The application prospects are broad! AI will become the ‘main force’ in the field of drug discovery

According to the Zhitong Finance APP, Lilly (LLY.US) has been using generative artificial intelligence to discover drug molecules. The data shows that the number of drug molecules discovered by artificial intelligence in five minutes is equivalent to the number of molecules synthesized by Eli Lilly in a traditional laboratory for a whole year, so testing the limits of artificial intelligence in the medical field is meaningful. We cannot know whether a large number of designs generated by artificial intelligence can be effective in the real world, which is exactly what skeptical company executives want to know more about.
Diogo Rau, Chief Information and Digital Officer of Eli Lilly, recently participated in some atypical experiments using artificial intelligence to generate drug molecules. He described these AI generated biological designs as molecules with “strange structures” that cannot be matched with the company’s existing molecular database, but appear to be potential powerful candidate drugs. These drug molecules generated by artificial intelligence were handed over to research scientists at Lilly and surprised them.
According to executives working in the intersection of artificial intelligence and healthcare, in the near future, the field will be entirely powered by AI generated drugs. Some industry insiders suggest that it will become a standard for drug discovery within a few years at most. Generative artificial intelligence is rapidly accelerating its application in the development and discovery of new drugs. This move will not only reshape the pharmaceutical industry, but also reshape the fundamental ideas that have been integrated into scientific methods for centuries.
Google DeepMind Becomes a Pioneer
The progress related to artificial intelligence is occurring in the field of biology, which is increasingly digitized at an unprecedented scale and resolution, as described by Kimberly Powell, Vice President of Healthcare at NVIDIA.
This change actually occurred a few years before OpenAI’s ChatGPT became familiar to the public. In 2021, Google’s DeepMind artificial intelligence division took the lead in applying artificial intelligence big language models to biology. Kimberly Powell said, “We can train these deformation models with very large datasets, from amino acid sequences to protein structures, which are the core of drug development and design
This is a medical revolution that includes spatial genomics that scans millions of cells within tissues in 3D, as well as AI model construction that benefits from already existing catalogs of chemical substances in digital form, allowing generative artificial intelligence transformer models to work on them now. Kimberly Powell said, “This kind of training can be accomplished through unsupervised and self supervised learning, and it can not only be done quickly, but also imaginatively. Artificial intelligence can ‘think’ about drug models that humans cannot achieve
The mechanism of ChatGPT can serve as an analogy for understanding artificial intelligence drug development. Kimberly Powell said, “It’s basically trained on every book, every webpage, and every PDF file. It encodes the world’s knowledge in such a way that you can ask it questions and it can provide you with answers
GPT version of drug discovery
Drug discovery is a process of witnessing biological interactions and changes, but it takes months or years in the laboratory and can be reflected in computer models that simulate traditional biological behavior. Kimberly Powell said, “When you can simulate their behavior, you can predict how they work together and interact.” “We now have the ability to represent the drug world – biology and chemistry – because we have artificial intelligence supercomputers, using artificial intelligence and GPT like methods, with all the digital biological data, we can represent the drug world in a computer for the first time
This is completely different from the classic empirical approach that dominated drug discovery in the last century: extensive experimentation, subsequent data collection, data analysis at the human level, and then another design process based on these results – conducting experiments within the company, followed by several decision points that scientists and executives hope will lead to successful clinical trials. Kimberly Powell said, “This is a very manual process. Therefore, it is a drug discovery process with a failure rate of up to 90%
Supporters of artificial intelligence believe that this will save time, increase success rates, transform classical processes into more systematic and reproducible engineering, and enable drug researchers to establish higher success rates. Kimberly Powell cited recent research published in the journal Nature and pointed out that Anjin found that with the help of artificial intelligence, the discovery process of a drug, which could have taken years, can now be shortened to a few months. More importantly, considering the cost of drug development (which may range from $30 million to $300 million per trial), introducing artificial intelligence into this process in the early stages significantly increases the success rate. After two years of traditional development process, the probability of success is 50%. Kimberly Powell stated that the success rate increases to 90% at the end of the faster artificial intelligence enhancement process.
