The most familiar artificial intelligence (AI) applications include text generation, facial recognition, and autonomous driving. But there are other AI applications that deserve their share of attention. Surely, one of these worthy but relatively unsung AI applications is drug development. Its lifesaving potential would be hard to exaggerate. To explore the status of AI-driven drug development, GEN talked with four experts in the industry.
Customized for chemistry
AI is far from a one-size-fits-all approach. “The AI methods that are needed for ‘killer apps’ like computer vision are not the same as those needed for natural language,” says Evan Feinberg, PhD, CEO of Genesis Therapeutics, headquartered in Burlingame, CA. “Generally, AI architectures were not designed for chemistry and physics.”
In the 2010s, while earning a doctorate in a laboratory led by Vijay Pande, PhD, at Stanford University, Feinberg did his part to expand AI’s remit. Specifically, he contributed to the development of machine learning methods that could advance the development of small-molecule drugs.
In 2019, Feinberg co-founded Genesis. He recalls that the company, from its very inception, has been “devoted to the development of novel, differentiated AI specifically for the purpose of drug discovery.”
The work at Genesis produced the Genesis Exploration of Molecular Space (GEMS) platform, which combines generative and predictive AI methods, allowing Genesis chemists to create, score, and rank molecules in silico. Feinberg and his colleagues use GEMS to explore disease targets that are difficult to drug or are not impacted by any known chemical compound.
“If you’re going to deploy an AI model for those targets, it must be able to extrapolate to new areas of chemical and biological space,” Feinberg insists. So, instead of interpreting existing data as most AI models do, GEMS leverages AI and physics to extrapolate data into the unknown. The company uses GEMS from hit identification through lead optimization and candidate nomination.
Currently, the lead candidate at Genesis is a small molecule that inhibits phosphatidylinositol 3-kinase α (PI3Kα), which is mutated in many forms of cancer and drives the growth and expansion of tumor cells. Historically, this target has proved difficult to drug selectively because many existing drugs poorly distinguish between the normal, wildtype PI3Kα protein and the cancer-causing mutant. Using GEMS, though, Feinberg’s team is optimizing a small-molecule compound that is, in Feinberg’s estimation, “selective for, and able to inhibit, all the most prevalent mutations of PI3K, while sparing the normal PI3K the body needs for things like regulating blood sugar.”
Although Feinberg adds that GEMS was essential to selecting and developing a PI3Kα inhibitor, he also gives credit to the biologists, chemists, and pharmacologists “who have all played a critical role.” He describes the process as AI and humans working together in “a steadily advancing way.”
A small-molecule score
Small molecules are also the focus of Exscientia, which is a drug design and development company headquartered in Oxford, U.K. “Internally, we are pursuing precision oncology indications,” says John Overington, PhD, Exscientia’s chief technology officer. “With partners, we are successfully pursuing neuroscience, immunology, and rare disease.”
In August, Exscientia announced that it would be a party in the closest form of business partnership, a merger. In this case, the merger involves Exscientia and clinical-stage biotechnology company Recursion Pharmaceuticals, headquartered in Salt Lake City, UT.
At present, Exscientia has drugs in studies from early discovery through Phase I/II trials. The company’s work on these drugs relies on two AI-based approaches: generative AI and large language models (LLMs). “Drug-like chemical space is currently too large to exhaustively search, and generative-AI methods allow for very efficient exploration and mapping of the chemical landscape active against the target,” Overington explains. “It is also important to ensure patent novelty in design, so timely data integration is essential in the highly competitive commercial area of drug design, and this AI-enhanced efficiency is essential for our drug optimization approaches.”
Although next-generation generative chemistry algorithms already exist, Overington believes that “the big challenge in the field now is predicting—scoring—the compounds for physicochemical, ADME [absorption, distribution, metabolism, excretion], and bioactivity properties.” When a compound gets a high score, though, it must be made for further testing. For that, Exscientia is “developing and combining AI methods to assess synthetic accessibility of the designs,” Overington says. “In our hands, this end-to-end ‘design-score-synthesize’ approach is an incredibly powerful way to make drug discovery more efficient and less costly.”
In addition to that method based on generative AI, Exscientia uses LLMs to “support, enhance, and automate the work of the drug design teams and validation biologists,” Overington relates. “We can leverage external documents, such as papers, patents, and abstracts, alongside internal proprietary data and documents. This capability has accelerated our progress in identifying strategies for drug optimization.”
Building a multifaceted method
Insilico Medicine, a clinical-stage biotechnology company headquartered in Boston, MA, uses a range of AI-based tools. These include components of the company’s Pharma.ai drug discovery suite such as PandaOmics (an analytical tool for therapeutic target and biomarker discovery), Chemistry42 (a platform for the de novo generation of novel small molecules), and inClinico (a data-driven multimodal platform for predicting the probability of successfully transitioning from Phase II to Phase III).
