In the near future, AI will independently create new drugs without human intervention.

• AI has generated novel molecule designs that have surprised scientists at Eli Lilly during hypothetical drug discovery research.
• In 2021, Google's DeepMind AI made a major precedent in biology by creating a novel protein called AlphaFold, demonstrating its creative thinking in various fields such as strategy games, music, video, and cloud computing.
• In the near future, AI experts predict that AI will develop drugs that humans cannot conceive.

Diogo Rau, Eli Lilly's chief information and digital officer, recently engaged in unconventional experiments in the office, deviating from the usual drug research work associated with a major pharmaceutical company.

Lilly has been utilizing AI to sift through vast amounts of molecules. With AI capable of discovering molecules at a rate five times faster than traditional wet labs in just five minutes, it makes sense to explore the full potential of AI in medicine. However, there is no guarantee that the abundance of AI-generated designs will be effective in the real world, and this is something skeptical executives wanted to investigate further.

The AI-generated biological designs, which Rau described as having "unusual structures" and were not found in the company's existing molecular database, were presented to Lilly research scientists. Despite the executives, including Rau, expecting the scientists to dismiss the AI results.

Before presenting the AI results, he wondered if they could be that good.

Lilly executives were taken aback by the scientists' response, which was a surprise. Instead of criticizing the AI-generated designs, the scientists said, "It's interesting; we hadn't thought about designing a molecule that way," according to Rau.

"Rau stated that an epiphany occurred for him when he realized that while humans often focus on training machines, another form of art is where machines generate ideas based on data sets that humans wouldn't have been able to visualize. This opens up new pathways in medicine development that humans may not have otherwise explored, thereby spurring even more creativity."

The pharmaceutical industry is on track to see medicines completely generated by AI in the near future, according to some executives working at the intersection of AI and health care. Generative AI is rapidly accelerating its applicability to the development and discovery of new medications, reshaping not only the pharmaceutical industry but ground-level ideas that have been built into the scientific method for centuries.

When Google's DeepMind broke the protein mold

In 2021, the "AlphaFold moment" marked a significant milestone in the development of AI large language models in biology, as Google's DeepMind AI unit demonstrated the potential of AI to revolutionize drug development and design.

Biology is becoming increasingly digitized at unprecedented scales and resolutions, with AI advances taking place within this field.

The medical revolution involves spatial genomics scanning of millions of cells in tissue, in 3-D, and AI model-building that benefits from a catalog of chemicals already in digital form. This allows generative AI transformer models to work on them. The training can be done using unsupervised and self-supervised learning, and the AI can 'think' of drug models that a human would not.

The development of AI drugs can be compared to the functioning of ChatGPT, which has been trained on a vast amount of information and can generate answers to queries.

The GPT-version of drug discovery

Simulating biological behavior in computer models can predict how drugs might interact and work together, saving time and resources in drug discovery. With AI supercomputers and GPT-like methods, we can now represent the world of drugs in a computer for the first time, using digital biology data.

The traditional drug discovery process involves extensive experimentation, data collection, analysis, and another design process based on the results. However, this approach has a 90% failure rate. According to Powell, the process is highly artisanal and involves experimentation within a company followed by several decision points.

AI backers believe that the classic drug discovery process can be transformed into a more systematic and repeatable engineering process, allowing drug researchers to build off a higher success rate. According to Powell, citing results from recent studies published in Nature, Amgen was able to cut down the time it takes to discover a drug from years to months with the help of AI. Additionally, Powell noted that the success rate jumped when AI was introduced to the process early on, with the probability of success rising from 50/50 to 90% after a two-year traditional development process.

"We predict that the progress of drug discovery will significantly increase," Powell stated. The flaws of generative AI, such as its tendency to "hallucinate," could prove beneficial in drug discovery. "Over the past few decades, we have focused on the same targets. However, by utilizing a generative approach, we can explore new targets," Powell added.

'Hallucinating' new drugs

Protein discovery is an example of how biological evolution identifies a protein that works well and then moves on, without testing all other potential proteins. In contrast, AI can begin its work with non-existent proteins within models, allowing for a much larger discovery set to explore. With a potential number of proteins that could act as a therapy essentially infinite, Powell said, the existing limit on working with proteins nature has given humanity is exploded. AI can use models to "hallucinate" proteins with all the functions and features needed, going beyond what a human mind could imagine.

Recently, the University of Texas at Austin acquired a massive NVIDIA computing cluster for its new Center for Generative AI.

"Just as ChatGPT can learn from strings of letters, chemicals can be represented as strings, and we can learn from them," said Andy Ellington, professor of molecular biosciences. AI is learning to distinguish drugs from non-drugs, and to create new drugs, in the same way that ChatGPT can create sentences, Ellington said. "As these advances are paired with ongoing efforts in predicting protein structures, it should soon be possible to identify drug-like compounds that can be fit to key targets," he said.

Daniel Diaz, a computer science postdoctoral fellow at UT's Institute for Foundations of Machine Learning, leads the deep proteins group and believes that the future of AI in drug discovery lies in the development of novel biologics, where he has already observed how AI can accelerate the process of identifying the most effective designs.

His team is currently conducting animal studies on a therapeutic for breast cancer that utilizes an engineered version of a human protein designed to degrade a critical metabolite that breast cancer relies on, effectively starving the cancer. Typically, when scientists seek a protein for therapeutics, they prioritize stable proteins that are less likely to break down easily. However, achieving this stability requires genetic engineering to modify the protein, a time-consuming process in laboratory work that involves mapping the structure and selecting the most suitable genetic changes.

AI models are helping scientists identify the best modifications to try more quickly, resulting in a sevenfold increase in protein yield in an experiment cited by Diaz. This protein is human-based, reducing the risk of allergic reactions in patients.

Nvidia has introduced "microservices" for AI healthcare, including drug discovery, as part of its efforts to promote the adoption of AI in the health sector. This technology enables researchers to screen for trillions of drug compounds and predict protein structures. Meanwhile, computational software design company Cadence is integrating Nvidia AI into its molecular design platform, allowing researchers to generate, search, and model data libraries with hundreds of billions of compounds. Additionally, Cadence is offering research capabilities related to DeepMind's AlphaFold-2 protein model.

"We've simplified AlphaFold so that biologists can easily use it without needing extensive training. Simply input an amino acid sequence into our webpage and the actual structure will be displayed. If you were to perform this task with an instrument, it would cost $5 million and require three full-time equivalent workers, taking a year to obtain the structure. We've made this process instantaneous through our webpage," Powell stated.

The success or failure of AI-designed drugs will depend on their performance in human trials, which is the final step in drug development.

"You still have to generate ground proof," Powell said.

Just like in the training of self-driving cars, where data is constantly collected to improve and enhance models, the same process is happening in drug discovery," she stated. "Using these methods, you can explore new spaces, refine and refine models, conduct more intelligent experiments, and feed the data back into the models in a continuous loop.

The biology models number in the tens of billions, while the AI industry is in the range of a trillion in areas of multi-modal and natural language processing.

"Powell stated that we are in the early stages, and an average word has less than ten letters. In contrast, a genome is 3 billion letters long."