A New Frontier in Drug Discovery

Sure enough, ‘Generative AI’ in the form of LLMs are great at organizing and presenting the worlds information. Whether neural nets, the technology underlying today’s generative AI, will also provide new, foundational knowledge outside the current lexicon remains to be seen. Subdomains such as mathematics, programming, physics and drug discovery seem suitable for achieving breakthroughs. These disciplines enable rapid testing and validation, opening the door for reinforcement learning, where the AI receives feedback in the form of rewards or penalties for its actions. Today, let's delve deeper into the promise of AI to transform drug development. At the same time, we’ll also look at several companies that are active within this domain.

Accelerating Healthcare Innovation with AI - Kimberly Powell, VP of Healthcare Nvidia

 

The current drug discovery process is lengthy, costly and risky; the average cycle lasts 8-10 years, with a success rate of 5-10%, and costs can easily exceed $ 1 billion. The journey of a new drug is a meticulous one, typically involving several key stages: identifying a target (~proteins), screening a library of compounds, identifying ‘hits’, optimizing them into ‘leads’, and finally selecting a promising candidate. Historically, this has been a laborious process, often relying on time consuming and expensive ‘wet lab’ experiments.

The first critical step in this process is identifying a druggable target—a protein in the body that can be manipulated to treat a disease. There are two primary approaches to this: forward pharmacology, which involves screening compounds to find those with a desired therapeutic effect, and reverse pharmacology, which starts with a known target and then finds compounds that can alter its function. While both methods have yielded important discoveries, they are often time-consuming and data intensive.

This, of course, is where neural networks excel, they analyze vast networks of biomedical data, revealing patterns that are invisible to the human eye. Companies like Relay Therapeutics look beyond the traditional ‘lock-and-key’ model of drug discovery. Instead of treating proteins (~’lock) as rigid structures, they view them as dynamic, constantly moving entities. Their approach leverages this understanding of protein motion to design drugs that target previously ‘undruggable’ sites.

Another example is Recursion Pharmaceuticals, a company partnering with major pharmaceutical firms like Bayer and Roche to create so called ‘phenomaps’ (i.e. massive, multidimensional maps of the biological system) that can uncover new targets for diseases like cancer.

Recursion Pharmaceuticals, Download Day 2024

Once a target is identified, the next step is to find molecules that can bind to it. This has traditionally been done through high-throughput screening, a costly and labor-intensive process that can take years. The advent of neural networks has given rise to virtual screening, where algorithms can test millions of compounds on a computer before a single one is synthesized (~ ‘wet lab’ experiments). This approach dramatically expands the pool of potential molecules while reducing both cost and time. Companies such as Schrödinger are combining machine learning with physics-based methods to accelerate drug development and produce high-quality candidates. Others, like Ideaya Biosciences, are leveraging advanced calculations to predict how a molecule will bind to a target, setting a new standard for rapid and accurate hit-finding.

Schrödinger company presentation - https://s203.q4cdn.com/609444515/files/doc_presentations/2025/Aug/06/Corporate-Overview-Deck-Aug-14-2025.pdf

Finally, a major cause of drug development delays and failures is toxicity. Predicting a compound's safety profile early in the process is critical to avoiding costly late-stage failures. AI is proving invaluable here, as it sometimes can predict potential safety issues, allowing companies to adopt a "fail early, fail fast" approach. Private companies like Generate:Biomedicines are partnering with major pharmaceutical firms to apply these new technologies to the creation of novel medicines.

Neural networks continue to receive significant attention and industry funding, despite the lack of major breakthroughs in drug development so far.

Recently, a glimpse of its potential was demonstrated with Rentosertib, widely considered the first experimental drug where both the biological target and the therapeutic compound were identified and designed using generative AI. Rentosertib, a drug to treat serious lung disease, has progressed through phase I and phase II clinical trials with promising results, and is currently undergoing larger pivotal trials for regulatory approval.

Drug Target review - https://www.drugtargetreview.com/news/157365/first-ai-designed-drug-rentosertib-named-by-usan/

The potential of AI has also not gone unnoticed by regulators. Dr. Marty Makary, the FDA commissioner, has recently noted the opportunity for companies to use AI. In the long term, this should also lead to new regulatory frameworks that are more accommodative.

Webstite FDA - https://www.fda.gov/about-fda/center-drug-evaluation-and-research-cder/artificial-intelligence-drug-development

For investors, these efficiencies could translate into billions in additional cash flow. A drug that gets to market even a year earlier could see a significant increase in its net present value. This offers a powerful incentive for companies to embrace these new technologies, ultimately allowing the ecosystem to flourish.