Prediction of opioid dose in cancer pain patients using genetic profiling: not yet an option with support vector machine learning

Objective Use of opioids for pain management has increased over the past decade; however, inadequate analgesic response is common. Genetic variability may be related to opioid efficacy, but due to the many possible combinations and variables, statistical computations may be difficult. This study investigated whether data processing with support vector machine learning could predict required opioid dose in cancer pain patients, using genetic profiling. Eighteen single nucleotide polymorphisms (SNPs) within the µ and δ opioid receptor genes and the catechol-O-methyltransferase gene were selected for analysis. Results Data from 1237 cancer pain patients were included in the analysis. Support vector machine learning did not find any associations between the assessed SNPs and opioid dose in cancer pain patients, and hence, did not provide additional information regarding prediction of required opioid dose using genetic profiling.


Introduction
Opioids are the basis in treatment of severe pain of both benign and malignant origin. Unfortunately, the clinical use is limited by large inter-individual differences in analgesic responses, and insufficient treatment is often seen. Unacceptable side effects may also appear, potentially reducing quality of life. Thus, identification of biomarkers that can predict the appropriate opioid type and dose for the individual patient is highly warranted. Currently, there is no well-validated objective means of identifying patients likely to experience adequate opioid analgesia, and quantitative sensory testing as well as clinical biomarkers have been applied with low success [1][2][3].
Various factors such as gender, age, and genetic variation may influence the analgesic response to opioids. Several single nucleotide polymorphisms (SNPs) in different candidate genes have been demonstrated to be associated with altered pain sensitivity and analgesic response [4]. The genetic variations can influence the pharmacokinetics and/or pharmacodynamics of opioids and potentially the effect. A large study, the European Pharmacogenetic Opioid Study (EPOS) included 2294 cancer pain patients and investigated the influence of genetic variability on multiple pain-related outcomes and required opioid dose [5,6]. No significant associations were found between 112 SNPs in 25 candidate genes and opioid dose, thus the findings did not support the use of genetics profiling to guide opioid treatment. However, statistical analysis including multiple testing of several factors may be a limitation. In contrast, machine learning can include many factors in a single analysis, limiting the risk of erroneous false-positive results [7]. Support vector machine (SVM) is a data driven method, which enables detection of subtle patterns in complex datasets, which are only visible when assessing multiple variables at once. This could be the case for genetic data, where certain combinations of genes could determine the effect. The objective of the present study was to use SVM analysis of various SNPs to predict the required opioid dose in cancer pain patients.

Study design and patient samples
Data from the EPOS study was used for analysis [6]. In brief, the study included patients from 17 centres in 11 European countries. Inclusion criteria were: age > 18 years; diagnosed with a malignant disease; using opioids for moderate to severe pain; treated with regular oral, subcutaneous, transdermal or intravenous opioids (morphine, methadone, fentanyl, hydromorphone, buprenorphine, or oxycodone) for a minimum of 3 days.

Study outcome
The median oral morphine equivalent dose in the full population of 2294 patients was 180 mg/24 h. Therefore, patients were divided into two groups: Group 1 requiring less or equal to 180 mg/24 h; group 2 requiring more than 180 mg/24 h.

Gene selection and genotyping
In the present study, nineteen SNPs; nine SNPs in the OPRM1 gene, one in OPRK1, three in OPRD1 and six in COMT, were selected as relevant. The SNPs were selected according to previous studies, in which genetic associations to opioid analgesic effects were found [4,8,9]. The genotype distributions are listed in Table 1. As minor allele frequency had to be higher than 10% to be included in data analysis, the OPRK1 was discarded. SNPs were coded to be binary variables for inclusion in the SVM, hence, they were coded so that carriage of the minor allele equals 1 and homozygous for the major allele equals 0 in the model.

Machine learning analysis
SVM is a binary classifier, previously used in other prediction studies in pain medicine [10,11]. In the present study, classification was performed using the libSVM toolbox (version 3.20) for Matlab [11], and a linear kernel function was used to avoid over-fitting of the model [12]. The analysis process is described in details elsewhere [13]. In brief, 10 features along with a label indicating to which opioid dose group the patient belonged to were analyzed in the SVM. The number of features was determined by calculating the accuracy of the classifier by gradually increasing the number of features up to 10. Accuracy was defined as the ratio between correctly classified subjects and total number of subjects in percentage. Based on this, the SVM calculated an optimal decision rule to separate the two groups in the most optimal way. This was done by leaveone-out cross-validation by extracting one patient for testing, and using the remaining patients to train the model. This process was repeated until all patients had been left out. Once a decision rule was determined, a classification accuracy for each of the 10 features was calculated.

