A new artificial intelligence tool developed by researchers at Northeastern University could dramatically shorten the time it takes to diagnose and start treatment for acute myeloid leukemia (AML), a rare and aggressive cancer, according to a patent awarded to bioengineering assistant research professor Kiran Vanaja.
AML, which affects both blood and bone marrow, currently requires oncologists to analyze samples from both sources to determine a patient’s specific genetic makeup and identify the most appropriate treatment. This process can often take a month or more, a critical delay given that the median survival rate after initial diagnosis is less than five years, Vanaja stated.
The AI platform, built at Northeastern’s Roux Institute in Portland, Maine, aims to accelerate this process by mapping genetic mutations within a patient’s AML cells. It then utilizes a neural network – a computational model inspired by the human brain – to suggest potential drugs and predict the likelihood of drug resistance. Vanaja believes the tool could reduce the preliminary diagnosis and treatment planning phase from weeks to a single night.
Understanding the complex interplay of genes within cancerous cells is a significant challenge. Vanaja likened genes to “LEGO blocks,” explaining that their various combinations and arrangements dictate cellular expression. Conventional methods like gene and RNA sequencing can reveal what’s inside a cell, but don’t always reflect how the cell behaves.
Vanaja’s team discovered that cancerous cells undergo a radical transformation when exposed to existing therapies, activating numerous survival mechanisms in a desperate attempt to resist treatment. This creates a disconnect between a cell’s genetic code (genotype) and its observable characteristics (phenotype).
“The cells undergo massive rewiring when subjected to these cancer therapeutics, as they’re literally trying to survive by turning on anything and everything possible,” Vanaja explained.
The neural network addresses this complexity by rapidly analyzing the vast number of possible gene combinations. With approximately 50,000 known genes, even considering only the roughly 20,000-30,000 genes in the human genome, the potential combinations are enormous. The neural network, with its interconnected processing layers, can efficiently sift through these possibilities.
The team trained the AI model using genotype and phenotype data from thousands of cells from approximately a dozen AML patients, supplementing this with data from existing scientific studies. This allowed the model to more accurately correlate genetic variations with cellular behavior.
While initially focused on AML, Vanaja emphasized that the tool’s core function – connecting genotype to phenotype – has broader applications. He indicated that future research will explore its potential use in diagnosing and treating solid tumors.
The next step, Vanaja stated, is to continue refining the model with additional patient samples and validate its predictions against real-world clinical outcomes.