As one example, Insilico Medicine targeted the TRAF2- and NCK-interacting kinase (TNIK) to treat kidney and pulmonary fibrosis by using a process known as a random walk on heterogeneous graphs. The company’s scientists used this method because it “allows for the exploration of potential connections between entities—such as genes, proteins, and diseases—across different types of biological and biomedical data,” explains Thomas Leichner, Insilico Medicine’s head of strategy. “This approach is particularly effective in identifying novel relationships that may not be apparent through conventional analysis methods.”
To handle the large and diverse datasets, Insilico Medicine uses “negative matrix factorization to decompose large biological datasets into lower-dimensional representations, facilitating the identification of hidden structures within the data,” Leichner points out. “This is especially useful in uncovering disease-related patterns and potential therapeutic targets that are not easily observable in high-dimensional space.”
Once all of the data gets analyzed, a company needs an efficient way to decide how to use it. Here, Insilico Medicine’s “platform employs a combination of score compositions and filters to generate a ranked list of targets,” Leichner relates. “These filters include disease-agnostic properties such as protein family, accessibility by small molecules or therapeutic antibodies, novelty, and crystal structure availability.”
As Insilico Medicine’s AI-based approach to drug discovery shows, multiple tools and techniques are required. So, this company’s multifaceted approach, says Leichner, “ensures that the targets identified are not only relevant to the disease of interest but also possess characteristics that make them amenable to drug development.”
Insilico Medicine is already accelerating drug discovery. For example, its AI-based tools picked 20 preclinical candidates in an average of 13 months. Moreover, Insilico Medicine’s lead TNIK-targeting asset for idiopathic pulmonary fibrosis “recently had a positive clinical readout for its Phase IIa trial in China,” Leichner says. “We are also running a U.S.-based Phase IIa trial for this asset.”
A multimodal transformer
At San Diego-based Iambic Therapeutics, chief technology officer Fred Manby, PhD, and his colleagues used a combination of an AI-driven drug-discovery platform and a highly automated experimental platform to go from launching a drug candidate to filing an Investigational New Drug application in just 24 months. This candidate, IAM1363, is a tyrosine kinase inhibitor that targets wild-type and mutant HER2, which is expressed in many cancers. In March, Iambic started a Phase I trial on IAM1363.
Iambic Therapeutics will accelerate drug discovery even more with its recently announced Enchant, which Manby describes as “a multimodal transformer model—an AI model that is trained very widely across different modalities of data and different sources of data from the whole slew of discovery activities.”
According to Manby, the key challenge in drug discovery is the scarcity of clinical data, which persists despite a wealth of laboratory data on potential drugs. “If there isn’t enough clinical data, the magic of AI can’t be leveraged to predict clinical outcomes—not as well as it can be leveraged to predict the preclinical properties of molecules,” he says. “The breakthrough of Enchant is that this model gets better at predicting clinical properties by being trained on more preclinical data.”
To do this, Enchant employ a range of technologies. First, the transformer architecture, best known as the engine for ChatGPT, is a neural network that seeks associations between data. Enchant’s multimodal capabilities allow it to learn from very different types of data, including molecular properties, genomics and other omics data, biomedical literature, knowledge graphs, and computed data.
“We’ve built a huge amount of our data-handling infrastructure with Amazon Web Services,” Manby remarks. “We also have a longstanding relationship with Nvidia on how you efficiently parallelize over multiple GPUs to train these models.”
The power of Enchant depends on access to large amounts of data in the right form for AI-based processing. Manby explains, “An enormous fraction of the effort on this project has simply been building the data infrastructure to be able to gather together these different types of data and then codify them into some systematic format that can be used for model training.”
The experimental platform developed at Iambic “creates hundreds or thousands of distinct molecular structures on weekly time scales,” Manby says. “Then, we can carry them to a whole suite of different biological assays, but also metabolic assays, and we do a set of biophysical measurements.”
Iambic Therapeutics measures multiple drug-related properties of each compound. The resulting data, which are collected automatically, inform decision making about the compound and sharpen the company’s computational models. For example, Iambic Therapeutics showed recently that Enchant can predict a drug candidate’s human pharmacokinetics.
Ultimately, drug discovery is about patients. Beyond creating more treatments at a much faster pace, AI-based drug discovery can also reduce the human burden during testing. As Manby says, “The more that you can accurately predict human dose and also accurately predict safety and efficacy in human patients, the less the burden of a clinical trial on the participants.”