Statistical analysis
The null hypothesis was that SVM analysis of various SNPs could not predict the required opioid dose in cancer pain patients. All data are reported as mean ± standard deviation. Results from SVM classification were analyzed using Chi square tests. P values below 0.05 were considered statistically significant.

Discussion
This study aimed to investigate whether SVM was able to identify associations between genetic variability and required opioid dose in cancer pain patients. None of the chosen 18 SNPs in the three candidate genes showed significant association with opioid dose, which support earlier findings from the EPOS study in which regular linear regressions were unable to identify correlations [6]. Hence, SVM analysis did not provide additional information regarding prediction of opioid dose using genetic profiling.

Limitations
A single SNP may only explain a minor part of analgesic variability. A recent study showed that combinations of genetic variants, e.g. in OPRM1 and COMT better explained variability in morphine consumption than single genetic variants [14]. Thus, one advantage of SVM is the possibility to include several SNPs in one analysis, compared to simple linear regression. However, for each SNP three genotypes exist. When using a binary variable which is necessary for SVM analysis, a dominant genetic model is assumed, which may not be optimal. Furthermore, the high number of SNPs included in the analysis may result in a lower accuracy. Additionally, if epistasis, which is the interaction between genes, is present, the effect of one SNP may be altered or masked by the effect of another SNP and thereby reduce the power to detect genetic associations. Thus, from the present study, it cannot be excluded that some SNPs in the selected genes are associated with required opioid dose.
Various statistical methods (based on general linear models) to predict and assess data relationship exist, but here, a SVM approach was selected, according to two reviews [15,16]. Machine learning differs from conventional statistics, in that there is no predefined model and assumption of data normality, and each patient is dbSNP single nucleotide polymorphism database identification, OPRD δ-opioid receptor, OPRM μ-opioid receptor, OPRK κ-opioid receptor, COMT catechol-Omethyltransferase, A Adenine, G Guanine; C Cytosine, T Thymine classified at the individual level rather than the group. Furthermore, SVM can find non-linear relationships in data, and assess complex associations between several parameters. The latter is different from the traditional one-at-time approach in statistics, where relatively few variables can be tested. On the contrary, the SVM model can be over-fitted to the data and thus loose generalizability. Moreover, the method is relatively new, and many clinicians and researchers are not familiar with the method and output from the model. Although the method of SVM presents itself with several limitations, methodological limitations of the study design itself may also have had an influence on the result. First, as opioid dose was the primary outcome in the EPOS study, it was used as outcome in the present study as well.
Here it was anticipated that opioid dose is related to opioid response, i.e. high dose = less responsiveness to opioid analgesia, and low dose = high responsiveness to opioid analgesia, however this is only a rough estimation and many other factors may be important. Hence, a composite score taking pain intensity, opioid dose and side-effects into consideration might be a better outcome for association analyses of opioid efficacy [17]. However, such a composite score has not been developed or validated for cancer pain patients. In addition, only three opioid receptor genes were included in the analysis. Future studies should include more, as well as SNPs within genes other than those related to opioid receptor signaling, e.g. genes coding for pharmacokinetic factors. For instance, associations between SNPs within the ATP-binding cassette transporter-and cytochrome P450 genes and opioid analgesia have been found in both healthy volunteer and patient studies [4,17].
Moreover, as human genetic factors only account for part of the inter-individual difference in pain sensitivity, several cofactors may also influence opioid consumption during the post-operative period. These include age, gender, mood, anxiety, drug-drug interactions and epigenetic factors. Thus, human experimental pain studies, conducted in a controlled setting, have shown to be of value to explore the genetic contribution to both pain sensitivity and analgesic responses [18].
Authors' contributions PK and FS collected, analyzed and interpreted data from the EPOS study. AEO, DG, MG and AMD analyzed and interpreted the SVM data regarding associations between opioid consumption and genetic profiling. All authors contributed to the final manuscript. All authors read and approved the final manuscript.