microRNA levels can infer something about a person’s physiological state

microRNAs (miRNA) are noncoding, single-stranded, RNA molecules that are 18-24 nucleotides long.¹ They play a key role in the regulation of gene expression by repressing translation and degrading mRNA.² Previous studies have shown that a change in miRNA expression profile correlates with the progression of many diseases, including neurodegenerative diseases.^3–5 In multiple sclerosis (MS) patients some miRNAs are also dysregulated and can be up- or down-regulated compared to healthy people.⁶ Through human practices we found that relapsing-remitting MS (RRMS) was the best type of MS to focus on. As miRNA can be found in the peripheral blood, they are potential biomarkers for minimally invasive diagnosis and reproducible testing.⁷ Their stability and accuracy in blood samples also makes them suitable biomarkers.^7,8 This part of our dry-lab aims to find an miRNA combination in the blood that is specific for RRMS. The RRMS-specific miRNA combination will be incorporated in our test design for the diagnosis of RRMS.

Data processing

First, we build a classification model where we classify miRNA expression data of MS patients and healthy controls. For this classification, we used the dataset obtained by Cox et al. (2010).⁹ In addition to the miRNA expression data of 37 healthy controls, this dataset contains the miRNA expression data of 18 primary progressive, 17 secondary progressive and 24 RRMS patients. The full dataset was filtered to obtain a dataset with only RRMS patients and healthy controls. The Illumina BeadArray reader was used to determine the miRNA expression values. This programme reports a detection p-value after comparing the miRNA expression data with the background values determined by the negative control probes. As the detection p-value indicates the significance of miRNA detection, the miRNA expression values with a p-value higher than 0.05 were replaced by "Not a Number" (NaN) values. After a binary logarithmic (log2) transformation, the data was processed according to the same procedure used by Cox et al. (2010).⁹ To account for deviation and outliers, a baseline transformation to the median was performed. A quantile normalisation was performed to decrease the technical variability between the samples.¹⁰
 

Classification of miRNA expression profiles of RRMS patients and healthy controls

The machine learning algorithm, random forest, was used for the classification of RRMS patients and healthy controls. By comparing the predicted values with the true values, our model obtained an accuracy of 0.69. The miRNAs with a prominent role in the classifications were determined through a method called permutation importance. This method permutates the miRNA expression values of each miRNA separately and recalculated the accuracy of the model. Positive importance values indicate a decrease in accuracy after permutating the sample, while negative values indicate an increase.¹¹ It was determined that there are seven miRNAs with a positive importance value playing a prominent role in this classification (Figure 1). We also found two miRNAs with a negative importance value. This means that randomly permutating the expression values of these miRNAs increases the accuracy of the model.

MS-specific miRNA combination excluding mimic diseases

Through human practices, we found that Dr. Pablo Villoslada (Professor of Neurology and MS researches, Barcelona) was concerned about the specificity of the miRNA for the diagnosis of MS. During progression of diseases similar to RRMS, referred to as mimic diseases, the same miRNA could be upregulated. This similarity in expression profile makes it difficult to distinguish between MS and these mimic diseases. To prevent our test from detecting a mimic disease, we introduced the use of the Human miRNA Disease Database (HMDD) in our pipeline. We searched for the seven miRNAs found to have a positive importance value in the previous step.¹² This database lists miRNAs found to be up- or down- regulated in disease. We were unable to identify two of them (HS 263.1 and HS 65) in this database, due to the use of an older nomenclature structure. Therefore, with the miRNA sequence obtained from the paper by Cox et al. (2010), we performed sequence alignments in the miRBase, which lists miRNA names and sequences.^9,13 Using this webtool with the default settings, we tried to find the names of the two miRNAs according to the new nomenclature structure. However, no similar miRNAs were found. Therefore, we continued the analysis with the five remaining miRNAs. By searching for their names with the default settings in the HMDD, we obtained a table showing all diseases where the five miRNAs are found to be dysregulated (Table 1). For hsa-miR-17-5p, the search term hsa-miR-17 was used as HMDD does not distinguish between two mature miRNAs derived from the opposite arms of a pre-miRNA (hsa-miR-17-3p and hsa-miR-17-5p).

As all five miRNA together indicate MS, we implemented so-called AND gates on our diagnostic test platform. By implementing these AND gates, we will only obtain an output when all RRMS-indicating miRNAs are present. To prevent the diagnostic test from detecting mimic diseases, we can implement NOT gates in our diagnostic test platform. These NOT gates should represent a miRNA dysregulated in the disease progression of mimic diseases where it is not in the progression of RRMS. If this miRNA is present, the NOT gate will not be activated and no output will be seen.

To find a proper NOT gate, we first needed to find the mimic diseases where all five miRNA are dysregulated. If all five miRNAs are also dysregulated in a mimic disease, we will be unable to distinguish between RRMS and this mimic disease by only applying an AND gate in our diagnostic test platform. Through a literature search, we obtained a list of diseases that mimic MS (Table 2). We found that in most mimic diseases not all five miRNAs were dysregulated. This means that the combination of these five miRNAs distinguishes RRMS from those mimic diseases, and means AND gates should be applied in our diagnostic test platform. However, all five miRNA are dysregulated in the mimic disease diabetes. Therefore, we searched for an miRNA that can be implemented as a NOT gate for diabetes. This miRNA should be involved in the disease progression of diabetes where it is not in RRMS. If this miRNA involved in diabetes is present in the sample, the gate is not activated and no output will be produced. Whereas the gate will be activated if the miRNA is not present.

By searching ’diabetes’ and ’multiple sclerosis’ in the HMDD with default settings, we obtained two lists of miRNAs dysregulated in diabetes and MS. After comparing these 2 lists, we found 84 potential miRNAs to function as a NOT gate. A criteria for this miRNA was to be involved in both men as well as in women. As MS is an autoimmune inflammatory disease, the miRNA in the NOT gate should not be involved in processes related to immune cells and inflammation as they might have overlap with MS. Through literature searches, we found hsa-miR-1287 to be the most suitable miRNA to make up the NOT gate for diabetes. This miRNA has not been found to be dysregulated in immune cells and inflammation.

Conclusion

To distinguish miRNA expression data of RRMS patients and healthy controls, seven miRNAs are needed. The accuracy of this classification was 0.69. With an additional miRNA, we found an miRNA combination specific for RRMS excluding diseases that mimic MS. By implementing the detection of these eight miRNAs in AND and NOT gates in our diagnostic test platform, we can diagnose RRMS and minimise the chance of false positives.