Pancreatic cancer remains one of the deadliest malignancies. Five-year survival rates hover around 12%, and most approved drugs show limited effectiveness when used alone. Finding effective drug combinations through trial and error would take decades.
Researchers at the National Center for Advancing Translational Sciences (NCATS), part of the National Institutes of Health, took a different approach. They trained machine learning models to predict which drug combinations would work synergistically against pancreatic cancer cells, then tested their top predictions in the lab.
The result: 307 validated synergistic drug combinations from a search space of nearly 1.6 million possible pairings. The research, published in Nature Communications, demonstrates that AI can dramatically accelerate the hunt for combination therapies in treatment-resistant cancers.
The Pipeline
The team started with a practical screen. They tested 1,785 FDA-approved cancer drugs and investigational compounds against PANC-1 cells, a standard pancreatic cancer cell line. This narrowed the field to 32 compounds showing meaningful anticancer activity.
Next came the combinatorial explosion problem. Testing all possible pairs of those 32 drugs would require 496 experiments. But the researchers wanted to explore a much larger space - what if they could predict synergies across compounds they hadn’t tested together?
Three independent teams from NCATS, University of North Carolina at Chapel Hill, and MIT each built machine learning models using the experimental combination data. The models learned patterns in how drugs interact based on their chemical structures and biological targets, then predicted synergy scores for nearly 1.6 million potential combinations.
When the researchers tested the top 88 predictions in the lab, roughly 60% showed synergistic effects - meaning the combination killed more cancer cells than either drug alone.
The Unexpected Findings
Two drugs stood out for their promiscuity. NSC-319726, a compound that reactivates mutant p53 tumor suppressor genes, showed synergies with multiple partners. So did AZD-8055, an mTOR inhibitor that targets cancer cell growth pathways. The two drugs also synergized with each other.
What made this interesting was the mechanism. Alexey Zakharov, who led the NCATS team, noted that “compounds are inactive by themselves, but mixed together, they show synergy.” The drugs target different pathways, and their combined effect exceeds what either achieves alone.
About 75% of pancreatic cancers carry p53 mutations. A drug that restores p53 function, paired with a compound blocking cancer cell survival pathways, represents a logical combination - but identifying which specific pairings work required experimental validation.
How the Models Performed
The research compared multiple machine learning architectures. Graph convolutional networks, which represent molecules as connected graphs of atoms, achieved the best hit rate. Random forest models showed the highest precision. The teams published their code on GitHub so other researchers can apply the same approach to different cancers.
The 60% confirmation rate for top predictions is notable. Random chance in drug combination screening typically yields synergy rates around 10-20%. The models tripled or quadrupled that baseline, focusing experimental resources on the most promising candidates.
The Path to Patients
Laboratory results in cell lines do not translate directly to clinical effectiveness. Cancer cells in a dish lack the immune system, blood supply, and complex tissue environment of real tumors. Many combinations that look promising in early testing fail when tested in animals or humans.
But the value here is in the screening approach itself. Pancreatic cancer has proven resistant to most treatments, and identifying effective combinations through exhaustive testing would take years. Machine learning provides a shortcut - not to clinical approval, but to identifying which combinations deserve further study.
The researchers suggest their pipeline could accelerate drug repurposing for other treatment-resistant cancers. The approach works with any cancer cell line where initial drug sensitivity data can be generated.
What This Means
The study demonstrates a practical workflow for AI-assisted drug combination discovery: screen drugs individually, train models on combination data, predict across a larger chemical space, and validate top predictions experimentally.
The 307 validated combinations represent a starting point, not an endpoint. Researchers now have a prioritized list of drug pairs to test in more sophisticated models - patient-derived cells, animal studies, and eventually clinical trials.
For a disease where single-drug approaches have largely failed, systematic identification of combinations that work together may offer a more promising path forward.