Kimberly Powell said, “We predict that the progress of drug discovery should significantly improve.” Some noteworthy flaws of generative artificial intelligence, such as its tendency towards “hallucinations,” may play an important role in drug discovery. Kimberly Powell added, “Over the past few decades, we have been searching for the same targets, but what if we could use generative methods to open up new ones
The New Drug Discovered through Illusion
Artificial intelligence can start working from proteins that do not exist in the model, which is untenable in classical empirical models. Numerically speaking, artificial intelligence has greater discoveries to explore. Kimberly Powell stated that the potential number of proteins that can be used as a therapeutic method is essentially infinite – to the power of 10 ^ 160, or to the power of 10 ^ 160- and the limitations of utilizing naturally endowed proteins in humans will be broken. You can use these models to generate ‘illusion’ proteins that may have all the functions and features we need. It can do things that the human brain cannot do, but computers can do them
The University of Texas at Austin recently purchased one of the largest NVIDIA computing clusters for its new generative artificial intelligence center. Molecular biology professor Andy Ellington said, “Just like ChatGPT can learn from letter strings, chemicals can also be represented by strings, and we can learn from them.” He said that artificial intelligence is learning to distinguish between drugs and non drugs and create new drugs, just like ChatGPT can create sentences. “With these advances combined with continuous efforts to predict protein structures, it should soon be possible to identify drug like compounds suitable for key targets
Daniel Diaz, a postdoctoral fellow in computer science at the Institute of Machine Learning Fundamentals at the University of Texas, stated that most of the current work on artificial intelligence in the field of drugs is focused on small molecule discovery. However, he believes that the greater impact will be on the development of new biologics (protein based drugs), and he has seen how artificial intelligence can accelerate the process of finding the best design.
Daniel Diaz’s research team is currently conducting an animal experiment to treat breast cancer, which is an engineered version of human protein. This protein can degrade a key metabolite that breast cancer depends on. Traditionally, when scientists need a protein for treatment, they look for several characteristics, including stable proteins that are not easily broken down. This requires scientists to introduce genetic engineering to adjust proteins, which is a tedious process in laboratory work – drawing structures and identifying the best choice from all possible gene modifications.
Now, artificial intelligence models are helping to narrow down possibilities, so scientists can know the best modifications to try faster. In the experiment cited by Daniel Diaz, the use of a more stable AI enhanced version resulted in an approximately sevenfold increase in protein production, thus allowing researchers to ultimately have more proteins available for testing and use. Daniel Diaz said, “The results look very promising.” And because it is a protein based on the human body, the likelihood of patients being allergic to this drug – an allergic reaction to protein based drugs is a big issue – is minimized.
Nvidia recently released so-called artificial intelligence medical “microservices,” including drug discovery – part of the company’s ambitious use of artificial intelligence in the healthcare field – enabling researchers to screen trillions of drug compounds and predict protein structures. Cadence, a computing software design company, is integrating Nvidia AI into a molecular design platform that allows researchers to generate, search, and model databases using billions of compounds. It also provides research functionality related to DeepMind’s AlphaFold-2 protein model.
Ultimately, the drugs designed by artificial intelligence will depend on the traditional final step of drug development: the performance of human trials. Kimberly Powell said: “You still need to show solid evidence.” She compared the current level of progress with the training of autonomous vehicle – autonomous vehicle are constantly collecting data to strengthen and re enhance the model. Kimberly Powell said, “The same thing happens in the field of drug discovery. You can use these methods to explore new spaces… hone it, hone it… conduct smarter experiments, obtain experimental data and feed it back into the model, and repeat the process
However, compared to the broader field of artificial intelligence models, the space in the biological field is still very limited. In the fields of multimodal and natural language processing, the artificial intelligence industry has models with one trillion or more parameters, while biological models have only tens of billions of parameters.