Ken Drazan: Building an LLM that speaks the language of the immune system
How ArsenalBio is targeting cancers and accelerating the speed of biomedical research
The platform that ArsenalBio has built to develop these therapies has been described as the building blocks for a large language model (LLM) that speaks the language of immunology. Can you explain what that means and how it works?
The immune system has developed over 500,000 years. It is very effective at some things, such as dealing with infection, but it’s not so good at tackling challenges that often come when humans get older: cancer, autoimmune diseases, and allergies. These are all complications of imperfect adaptation to our environment. We observed that as the human immune system evolves, its programming language is actually evolving, too. So, we asked how we might accelerate that programming language and apply it to a specific problem, such as recognizing and eliminating cancer. We thought that by inserting huge libraries of different types of code mimicking what would happen over the course of evolution, we could build an LLM that would essentially accelerate the activities of evolution. By capturing that data, in the same way that AI captures all of the spoken English language and turns it into a tool to compose an email, we could potentially use that immunology data set to interpret what the immune system needs to do, or does, when a disease is complicating someone's life. Given a sample from a patient, the system could potentially tell us which therapies would be powerful enough to reverse or redirect that process.
In addition to applying this to patients, our idea is to make this platform widely available for others studying human disease. That democratization of scientific research using an LLM, in this case, focusing on diseases where T cells play a significant role — which includes COVID, solid organ cancers, allergies, and many other diseases — could be really powerful.
So you originally set out to develop therapies for cancer, but in the process you leveraged your data and AI to build this platform. And now you seem on the cusp of something that could have much broader applicability. What are the lessons in that?
First of all, this opportunity started in a company that had very big ambitions. We started with more of an engineering approach than a classic wet lab research approach. And, as our capabilities became operational, they started to generate large data sets. At the time, AI in the life sciences was nascent, and frankly, not necessarily accessible. Fast forward several years, and our datasets got bigger and better, and we became more sophisticated at handling that data.
In parallel, the cost of GPU started to come down, and hyperscalers became increasingly accessible for small customers. That’s what began the platform shift. Instead of us trying to do things by hand, we could do more in the cloud. And things ballooned from there, as we were able to generate tons more data. And when you start to think big, new business partners, investors, and thought leaders show up because they want to enable the platform shift of the future. So, we basically created a startup within a startup. By having a culture that allows inquiry, as soon as our platform started to answer questions that seemed impossible to answer in our laboratories, we started to ideate how this could become a platform for many.
.png)
.png)
