11 Speaker 00 But even though Felice, Jim and Jonathan have streamlined the machine learning process, drug development can still come to a screeching halt during the second and third phases, clinical trials.
23:23 Speaker 03 The longer you get down that pipeline going from initial discovery phase to actually clinical testing, you end up finding failures, particularly around safety, that can really slow down the whole pipeline. So one of the premises of the computational platform has been this idea that we want to try to evaluate these molecules upfront as early in the discovery process as possible for all of these different criteria that go into making a good drug. The idea is that if we can initially explore different parts of chemical space that meet all of our design criteria, then we will save some time later on in terms of having to go back and start over again.
24:07 Speaker 00 Usually what causes a drug to fail at the clinical level is a safety concern.
24:12 Speaker 03 A lot of what we're trying to do is predict, number one, is whether this chemical will bind and affect the activity of the protein target primarily, but then secondarily, is it going to be safe in the human body? Small molecules that are associated with cardio toxicity or liver toxicity and can typically make drug discovery fail later in the process as it gets closer to clinical trials.
24:39 Speaker 00 So by using AI to test these molecules before testing, Livermore is ensuring they don't have to go back to the drawing board multiple times. That's a great advantage, and they're applying the same techniques to address other drug problems, ones you're probably all too familiar with. Common side effects may include headache, dizziness, dry mouth, fatigue, upset stomach, increased heart rate, mild rash.
25:03 Speaker 03 But it can make small molecule drug discovery so challenging. There's a lot of things that we can ingest in the body that our body doesn't like, or can make us have negative side effects.
25:14 Speaker 02 If you're taking a drug and it's in your bloodstream, you want to make sure it hits its protein target and it doesn't go everywhere else in your body. This is where you would say you have side effects, and side effects are where it shouldn't be going. Like, you might feel nauseous, or you might have a rash or something like that, but you still want the drug to really work in where it's designed to work.
25:37 Speaker 03 We've been demonstrating this computational pipeline with this selective histamine inhibitor, a novel antihistamine that prevents drowsiness. The idea is to design a molecule that hits this histamine receptor target, but doesn't hit some of these other receptors that look like the target that have this negative side effect of drowsiness.
25:58 Speaker 00 That's just one example. Often, which targets companies are working on is kept tightly under lock and key to prevent competition. But there's one protein that everyone in the drug development space knows about.
26:12 Speaker 02 It's been called the undruggable target.
26:14 Speaker 00 It's the RAS protein.
26:16 Speaker 01 It's a protein that you actually need it. It actually initiates cell division. But when it's mutated, it can get stuck in an on state where it will just continually initiate cell division.
26:29 Speaker 00 And it's responsible for 30% of all cancers.
26:33 Speaker 01 It's implicated as the triggering point in some of the worst and most untreatable cancers, like 90% of pancreatic cancer, gastrointestinal and lung cancers.
26:42 Speaker 00 When a RAS protein mutates, it doesn't know when to stop creating more cells, in this case, cancerous ones. To make matters worse, some mutations can also create resistance to cancer treatments, making the disease even harder to combat.
26:59 Speaker 02 If we can actually find an effective drug that would target all of those proteins that are mutated, then there's some vision that it would cure 30% of all human cancers.
27:12 Speaker 00 That would be revolutionary.
27:14 Speaker 02 You have to have an open mind that many things are possible. And the beautiful thing about Lawrence Livermore is that if you really want to do something that is within the mission of the lab, you can do it.
27:27 Speaker 00 And they've partnered with other research facilities that are just as passionate about developing a treatment. Working with Frederick National Laboratory, Lawrence Livermore National Laboratory has developed methods to model cancer cell interactions with proteins.
27:42 Speaker 01 Some of that work has now led to work with a small company called Bridge Bio.
27:48 Speaker 00 They're a pharmaceutical company focused on addressing cancer.
27:52 Speaker 02 Bridge is providing great chemists, great insight in the cancer space and the ability to push and get to a new drug entity. Livermore is contributing the computational.
28:04 Speaker 02 prowess of the Department of Energy Labs. And to show that we can develop new methods and new technologies. What that comes down to is the people on both sides. It's the interaction at the human level. We have great experts in the computer space. We have great experts in the chemistry space and the ability to communicate and actually convince each other that we are able to do what we're able to do and produce new drugs.
28:33 Speaker 01 We actually have two new molecular designs now that are coming from Felice Lightstone and her group that are going into clinical trials right now. And so that's very exciting. So if this works, this will be transformational for cancer treatment.
28:47 Speaker 00 Shortening the time between drug discovery and getting pills to patients is a lofty goal, but Felice, Jonathan, Jim, and their incredible teams are tackling them head on.
28:59 Speaker 03 It's something like 750 FDA-approved drugs right now that are available on the market. What I would love to see is our ability to expand that pipeline of molecules and therapeutics that can get towards FDA approval much more efficiently and more quickly and cheaply. I'm hoping we will have a whole set of new molecules that we can bring to bear on different disease targets.
29:28 Speaker 00 By using state-of-the-art computing power and combining traditional chemistry and physics with machine learning and AI, they're already making huge strides in achieving that goal.
29:40 Speaker 02 In the last two years, we've had great success in being able to make a few drugs.
29:46 Speaker 01 One was just FDA-approved for the phase one trials, and those will be starting up in the fall. These are small molecule drugs, so they go through the standard stages of clinical trials. That will take up to a couple of years to have real feedback on those.
30:02 Speaker 03 I'm optimistic that we'll have a lot more therapeutic tools in the toolbox, so to speak, to treat various diseases in the next five to 10 years.
30:13 Speaker 00 We've come a long way from the Willow Tree, and the folks at Lawrence Livermore are confident that the lab's state-of-the-art computing technologies will bring better, faster drugs in the next few years. Whether it's a headache, virus, genetic disease, or cancer, no challenge is too great. Helping humanity, that's the priority.
30:56 Speaker 00 Thank you for tuning in to Big Ideas Lab. If you loved what you heard, please let us know by leaving a rating and a review. And if you haven't already, don't forget to hit the follow or subscribe button in your podcast app to keep up with our latest episode. Thanks for listening.
Previous Optics Recycle Loop Next Community Giving Page updated November 3, 2026 · transcript and summary are machine-generated