Experts have been raising concerns about “AI washing.” One of these experts, futurist Bernard Marr, has said that AI washing “involves exaggerating the capabilities of a product or service that is sold as ‘AI’ in order to make it look more sophisticated, innovative, or intelligent than it actually is.” This statement, which appeared in Forbes on April 25, 2024, was accompanied by a warning that AI washing could lead to unrealistic expectations, bitter disappointments, and gnawing doubts. If that were to happen, even true AI innovations would be met with cynicism.
So, is AI washing adversely impacting the field of drug design? Less than one might suppose, suggests Kfir Schreiber, co-founder and CEO of DeepCure. He points out that emerging technologies in drug discovery often follow a familiar pattern: they start with great excitement, face a period of skepticism, and eventually gain renewed interest as they begin to deliver real results.
Schreiber believes that AI, especially machine learning, has moved past the skepticism phase and is now delivering results. Such results are discussed in this article. It describes how four companies are leveraging AI to engage previously undruggable targets, to reduce reliance on animal testing, and to predict adverse effects that could occur in clinical trials.
Solving the most challenging problems with AI
“First-generation, first-wave AI technologies for drug design were great at improving efficiency,” Schreiber remarks. “It is the novelty that we are missing, and that is exactly what we are trying to do at DeepCure.
“Typically, scientists and drug discovery experts start by screening massive libraries of small molecules with the hope of finding a starting point for a long optimization process. [They] throw everything but the kitchen sink and hope something sticks.”
Schreiber points out that DeepCure prefers to flip the process. Rather than trust in blind searches, the company uses intentional engineering when developing small-molecule drugs for challenging therapeutic targets. He says, “Instead of just screening massive libraries and hoping that something will stick, we design the drug from the beginning to meet specific requirements.”
Because there are many parameters that go into the design of small-molecule drugs, DeepCure employs a multiparametric approach, one that simultaneously optimizes multiple drug characteristics. “The drug obviously needs to bind to the target of interest,” Schreiber adds. “Beyond that, it must be selective, soluble in various fluids, orally bioavailable, and safe against multiple toxicology endpoints.”
DeepCure’s platform, called MolGen, builds custom libraries for a specific set of requirements. For instance, MolGen might consider the target of interest, the desired safety profile, the route of administration, and the pharmacokinetics profile. “The key,” Schrieber emphasizes, “is that we take these considerations into account from the very beginning of the design cycle and not in the advanced stages of discovery.”
DeepCure’s lead program involves a third-generation BRD4 BD2 inhibitor. It is an epigenetic modulator that has been heavily implicated in rheumatoid arthritis. “By applying our technology, we were able to engineer a molecule that is selective to a very specific site for one member of the BET protein family,” Schreiber asserts. “By doing so, we have mitigated the historical toxicology and safety liabilities of this class of compounds.” The company is also developing novel drugs to target transcription factors, a protein class traditionally considered undruggable.
Calculating “credit scores” to predict clinical success
Jo Varshney, PhD, founder and CEO of VeriSIM Life, explains that, unfortunately, translatability between animal studies and clinical studies is often very poor, with about 5% success in many cases. “There is a lot of inefficiency in doing animal experimentation,” Varshney observes. “But if you can take that work out, you can eliminate unnecessary animal testing, enhancing both efficacy and animal welfare.”
VeriSIM Life combines AI methods with mechanistic mathematical models. The result: hybrid models. “If you have a hybrid car, you use gas and electricity,” Varshney says. “Likewise, in our case, we combine mathematical models and machine learning into our hybrid AI mechanistic models.”
VeriSIM Life, a company that refers to itself as a “de-risker for breakthrough drug design and development,” has developed BIOiSIM, a hybrid drug development platform that combines AI methods and mechanistic models. BIOiSIM is designed to predict the likelihood of a drug candidate’s success in clinical trials. For example, the platform can process preclinical data to calculate a Translational Index score, which is analogous to a FICO credit score.
The company has created mathematical representations of different systems and organs across various animal species, including humans. “We have codified all these different organ systems like the heart and the kidney,” Varshney details. “We have also codified the key differentiation between rats, monkeys, mice, and humans.” The company’s platform, BIOiSIM, then runs simulations with real-world data to clarify how different molecules interact within these codified systems.
VeriSIM has also developed a unique scoring system to gauge its confidence in a drug. “Our score is like a credit score that informs you about your financial health,” Varshney says. “However, our score is designed to tell you the likelihood of clinical success of this drug before the drug enters the clinic. Does there need to be a change in the formulation, the dosing, or the entire molecule?”
Among other advances, the company has received FDA orphan drug designation for a new inhalation formulation to treat pulmonary arterial hypertension, a rare disease whose mechanisms are unclear. According to Varshney, the formulation is “convenient for patients and eliminates all toxicities.”
