Bayesian networks based on single nucleotide polymorphisms (SNPs), developed from saliva-sourced DNA, can be used to predict the occurrence of adverse effects associated with chemotherapy, according to investigators who presented studies at both the 2012 CTRC-AACR San Antonio Breast Cancer Symposium and the 54th Annual Meeting of the American Society of Hematology (ASH).1,2
Given the biological basis for many toxic adverse effects of chemotherapy, genotypically based risk assessments make sense, the investigators said.
“It is clear that genetics plays a significant role in determining individual patient risk for side effects, and it seems most likely that the genetic impact on risk is a result of teams of genes functioning synergistically,” said Lee Schwartzberg, MD, of the West Clinic in Memphis, Tennessee. “In these investigations, we determined if we could construct SNP networks using an innovative Bayesian methodology that would be able to identify patients being treated for cancer who were at risk for side effects.”
“The development of SNP-based Bayesian networks for risk assessment offers significant advantages over classic candidate gene and genome-wide association studies (GWAS),” added Ed Rubenstein, MD, president and chief executive officer of Inform Genomics, Inc., Boston, Massachusetts, which is developing the test. “Instead of looking for a single ‘master gene or SNPs’ or individually acting genes that ‘fit’ with phenotype (the GWAS approach) to predict side effects, the new concept recognizes a network of SNPs working cooperatively,” he said.
The investigators have found a range of around 20 to 100 SNPs per network, and each network is associated with a particular moderate-to-severe toxicity. Advances in bioinformatics and computational power have enabled the development of learned networks using a Bayesian methodology, which is a means of estimating probabilities.
Rubenstein explained the approach. “We start out with 2.5 million SNPs per person plus the clinical data, and all this goes into the network. The Bayesian algorithm removes the 99% of the SNPs that have no association with the side effect, ie, the ‘noise.’ We then take the remaining SNPs and the clinical data and build an interaction SNP network, and this is then cross-validated.”
Predicting Breast Cancer Risk
The study presented in San Antonio involved 78 patients with breast cancer who received at least 3 cycles of dose-dense doxorubicin/cyclophosphamide plus paclitaxel and the recommended prophylaxis and supportive care. A validated patient-reported symptom assessment tool was used to measure oral mucositis, chemotherapy-induced nausea and vomiting, the severity and frequency of diarrhea, fatigue, cognitive dysfunction, and peripheral neuropathy.
Predictive SNP networks were developed for each of the 6 adverse effects. The primary end point was an area under the receiver operating characteristic curve (AUROC) >0.80.
The investigators were able to identify SNP networks associated with moderate-to-severe adverse effects of the standard breast cancer regimen. The occurrence of toxicity and the model’s accuracy at predicting it are shown in the Table.
Schwartzberg explained that the AUROC shows tight correlation between the SNP networks and the occurrence of these adverse effects. For example, the presence of the SNP network associated with oral mucositis can predict with 97% accuracy that an individual is likely to be among the 49% who will develop this adverse effect, he explained.
He said the test could possibly be very useful in the clinic. “We have alternative regimens for many kinds of cancer, especially breast cancer, and they differ substantially in toxicity, if not necessarily in efficacy. In an attempt to personalize treatment, this is a potentially exciting tool for selecting patients for particular regimens that will be less toxic for them,” he said.
Schwartzberg added that the model reveals intrinsic predispositions based on different germline SNPs. “It makes biological sense, and that is key,” he said.
Predicting the Risk for Oral Mucositis
The oral mucositis study presented at ASH comprised 153 patients in the discovery set, including 82 patients with myeloma and 71 with Hodgkin disease or non-Hodgkin lymphoma undergoing conditioning regimens.
The genetic analysis revealed 82 SNPs within a network, and these identified patients developing mucositis with an accuracy of 99.3%, reported Stephen Sonis, DMSc, professor of Oral Medicine and Diagnostic Science at Harvard School of Dental Medicine, Boston, Massachusetts. Sonis has pioneered much of the research on oral mucositis and its prevention and treatment.
In a prospective validation study of 16 patients who were demographically similar to the discovery cohort, the network predicted mucositis with an accuracy of 81%. Of 8 patients without mucositis, all were accurately identified and there were no false-positives. Of 8 patients who did not develop mucositis, 5 were accurately identified and there were 3 false-negatives.
The SNPs network is “more robust” and “more likely to be accurate in subsequent validation studies.” There is less chance for false-positives compared with the “old paradigm,” Sonis said.
According to Rubenstein, having a predictive test would streamline the selection of patients who could benefit from expensive prophylaxis such as palifermin. “You don’t want to give this expensive drug to patients who will never need it.” The commercial custom chip could become available for routine patient care if results from these studies can be validated in a multicenter study, he added.
References
1. Schwartzberg LS, Sonis ST, Walker MS, et al. Single nucleotide polymorphism (SNP) Bayesian networks (BNs) predict risk of chemotherapy-induced side effects in patients with breast cancer receiving dose dense (DD) doxorubicin/cyclophosphamide plus paclitaxel (AC+T). Presented at: 2012 CTRC-AACR San Antonio Breast Cancer Symposium; December 4-8, 2012; San Antonio, TX. Poster P1-15-12.
2. Sonis ST, Antin JH, Alterovitz G. SNP-based Bayesian networks define oral mucositis risk in patients receiving stomatotoxic conditioning regimens for autologous hematopoietic stem cell transplantation. Presented at: 54th Annual Meeting of the American Society of Hematology; December 8-11, 2012; Atlanta, GA. Abstract 735.