How generative AI could help to cure ‘untreatable’ diseases

Enter the right prompt and a generative AI program can describe the symptoms of Parkinson’s disease pretty accurately. But can it help to find a cure for it?

That’s the expectation of a new generation of bioscientists who are using technologies similar to those behind ChatGPT and DALL-E 2 to create novel medicines that can repair the genetic causes of diseases. Alongside recent advancements in our understanding of genetics and protein biology, generative AI can allow us to “program” medicines for diseases that were once considered untreatable.

RNA editing is emerging as a promising field to treat or cure diseases with a genetic cause. RNA is a molecule present in all living cells that carries instructions from DNA for producing proteins that determine how cells behave and develop. If a person has a genetic mutation that makes them susceptible to a disease, scientists can edit their RNA to change that message and fix the underlying problem, without unintended permanent changes to the DNA. They can do this by designing and delivering what’s called a guide RNA molecule. Guide RNAs redirect a type of protein that naturally edits RNA (and is present in all human cells) to correct the genetic cause of the disease.  

The design of guide RNAs is a massive undertaking. RNA is composed of 4 letters of the genetic code: A, G, C, and U. A short strand of RNA that is 20 letters long has over a trillion possible sequences. The sheer magnitude of possible guide RNA sequences poses a significant challenge, as finding the precise sequence for each target would require exhaustive computations that would take years, even on the most advanced supercomputers. 

Enter generative AI

Similar to how large language models are trained on a vast corpus of data to interpret prompts and generate likely answers, bioscientists are training models using millions of experimentally measured data points to teach computer programs to generate candidates for RNA therapies. These likely candidates can then be tested in the lab, greatly accelerating the time to results. 

One of the joys of a program like DALL-E 2 is that it can produce unexpected, fantastical results—a fairy riding a unicorn or a rabbit at a dinner party. But scientists need results that can be applied practically in the real world, so we introduce constraints to our models to ensure that the AI generates only realistic RNA designs. We feed “images” of actual guide RNA sequences to “condition” the model along with information about how effective they were when tested. With this data, the model can generate brand new guide RNAs that it hasn’t seen before but that are plausible. 

Expanding possibilities of generative AI

In a similar vein, other groups are using generative AI to take on different challenges. Just last month, a research team at the University of Toronto published their findings on using generative models to design new proteins. Stability AI, which created the text-to-image AI system Stable Diffusion, is backing an effort called OpenBioML, which aims to develop an AI that can generate functional DNA sequences and chemicals from text prompts.

Further, pharmaceutical giants including Roche and Merck are pursuing new drugs using generative AI and natural-language processing approaches. Meanwhile, biotechnology companies, such as Recursion Pharmaceuticals, Generate Biosciences, and Insilico Medicine, are harnessing generative AI in creative ways to accelerate novel protein design and drug discovery.

We use diffusion-type models at my company, but we’re not yet concerned with text prompts. Instead of taking a corpus of text and associating it with an image, we take a corpus of experimental editing profiles and associate them with guide RNA “images” and further use those to generate new sequences. The model then can generate even better guide RNA solutions that fix mutations with exceptional accuracy.

Some of our current work focuses on a protein called amyloid precursor protein, or APP, that’s associated with Alzheimer’s disease. In certain individuals, APP degrades into a group of malformed proteins known as amyloid plaques, which are thought to play a central role in Alzheimer’s. Certain Icelandic populations have a very low incidence of Alzheimer’s due to a mutation in APP that protects it from degrading in this way. One of our goals is to develop a therapeutic guide RNA that can replicate this Icelandic variation to protect against Alzheimer’s in at-risk populations. 

Separately, we’re partnering with Roche and using RNA technology to discover potentially transformative treatment options for people living with Alzheimer’s, Parkinson’s, and other diseases. 

What’s next

In the long term, generative AI will help make drugs both more effective and more accessible. Today, many drugs and gene therapies are prohibitively expensive. Generative AI-enabled drug discovery and design, as well as innovations in manufacturing and drug delivery, will change that. With scientists working together with AI tools, major breakthroughs that will make a big difference for patients’ health are just around the corner. 

Ron Hause is a senior vice president and head of AI at Shape Therapeutics.