RNA Foundation Models to Improve Pharma R&D Productivity

It’s no surprise that pharma research and development (R&D) remains one of the most important hubs for innovation. Every pharma company invests heavily in and relies on R&D teams to create new therapeutics to help the company grow and profit. From 2012 to 2022, inflation-adjusted industry R&D spending increased 44 percent, from about $170 billion to $247 billion. Despite this investment, the number of US novel drug approvals remained flat at an average of 43 per year. Productivity challenges have come to the forefront as pharma companies seek to leverage every available resource.

Changing role of AI 

As an industry, we can no longer tolerate significant productivity issues. Ten years ago, AI was promised as a solution to R&D productivity, but so far this promise has fallen short. We have learned the reasons for this shortfall, but more importantly, there is a new frontier of AI that addresses these reasons and offers hope. The next frontier of AI is called ‘foundation models’ and they are on the cusp of impacting R&D. AI foundation models are akin to ‘ChatGPT for drug discovery.’ Over the past two years, foundation models pertinent to pharma R&D have emerged and are on a trajectory to transform the pharma industry.

The first ten years of AI in biotech revolved around deep learning AI models that were trained to solve specific tasks using task-specific data. For example, a model that predicts the effects of gene editing on splicing, one of dozens of biological processes that are crucial for cell function. Many academic labs and companies, including ours, showed that they could use these systems to make accurate predictions to support specific tasks in drug discovery. Unfortunately, this approach has generally fallen short because genomics, biology, physiology, therapeutics, and chemistry are complex, multifaceted, and interrelated, so accurately accounting for only one task while leaving the many other tasks unaddressed results in failure modes during R&D. Even if many different AI models are built for the different tasks, they aren’t trained to take into account the interrelatedness of the tasks and the complexity of having many models makes scaling difficult or impossible.

Foundation models are a new approach that changes all of that. They are extremely large, deep-learning AI models trained on massive datasets that cover a broad field, and can be applied to a wide range of tasks. They achieve synergies between these tasks during training and when deployed in R&D applications. They exhibit ‘emergent intelligence,’ meaning they discover and apply relationships and learnings that were neither specifically included in the training data nor, in fact, conceived of by the creators of the AI. It is common for an AI foundation model to surprise its creators by making predictions or revealing patterns outside the intended scope of application. An example of an AI foundation model outside of drug discovery is ChatGPT. AI foundation models work because they are trained on very large and broad datasets, with billions or even trillions of data points, making them versatile.

Decoding RNA biology to drive R&D productivity improvements 

As RNA biology and RNA therapeutics emerge in the mainstream of pharma, RNA foundation models offer massive potential for revolutionizing R&D.

RNA foundation models can assist R&D by providing predictions for the effects of patient mutations, protein changes, and therapeutic interventions on a wide range of tissue-specific RNA biology processes, including transcription, splicing, polyadenylation, protein-RNA interactions and microRNA-RNA interactions. Therapeutic interventions may include oligonucleotides, DNA editing, RNA editing, mRNA and gene therapies. The approach can be applied to a wide range of disease and tissue areas, ranging from common disease, to rare Mendelian disorders, to so-called ‘N=1’ therapeutics designed for individual patients.

Notably, technology is not the only factor. One of the most critical elements for productivity improvements is instilling a culture of multilingualism, especially when it comes to experimental biology and AI. Future success hinges on breaking down departmental silos and replacing them with collaboration between computational scientists and biologists. These two diverse, interdependent teams must understand each other to ensure open, effective communication. Computational experts must consider how every step can help find more molecules and targets, and biologists need to be conscious of scale and data collection. 

The business case for foundation models 

With this revolutionary TechBio technology now being accessible, pharma companies can leverage foundation models to accelerate innovation. Patents will soon run out for many top-tier drugs, putting added pressure on the TechBio industry. For example, Nova Nordisk’s patent on semaglutide, the active ingredient in both Wegovy and Ozempic, expires in China in 2026.

According to McKinsey, with looming patent cliffs on blockbuster drugs, a constrained pricing environment, and the prospect of pricing reform (especially in, but not limited to, the United States), many companies are prioritizing R&D productivity improvement and pushing to get more out of every dollar invested in R&D. 

Fortunately for all pharma companies, modern advancements in AI and biology demonstrate the potential for accelerating drug discovery, optimizing clinical trials, and predicting possible adverse effects of drugs. New AI foundation models, such as those designed to decode RNA, are poised to improve R&D productivity, upend drug discovery methods, and ensure a better future for pharma.

Photo: Christoph Burgstedt, Getty Images

This post appears through the MedCity Influencers program. Anyone can publish their perspective on business and innovation in healthcare on MedCity News through MedCity Influencers. Click here to find out how.