PARP inhibitors — substances that block certain cellular enzymes — hold promise for cancers caused by defects in homologous recombination (HR), the microscopic machinery that orchestrates repairs of harmful DNA breaks. But they’re underprescribed, because most clinical tests don’t reliably detect HR.
Encouragingly, though, scientists at Harvard Medical School have developed an AI screening system — SigMA — which they claim can successfully “read” the molecular signature of HR deficiencies highly accurately and efficiently, and which furthermore works with existing screening methods. It’s described in a report published today in the journal Nature Genetics.
“Pinpointing actionable genetic biomarkers and treating patients with drugs that specifically target the relevant cancer-driving pathways is at the heart of precision medicine. We believe our algorithm can greatly enhance physicians’ ability to deliver such individualized therapy,” said study senior author Peter Park, a professor of biomedical informatics in the Blavatnik Institute at HMS, in a statement. “We suspect there are many more patients without BRCA mutations who could benefit from PARP inhibitors, but doctors do not know which ones they are. Our approach could help close that gap.”
As Park and colleagues explain, PARP inhibitors are commonly given to patients with breast, ovarian, pancreatic, and other cancers who have mutations in their BRCA genes. But not every patient with an HR deficiency has a BRCA mutation, so most standard assays miss them. By comparison, SigMA can identify patterns characteristic of HR defects — patterns that emerge in DNA components scrambled by cancerous malformations — even in clinical tests that analyze only a subset of genes.
The researchers culled from thousands of fully sequenced tumor genomes to compile a corpus and train the model, after which they measured its performance against 730 samples analyzed by whole-genome sequencing. They report that it correctly identified samples 74% of the time — an improvement compared with current algorithms, which detect HR-deficient cancer cells at a rate of 30% to 40% — and that in subsequent experiments involving 878 breast tumor samples from patients who had previously undergone genetic testing, it detected 23% of the samples bearing signs of HR deficiency. Moreover, it successfully sussed out previously unidentified defects in other types of cancers, ranging from 5% in esophageal cancers to 38% of samples in ovarian cancers.
In a third experiment designed to determine whether the model could predict cancer cells’ response to PARP inhibitors, the scientists sourced results from tests on 383 tumor cell lines across 14 cancer types treated with four different PARP inhibitors. They said that breast cancer cell lines (and even other tumor types) identified by SigMA as having an HR defect responded better to the PARP inhibitors than cells that didn’t have it.
“Tens of thousands of patients with cancer are profiled with gene panels across many hospitals and we believe our algorithm can detect the molecular footprints of the underlying cancer-causing defects with much greater sensitivity,” said study first author Doga Gulhan, a post-doctoral researcher in the department of biomedical informatics at HMS, in a statement. “The overarching goal of such testing is to help clinicians determine the optimal treatment for each patient based on the absence or presence of a given gene defect.”
The researchers believe that if SigMA were to be incorporated into genetic tests already used in hospitals, it could benefit the roughly 270,000 people diagnosed with breast cancer each year, an estimated 5% to 10% of which have BRCA defects. (In one simulated analysis, the model identified twice as many cases of breast cancer without BRCA mutations but driven by HR defects.)
The team cautions that the SigMA can’t detect HR deficiencies in cancers with few mutations, such as medulloblastoma (a type of brain cancer) and Ewing sarcoma (bone cancer). But they say it could be trained on other fully sequenced genome corpora to detect a greater variety of mutations.
“We have spoken with many clinicians in the past months and we have started multiple collaborations in which additional patients in clinical trials will be given the drug based on our predictions. We think we could make a real impact in cancer care with this computational method,” Park said. “The accuracy of the algorithm will vary by cancer type. But even when the detection rate is not as high, there still will be additional cases identified that would be otherwise missed. What this ultimately means is better-targeted treatments for more people.”