Varshney notes that using BIOiSIM has rapidly accelerated programs from early-stage to IND-enabling studies, resulting in up to 50% reductions in R&D costs. The company has also expanded its technology to investigate antibody-drug conjugates, peptides, and bacteriophages for phage therapy.
Using robots to perform AI-designed experiments
Quentin Perron, PhD, co-founder and CSO of Iktos, says that his company generates innovative molecules by using an AI-driven retrosynthesis platform. Basically, the company uses generative AI to find ways of breaking down a complicated target molecule and identify simpler precursor structures that could be channeled into an efficient synthetic pathway.
Perron explains that Iktos uses medicinal and computational chemistry knowledge to enable AI-driven chemical space exploration: “Our proprietary generative AI, trained on millions of organic reactions, generates molecules like a chemist by leveraging commercial building blocks and organic reactions over several steps.” He also asserts that Iktos has developed the first AI-driven platform for de novo drug design that is user-friendly and focused on multiparameter optimization.
Iktos applies AI and robotic technologies to research in medicinal chemistry and drug design. The company’s technologies include Spaya, an SaaS platform for retrosynthesis; Makya, an SaaS platform for generative drug design; Ilaka, a workflow orchestration AI platform; and Iktos Robotics, an AI-driven synthesis automation platform that can accelerate the design-make-test-analyze cycle in drug discovery.
Perron says that the platform, which is called Makya, enables the “design of novel and easy-to-make compounds with a multi-objective blueprint and unprecedented speed.” He adds that Makya leverages advanced 3D structure–based design to enhance its drug discovery capabilities.
Iktos is also leveraging a sophisticated robotics platform, which makes molecules and generates real-life data faster. Perron says that the platform automates compound synthesis, facilitates high-throughput synthesis, and is already capable of managing 42 main reactions. He adds that the platform can manage about 100 reactions in a day, whereas a chemist in a traditional laboratory can manage only 2 or 3 reactions in a day.
Should chemists worry that they will one day be replaced with Iktos’s robots? Not according to Perron. He says, “Thanks to our automated platform, the chemist will be able to oversee more experiments and evaluate more molecules than ever before.”
Seeing drug design as a learning problem, not a screening problem
Like other AI-driven drug development companies, Exscientia employs a multiparametric approach to solve multiple problems simultaneously. The company’s vice president of medicinal chemistry, Thorsten Nowak, PhD, says, “We aim to design the smallest possible set of small molecules needed to identify a preclinical candidate molecule that hits the required target and has the overall properties required to make a drug to treat patients.”
In a recent Securities and Exchange Commission filing, Exscientia contrasted its multiparameter optimization approach with conventional optimization approaches, which usually start with improvements to target potency, then selectivity, then other properties. In Exscientia’s view, working sequentially “often leads to suboptimal molecules.” The company asserted that its platform “can design against more complex endpoints than have been conventionally possible.”
“When designing truly innovative drugs, there will be insufficient information available at the start of the project, and the right solution will almost certainly not already exist in big datasets or screening libraries,” Exscientia continued. “In other words, drug design is a learning—not a screening—problem. This is true for both novel targets, where no work has been done before, and established targets, where new approaches must be devised that are distinct from existing efforts. As we start to explore novel chemical spaces, we are likely to be at the limit of predictive power, or the domain of applicability, for current models. Our systems and models are designed to learn and evolve, which, like nature, allows them to find optimized solutions to problems.”
In general, Exscientia couples predictive modeling with experimentation in tightly knit design-make-test-learn cycles. To progress through these cycles at ever faster rates, the company relies on its recently launched automation facility, which has capabilities in compound management, automated chemical synthesis, and automated biological screening. The company expects that the facility will eventually enable the production of proteins and the development of drug metabolism and pharmacokinetics assays.
Exscientia’s drug pipeline highlights the early clinical success of AI-designed drugs. Upcoming milestones for the company include final results for Phase I/II trials for its cyclin-dependent kinase 7 inhibitor later in 2024. Promising Phase I results were also recently released for the company’s protein kinase C-theta inhibitor, which was in-licensed by Bristol Myers Squibb. The company is also planning 2025 clinical trials for a lysine-specific demethylase 1 inhibitor and a mucosa-associated lymphoid tissue lymphoma translocation protein 1 (MALT1) inhibitor.
Nowak highlights Exscientia’s success in chemically profiling compounds before they enter clinical trials. For example, the company was able to predict that a competitor’s MALT1 inhibitor compound would result in hyperbilirubinemia before the clinical trials readout. In contrast, the company’s MALT1 inhibitor has been designed without any markers for this adverse effect.
Schreiber believes that AI-driven drug design is on a new path, one where the milestones will be unprecedented drugs as opposed to incremental gains in efficiency and speed. “AI success is all about delivering a drug candidate that will show great data and be truly novel,” he declares. “In other words, a human expert will look at it and say, ‘Okay, I probably would not have thought about this by myself.’”
