This is How AI Can Help Us Make New Medicines Faster

Whether it’s cancer, diabetes, or Alzheimer’s disease — badly behaved proteins are behind dozens of the most devastating diseases.

If our bodies are an orchestra, proteins are our maestros. As hormones, they cue processes like growth, metabolism, and reproduction. As enzymes, they decide the tempo of the chemical reactions needed for digestion or DNA replication. As antibodies, they lead our immune systems — and, in still other forms, they preside over the life and death of cells.

But when proteins misbehave in form or function, they cause disease.

Tanja Kortemme, PhD, is a UC San Francisco bioengineering professor and vice dean of research at UCSF’s School of Pharmacy. Her lab recently created the world’s first shape-shifting synthetic proteins. She tells us how UCSF scientists are leveraging decades of federal funding to create never-before-seen proteins using artificial intelligence (AI) — think ChatGPT but for proteins. One day, these proteins may put their misbehaving brethren in their place.

How can proteins be turned into medicines?

We’d want to use proteins as medicine if a particular protein isn’t functioning correctly in a person’s body. Let’s take insulin: Insulin is a hormone made of protein that regulates metabolism, particularly how our bodies metabolize sugar. But in people with type 1 diabetes, their bodies don’t produce enough insulin to metabolize sugar and control their blood sugar levels, so they’re prescribed insulin.

Other protein-based medicines include GLP-1 weight-loss drugs like Ozempic and Wegovy, and antibody treatments like Herceptin for cancer.

How can AI develop new medicines like these?

If you design proteins completely from scratch, you’re not limited anymore to the proteins that already exist. We can build proteins with completely new properties, and that could be incredibly powerful in solving the challenges we face in medicine.

The possibilities are almost limitless.

You’re using AI to make new proteins from scratch. How?

We have AI models that learn about protein structures, meaning where all the atoms in a protein are arranged and in three dimensions. Once the AI knows protein structures, we can ask it to generate a protein to block a different protein that causes a disease, like cancer, to spread.

Where does the data come from that you train the AI on?

You can train AI models on many things, but they’re only successful if you train them on large databases. For many decades, a worldwide community of researchers has worked to determine protein structures. Scientists deposited their findings into the open-access Protein Data Bank to advance this work together. This database is made possible by federal research funding from the National Institutes of Health (NIH) and the National Science Foundation, among others.

My lab is using large-scale data from the bank and other sources to produce novel proteins with new functions. We’re doing this both by using existing generative AI models and creating new AI models of our own.

How does a protein move from an AI model to real life?

This is made possible by DNA synthesis and recombinant DNA technology that allows scientists to compose DNA segments to produce proteins.

Recombinant DNA technology is one of UCSF’s major innovations.

Are there any AI-designed protein medicines already on the market?

Many companies are using AI methods to help them discover and optimize potential medicines. But a purely AI-designed protein from scratch? We don’t have that yet.

Still, there’s an explosion of biotechnology industry efforts to develop AI-generated proteins with therapeutic implications. I expect we’ll see a lot of those designed proteins enter preclinical development in the next five years and then hopefully, into the clinic to actually help people.

Is UCSF at the forefront of AI-designed proteins?

Yes, for more than 20 years, UCSF has had a lot of strength in pioneering protein engineering technologies thatmodify proteins in a more classical way of modifying existing proteins to turn them into better medicines. Today, groups at UCSF – particularly our incredibly innovative graduate students and postdoctoral scholars – have developed advanced computational methods, including AI, to design proteins from scratch.

What’s the secret to this kind of innovation?

Much of progress at UCSF is due to interdisciplinary science and collaborations, creating opportunities for innovation when different fields come together — and our postdocs and graduate students are a big part of that.

Many students who apply to UCSF are interested in artificial intelligence. Often, they have a background in computer science, engineering, or mathematics — all the disciplines that provide the right background to make AI advances. But these students are also fascinated by the biological and biomedical problems that drive our research at UCSF.

UCSF provides an amazing environment for integrating computer science and engineering and other disciplines, with fundamental biology, discovery and biomedical sciences, and pharmaceutical development.