AI@HHMI: Accelerating the development of new biological sensors

Key takeaways:

• A team led by Alison Tebo and Srinivas Turaga, group leaders at HHMI’s Janelia Research Campus, is creating new AI tools to speed up the development of new biosensors used to track key cellular processes.
• The team will use information from existing and new protein biosensors to train AI models capable of predicting and prioritizing molecular changes most likely to lead to promising new biosensors.
• The researchers aim to use these models to create a new AI-powered tool that can be used by the entire scientific community to generate new sensor candidates for novel biological needs, potentially helping to accelerate scientific discovery. 
• The project is part of AI@HHMI, the institute’s $500 million initiative to support AI-driven projects and embed AI systems throughout the scientific process.

Bespoke manufacturing might be great for suits and watches, but it’s not ideal for widely used biological tools.

Over the past three decades, protein biosensors that illuminate the processes happening inside cells have become a staple of biology labs worldwide. These sensors enable scientists to detect chemical signals linked to important biological processes as they happen, allowing them to understand how the brain and body work like never before.  

But designing and building each new state-of-the-art probe is still a custom job. Each sensor is engineered from a protein that has been modified to detect a specific molecule. When the sensor senses or binds its target molecule, it changes shape. This change causes a domino effect, leading a reporter molecule attached to the protein to change its fluorescence, enabling scientists to detect the signal change under a microscope.

Because each type of sensor accomplishes this process differently, tool developers must start from scratch when developing a new one. They rely on testing thousands of new protein variants – and a bit of serendipity – to develop and improve each new type of biosensor, a process that can take years.

“What we learn from making one biosensor is transferable in only a limited capacity to others, at least for humans,” says Alison Tebo, a group leader at HHMI’s Janelia Research Campus. “We sort of have to re-hack the system every single time.”

But, say Tebo and Janelia Group Leader Srinivas Turaga, it doesn’t have to be that way.

Using AI to optimize biosensor development

Tebo and Turaga are leading a new project to apply AI to the design of new biosensors for measuring metabolites – the molecules that are assembled and broken down during chemical processes in the body. The team is developing new AI models that use information from new and existing biosensors to predict specific changes that affect sensor function – a task that is difficult for humans but ready-made for artificial intelligence.

The project is part of AI@HHMI, the institute’s $500 million initiative to support AI-driven projects and embed AI systems throughout the scientific process.

By enabling novel probes to be built and modified faster and more efficiently, the team hopes to lower the barrier to developing new sensors and improving existing ones. This will facilitate the development of sensors to meet the ever-evolving needs of biologists, potentially resulting in new scientific discoveries.

“It is a high effort proposition to make a new biosensor, but it really shouldn’t be, and that is what the project started with: We should make this easier to do,[GS1] ” Tebo says. “It shouldn’t take you years to make a good biosensor; it should take months.”

Applying AI to large biological datasets

The new project takes advantage of the vast amounts of data generated from the development of new biosensors at Janelia over the past 20 years.

While it is nearly impossible for humans to synthesize and learn from these humungous data sets, it is a job well-suited to AI. The team plans to train their AI models on this data, along with new protein engineering data generated by the research team.

The models will use this information to generalize across sensors, enabling them to make predictions about how to develop biosensors with new properties, like being able to detect a particular molecule or fluoresce in a particular color. AI will also help the team prioritize which changes to test, optimizing the overall development process.

The team plans to build upon other AI tools for protein prediction, like AlphaFold, which have already been trained on huge amounts of data. 

“The idea is that using the foundation models plus machine learning on top of that, we can hope to learn more from the sensors that have been designed to be able to design new ones from scratch more quickly,” Turaga says.

A new AI tool to share with the scientific community

The researchers envision creating an AI tool that can be used by the entire scientific community to generate their own sensor candidates that address their specific biological questions. 

“If a researcher has a new molecule they’re interested in sensing, the dream would be that they would then run it through our model and out will come a potential sensor and they can then try it out,” Turaga says. 

Tebo says the AI@HHMI Initiative has allowed the team to expand their vision beyond their initial idea for the project, which was originally hatched as a collaboration between the two Janelia labs.

“This idea of having something really generalizable that allows us to be flexible in the types of sensors that we design, that was probably something that would not have come about, at least not as quickly, if it weren’t for the weight of the AI@HHMI Initiative,” Tebo says.

“It’s adding a lot of resources and momentum to something that I was already really excited about and spending a lot of time and effort on. Having the extra support from the initiative is really exciting and I think it’ll allow us to get further in a shorter period of time.”