Diogo Lau, Eli Lilly's chief information and digital officer, was recently involved in some experiments in his office, a typical pharmaceutical research job you'd expect from tinkering in a lab within a major pharmaceutical company. It wasn't.
Lilly has used generative AI to search millions of molecules. Testing the limits of artificial intelligence in medicine as AI can move at the speed of discovery that allows him to generate as many molecules in five minutes as Lilly takes him a year to synthesize in a traditional wet lab. makes sense. But there was no way to know whether the rich designs generated by AI would work in the real world, and skeptical executives wanted to know more about it.
The top AI-generated biological design has a “strange structure” that doesn't match well in the company's existing molecular database, but it looks like a potentially strong new drug candidate, Lau said. The molecule was brought to Lilly's research scientists. Lau and other executives expected scientists to reject the AI results.
“They can't be this good?” he recalled thinking before presenting the AI results.
The scientists were expected to point out all the flaws in the AI-generated design, but their responses came as a surprise to Lilly executives. “I said, 'That's interesting. I hadn't thought of designing molecules that way,'” Lau recalls. They said this while telling attendees this previously unreported story at the CNBC Technology Executive Council Summit last November.
“That was an epiphany for me,” Lau said. “We always talk about training machines, but another art is for machines to generate ideas based on data sets that humans can't see or visualize. It opens up new avenues for drug development and further stimulates creativity.''I haven't looked into anything else. ”
Executives working at the intersection of AI and healthcare say the field is on track to see medicines generated entirely by AI in the near future. Some say it will become the standard for drug discovery within a few years at most. Generative AI is rapidly accelerating its application to new drug development and discovery, not only in the pharmaceutical industry, but also as a movement to reshape fundamental-level ideas that have been embedded in the scientific method for centuries. It has become.
When Google's DeepMind broke the protein mold
This trajectory first became apparent several years before ChatGPT entered the public consciousness. According to Kimberly Powell, Nvidia's vice president of healthcare, 2021 was “AlphaFold's moment.” Google's DeepMind AI division became famous for showing how creative thinking in AI differs from humans in Chinese strategy games. Go — Pioneers the application of AI large-scale language models to biology. “AlphaFold was a pivotal moment in being able to train these transformer models using very large datasets and go from amino acid sequences to protein structures. and core,” Powell said.
Advances related to AI are occurring in the increasingly digital field of biology, which Powell describes as “unprecedented scale and resolution.”
This will particularly benefit from spatial genomics, which scans millions of cells in a tissue in 3D, and AI model building, which will particularly benefit from a catalog of chemicals already in digital form that will enable generative AI Transformer models to run. It is a medical revolution that includes. they. “This training can be done using unsupervised learning and self-supervised learning, and can be done not only quickly but imaginatively. AI can ‘think’ drug models that humans cannot think of. We can,” Powell said.
A parallel for understanding AI drug development can be found in the mechanics of ChatGPT. “It's basically been trained on every book, every web page, every PDF document, and it encodes the world's knowledge in such a way that users can ask questions and generate answers,” Powell said. Stated.
GPT version of drug discovery
Drug discovery is a process of observing interactions and changes in biological behavior, but work that would take months or even years in the laboratory can be represented by computer models that simulate traditional biological behavior. “And if you can simulate their behavior, you can predict how things will work together and interact,” she said. “We now have the ability to represent the drug world, including biology and chemistry. We have AI supercomputers that use AI and techniques like GPT, and we use all the digital biological data to , because it allows us to represent the world of drugs within a computer.''
This is due to the classic empirical approach that has dominated drug discovery for the past century: extensive experimentation, subsequent data collection, human-level data analysis, and then a separate design process based on those results. It's a fundamental departure. In-house experimentation and subsequent several decision-making points are where scientists and executives count on successful clinical trials. “It's a very artisanal process,” Powell said. The result is a drug discovery process with a 90% failure rate.
AI advocates believe that AI can save time and improve success rates, converting classical processes into more systematic and reproducible engineering, and helping pharmaceutical researchers build higher success rates. I am. Citing recent research published in Nature, Powell noted that Amgen has discovered that the drug discovery process, which once might have taken years, can be shortened to months with the help of AI. More importantly, given that the cost of drug development ranges from $30 million to $300 million per clinical trial, early on he was confident that the introduction of AI into the process would improve success rates. It jumped up. Traditionally he had a 50/50 chance of success after going through a two year development process. According to Powell, at the end of the faster AI expansion process, the success rate rose to 90%.
