React-to-Me brings a chatbot interface to Reactome, an expert-curated, open-access database of biological processes
Helia Mohammadi was just beginning her master’s studies at the University of Toronto in September 2023 when use of AI chatbots such as ChatGPT was becoming mainstream. Intrigued by the potential of this technology, Mohammadi and her colleagues set out to create what would eventually become React-to-Me, an AI chatbot that makes it easier to explore the massive amount of data in the Reactome database, which is curated by human experts.
Mohammadi spoke to OICR News about the project, which she completed under the supervision of Dr. Lincoln Stein at OICR and recently detailed in the journal BMC Biology.
For those who aren’t familiar, what is Reactome?
Reactome is a free, open-access database that helps researchers understand how biological processes work at the molecular level. It organizes information about genes, proteins and their interactions into pathways, which are sequences of events that take place inside our cells. The data is carefully curated and reviewed by experts, making Reactome a trusted resource for studying everything from normal cellular functions to disease mechanisms, including cancer.
Where did the idea for an AI interface for Reactome come from?
Given its size, finding what you want and navigating Reactome isn’t always straightforward, especially for someone who isn’t already familiar with the database, but it offers the advantage of high quality, curated data. Tools like ChatGPT offered almost the opposite experience: you could ask them just about anything in everyday language, but you couldn’t always trust the answers.
My supervisor, Dr. Lincoln Stein, recognized an opportunity to combine the strengths of both approaches. He proposed developing a chatbot that would allow people to explore Reactome by asking questions in everyday language, without needing to know exactly what to search for, while maintaining the quality and reliability of Reactome’s expert-curated information.
So that became the starting point: could we make exploring Reactome as simple as having a conversation, while ensuring that the answers remained grounded in reliable biological knowledge?
Who did you envision using React-to-Me?
We thought of React-to-Me as a tool for anyone who wants to explore biological pathways, whether they’re experienced researchers, students or even scientists working outside of their usual area of expertise.
Ultimately, the goal was to make Reactome easier to explore, without requiring users to know the terminology or structure of the database. We also wanted to make it easier for people to ask follow-up questions and explore related concepts as they go.
Can you tell us about the process of creating React-to-Me?
There were a lot of interesting aspects to the project but one of the biggest challenges was figuring out how to make the biological knowledge stored in Reactome accessible to a language model. Unlike a textbook or scientific article, where information is primarily presented in sentences and paragraphs, Reactome organizes much of its knowledge as a complex network of interconnected molecules, reactions and pathways.
A major part of my work involved figuring out how to retrieve and organize that information in a way that preserved the biological context and relationships between different pieces of data. We didn’t want the model to simply pick up on relevant gene names or biological terms; we wanted it to be able to explain how those pieces fit together to describe a biological process.
What about React-to-Me are you particularly proud of?
One thing I really like about React-to-Me is that we designed it with scientific reliability in mind from the beginning. With general-purpose AI chatbots, it can sometimes be difficult to tell where an answer came from or whether the information is accurate. We wanted React-to-Me to be different in that respect.
Another important feature is how the system handles questions it can’t adequately answer using Reactome alone. Rather than trying to fill in the gaps with potentially unreliable information, it’s designed to point users toward other trusted biomedical resources.
I think that balance between making information easy to access and keeping it scientifically grounded is what makes React-to-Me particularly interesting and useful.
OICR researchers Gregory Hogue, Adam Wright, Marija Orlic-Milacic, Nancy Li, and Lincoln Stein also contributed to the development of React-to-Me as did Fatemeh Almodaresi (Natera Inc.) and Amin Mawani (York University).
