University of California San Diego School of Medicine and Moores Cancer Center say they have created a new artificial intelligence (AI) system called DrugCell that reportedly matches tumors to the best drug combinations, but does so in way that clearly makes sense.
“That’s because right now we can’t match the right combination of drugs to the right patients in a smart way,” said Trey Ideker, PhD, professor at University of California San Diego School of Medicine and Moores Cancer Center. “And especially for cancer, where we can’t always predict which drugs will work best given the unique, complex inner workings of a person’s tumor cells.”
Currently, Only four percent of all cancer therapeutic drugs under development earn final approval by the FDA.
In a paper “Predicting Drug Response and Synergy Using a Deep Learning Model of Human Cancer Cells” published in Cancer Cell, Ideker, Brent Kuenzi, PhD, and Jisoo Park, PhD, postdoctoral researchers in his lab, published a paper on their work.
“Most drugs entering clinical trials fail, often related to an incomplete understanding of the mechanisms governing drug response. Machine learning techniques hold immense promise for better drug response predictions, but most have not reached clinical practice due to their lack of interpretability and their focus on monotherapies,” write the investigators.
“We address these challenges by developing DrugCell, an interpretable deep learning model of human cancer cells trained on the responses of 1,235 tumor cell lines to 684 drugs. Tumor genotypes induce states in cellular subsystems that are integrated with drug structure to predict response to therapy and, simultaneously, learn biological mechanisms underlying the drug response. DrugCell predictions are accurate in cell lines and also stratify clinical outcomes.”
“Analysis of DrugCell mechanisms leads directly to the design of synergistic drug combinations, which we validate systematically by combinatorial CRISPR, drug-drug screening in vitro, and patient-derived xenografts. DrugCell provides a blueprint for constructing interpretable models for predictive medicine.”
“Most AI systems are black boxes; they can be very predictive, but we don’t actually know all that much about how they work,” said Ideker, who is also co-director of the Cancer Cell Map Initiative and the National Resource for Network Biology.
He gave the example of the way an internet image search for “cat” works. AI systems working behind the scenes are trained on existing cat images, but how they actually label a new image as “cat” and not “rat” or something else is unknown.
For AI to be useful in health care, Ideker said, we have to be able to see inside the black box to understand how the system comes to its conclusions. “We need to know why that decision is made, what pathways those recommended drugs are targeting and the reasons for a positive drug response or for its rejection.”
The team’s work on DrugCell began several years ago in yeast. In a previous study , they built an AI system called DCell using information about a yeast cell’s genes and mutations. DCell predicted cellular behaviors, such as growth, all outside the black box.
DrugCell, a next-generation version of DCell, was trained on more than 1,200 tumor cell lines and their responses to nearly 700 FDA-approved and experimental therapeutic drugs—a total of more than 500,000 cell line/drug pairings. The researchers also validated some of DrugCell’s conclusions in laboratory experiments.
With DrugCell, the team can input data about a tumor and the system returns the best-known drug, the biological pathways that control response to that drug, and combinations of drugs to best treat the malignancy.
Precision cancer therapy is already available at Moores Cancer Center at UC San Diego Health, where patients may have a biopsy of their tumor sequenced for mutations and assessed by the Molecular Tumor Board, an interdisciplinary group of experts. The board recommends personalized therapies based on the patient’s unique genomic alterations and other information. A recent study showed these patients have better outcomes. In a way, DrugCell simulates the human Molecular Tumor Board.
“We were surprised by how well DrugCell was able to translate from laboratory cell lines, which is what we trained the model on, to tumors in mice and patients, as well as clinical trial data,” Kuenzi said.
The team’s ultimate goal is to get DrugCell into clinics for the benefit of patients, but the study authors caution there’s still a lot of work to do.
“While 1,200 cell lines is a good start, it’s of course not representative of the full heterogeneity of cancer,” Park said. “Our team is now adding more single-cell data and trying different drug structures. We also hope to partner with existing clinical studies to embed DrugCell as a diagnostic tool, testing it prospectively in the real world.”