Northeastern scientists from the Neural Dynamics Group develop an algorithm that allows machines to link information across domains and extract more nuanced answers.
Ever notice how chatbots often insist on making up an answer instead of admitting defeat? Like a student desperately scribbling nonsense to avoid handing in a blank page, large language models (LLMs) seem to live by the make-it-up-as-you-go principle.
Many also stumble when bridging two areas of knowledge that don’t have an obvious link. They tend to give a superficial response instead of following the trail.
Sure, it’s annoying when a search snafu derails your goal of unlocking the flying car in Grand Theft Auto VI. But when the fumbled response involves synthesizing medical research findings, the stakes get a lot higher.
Given that PubMed lists over 40 million citations and abstracts, finding the right information in the vast database is a job that begs to be outsourced to AI. But more often than not, a query spanning more than one knowledge realm will result in answers that are outdated, incomplete or plain wrong.
“Since knowledge is so very fragmented, data is all fragmented,” said leader of the Neural Dynamics Lab Ayan Paul in an interview with Northeastern Global News. Paul is a research associate professor at the Institute for Experiential AI.
Headed by the lab’s machine learning engineer Nihar Sanda, a group of researchers found a way to teach AI to dig deeper.
In a paper that appeared in Bioinformatics in July, the team presented an algorithm called eGoT. Standing for “enhanced graph of thoughts,” it teaches LLMs to answer biomedical questions by pulling together evidence from different domains. An algorithm like eGoT works like a set of instructions for the computer system. It tells the machine how to perform a task — in this case, information retrieval.
The process that eGoT sets in motion starts with a user query. For example, the researcher might ask how a mutation contributes to a disease. They might also inquire why a certain set of environmental conditions makes symptoms worse.
Next, the algorithm tells the system to scan available resources for relevant information. It connects the dots in a way that makes sense, Paul explained.
The heart of eGoT’s approach is in the so-called graph-of-thoughts — a map of linked concepts relevant to a particular question. The concepts form so-called nodes. Promising ones can be expanded, sprouting new branches. Separate branches can also merge into new nodes, or ideas, that synthesize information in a useful way.
During the knowledge retrieval stage, the algorithm tells AI to mine the web of connected data and extract an answer that’s both accurate and comprehensive.
Editor’s Picks #
The trick is to grab the right mix of facts from the get-go, Paul explained. He used an analogy to illustrate the process: imagine you have a bag filled with multicolored balls, he said. You want to pull out at least five colors. Reach in and grab just one ball, and you’ve already lost. To give yourself a fighting chance, you need to grab a big handful — maybe even all of the balls.
The same is true for information systems. An LLM that doesn’t pull in enough useful data can’t give a solid, well-rounded answer. With eGoT, however, it becomes “really good at grabbing,” Paul said.
Next came the test: would this actually work?
Sanda, Paul and their team zeroed in on two areas of knowledge that tend to live in separate research silos — lupus and ultraviolet (UV) radiation. The first is an autoimmune disease that flares up when the body mistakenly attacks its own cells. The second — a form of energy emitted by the sun that can damage skin, and the reason why tanning beds have a risky reputation and have started to fall out of favor.
As Paul explained, “everyone knows that people with lupus are affected by UV,” which tends to make outbreaks worse. When UV radiation amps up, so does the autoimmune response, leading to rashes, fatigue and joint pain.
Rheumatologist, chair of the health nonprofit Lupus Foundation of America and author of “The Lupus Encyclopedia” Donald Thomas talked to Northeastern Global News about the connection. He said that lupus patients are prone to sun-sensitive rashes, even if some are more sensitive to them than others. Thomas advises his patients to “wear UV protection religiously,” switch to LED light indoors and make sure their clothing provides ample coverage. His Lupus Encyclopedia blog explains the chemistry behind the link. It shows that UV light interacts with proteins on the skin surface.
However, while most agree about the basic connection, Paul said that experts in the fields of lupus and UV radiation will still look at the respective subjects through different lenses. For example, the lupus crew won’t necessarily focus on “how UV is going to change with weather patterns” — a question that’s more likely to interest UV experts, he explained.
Bring the two bodies of research together, however, and less obvious links start to emerge. Sanda said with eGoT you can find new patterns about the way UV radiation spikes affect lupus patients — ones that may have otherwise been unexplored. The AI system can follow these threads even when they wander surprisingly far from the original question.
Take greenhouse gas emissions, for instance. Because they can increase UV radiation, they have an indirect relationship to lupus outcomes, Sanda said. A traditional search retrieval model might lose this thread along the way, but eGoT is designed to keep following it.
Lupus wasn’t the only proving ground for eGoT, which passed additional tests.
Small cell lung cancer — another devastating disease — also bridges knowledge and strategies from different subfields. With a less than 5% survival rate, the stakes are high.
When a co-author from the medical field, Northeastern systems biology professor and strategic adviser to the Provost Vito Quaranta, gave an eGoT-armed system a set of questions about small cell lung cancer to tackle, the outcome was promising.
“We were able to get some really cool results,” Sanda said. For example, it was able to draw some specific connections between different transcription factors — proteins that control gene expression — and specific clinical features of the disease.