“I predict that there will be significant advances in drug discovery,” Powell said. Some noted flaws of generative AI, such as its propensity for “hallucinations,” could prove powerful in drug discovery. “For the past decades, we've been looking at the same targets, but what if we could use a generative approach to explore new targets?” she added.
New drug that causes hallucinations
The discovery of proteins is one example. Biological evolution works by identifying proteins that work well, and then nature evolves. We do not test all other proteins that may function similarly or more effectively. On the other hand, AI can start working from proteins that are not present in the model, an approach that cannot be sustained with classical empirical models. By the numbers, even bigger discoveries await AI. The number of proteins that could potentially act as a therapeutic is essentially infinite, so 10 to the power of 160, or 0 plus 160, is 10, Powell said, adding that the number of proteins that nature has given humans is He said that the existing limitations in dealing with the world are exploding. “With these models, we can hallucinate proteins that might have all the functions and properties that we need. We can do it in places that human thinking can't, but computers can. ,” Powell said.
The University of Texas at Austin recently purchased one of the largest NVIDIA compute clusters for its new Center for Generative AI.
“Just as ChatGPT can learn from strings, chemicals can be represented as strings and we can learn from them,” said Andy Ellington, professor of molecular biosciences. Just as ChatGPT can create sentences, Ellington said, the AI is learning to distinguish between drugs and non-drugs and create new drugs. “These advances, coupled with continued efforts to predict protein structure, will allow us to identify drug-like compounds that match key targets,” he said.
Daniel Diaz, a postdoctoral researcher in computer science who leads the deep protein group at UT's Machine Learning Fundamentals Laboratory, said that while most of the current AI drug research is focused on small molecule discovery, the larger impact will be on drug discovery. He said he thinks it is due to development. We are already seeing how AI can speed up the process of finding optimal designs for new biologics (protein-based medicines).
His group is currently conducting animal experiments on breast cancer drugs. The treatment is a modified human protein that breaks down key metabolites that breast cancer depends on, essentially starving the cancer. Traditionally, when scientists need a therapeutic protein, they look for a few characteristics, such as a stable protein that doesn't fall apart easily. This requires scientists to introduce genetic engineering to fine-tune proteins, a laborious process in the lab that involves mapping the structure and identifying all possible genetic modifications. The idea is to identify the best option.
Now, AI models are helping narrow down the possibilities so scientists can learn more quickly about the best changes to try. In experiments cited by Diaz, using his more stable AI-enhanced version improved protein yields by about seven times, allowing researchers to ultimately test and use more proteins. became. “The results look very promising,” he said. Also, because this is a protein of human origin, there is a possibility that patients will be allergic to the drug (allergic reactions to protein-based drugs are big problem) is minimized.
Nvidia's recent release of what it calls “microservices” for AI healthcare, including drug discovery (part of the company's aggressive ambitions to bring AI into the medical field), will help researchers generate trillions of drug compounds. It is now possible to screen and predict the structure of proteins. Computational software design company Cadence is integrating her Nvidia AI into a molecular design platform that allows researchers to generate, search, and model data libraries containing hundreds of billions of compounds. He also provides research capabilities related to his AlphaFold-2 protein model in DeepMind.
“AlphaFold is difficult for biologists to master, so we simplified it,” Powell said. “You go to a web page and enter an amino acid sequence, and it shows you the actual structure. If you were to do that with a machine, the machine would cost him $5 million, and he would need three of them. [full-time equivalent workers] If done in FTE, the structure could be completed within a year. We did that instantly on a web page,” Powell said.
Ultimately, AI-designed drugs will succeed or fail based on their performance in human trials, the traditional final step in drug development.
“We still need to make sure it's earth resistant,” Powell said.
She likened the current level of progress to training for self-driving cars. In self-driving cars, data is continuously collected to enhance and re-enrich models. “The exact same thing is happening in drug discovery,” she says. “Using these methods, we can explore new spaces… honing, honing… doing more intelligent experiments, taking that experimental data and feeding it back into our models, looping… to rotate.”
However, the biological space within the broader AI modeling field is still small by comparison. The AI industry is in the range of over 1 trillion models in the field of multimodal and natural language processing. By comparison, the number of biological models is in the tens of billions.
“We're still in the early stages,” Powell said. “The average word is less than 10 characters long. The genome is 3 billion characters long.”
