miRNA profiling as a complementary diagnostic tool for amyotrophic lateral sclerosis

Data source and batch correction

The raw counts of miRNA were downloaded from the National Center for Biotechnology Information (NCBI) Gene Expression Omnibus (GEO) database²¹ of the accession number GSE168714, contributed by Magen and colleagues²⁰. The dataset was accessed at https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE168714 from March 1st to 3rd, 2023. The dataset comprises the annotated counts by small RNA-seq of the RNA extracted from the plasma of 253 ALS and 103 control²⁰. The batch number of data collection and ALS/control state were also available with the raw count. To account for the differences in our study design compared to the original work of Magen and colleagues, we performed our own batch correction using the following method. Firstly, we temporarily excluded highly present miRNAs whose raw count represented more than 2% of the total counts (13 miRNAs excluded, as detailed in Supplementary File 1) during the calculation of batch correction coefficient. Secondly, we calculated the sum of miRNA counts for each sample. Thirdly, we determined the average sum of miRNA counts for each batch. We then defined a batch coefficient as the quotient of the maximum average sum of all batches divided by the average sum of the batch. Finally, we calculated the corrected count, including the temporarily excluded highly present miRNAs, by multiplying the raw counts by the batch coefficient.

Machine learning

The strategy to identify the key miRNAs was adopted from our previous studies^22,23. Briefly, the transposition of the corrected miRNA counts served as the input file to the machine learning program RapidMiner²⁴. For clarity, bold italicized text is used to denote the terminology in RapidMiner. In the RapidMiner process setting, patient ID was used as ID; patient ALS state was used as Label, and the corrected miRNA counts were used as Attributes to perform training. Rule Induction was adopted as the algorithm, and the overall training and validation program was shown in Supplementary File 2, with the following parameters. The Split Data operator separated the sample library into 80/20 sets by shuffled sampling for model building and independent validation, respectively. Rule Induction was performed with the criterion of information gain, sample ratio of 0.9, pureness of 0.9, and minimal prune benefit of 0.25, while ten times of Cross Validation was used to improve the model. Finally, the Apply model operator utilized the 20% blind set to validate the generated Rule Model, and the Performance operator demonstrated the performance of the model.

mRNA targets of miRNAs

DIANA-TarBase v7 was used to find the experimentally validated mRNA targets of miRNAs²⁵. Data were accessed at http://diana.imis.athena-innovation.gr/DianaTools/index.php?r=tarbase/index from April 13th to 20th, 2023.

Interaction network and enrichment

String-db version 11.5 was used to generate the interaction network and pathway enrichment²⁶. Data were accessed at https://string-db.org/ from April 20th to 23rd, 2023. Interaction network was generated with highest confidence (0.9) and disconnected nodes in the network were hidden.

Intersection analysis

Venny 2.0 was used to generate the Venn diagram²⁷. Data was accessed at https://bioinfogp.cnb.csic.es/tools/venny/index2.0.2.html from April 20th to 23rd, 2023.

Statistics

Student’s T-test was used to estimate the significance of the difference between two groups.

Heterogeneity of ALS genetics

Recent advancements have greatly enhanced our comprehension of the genetic origins of familial ALS. Approximately 40–55% of familial ALS cases can be accounted for by variations in well-known ALS-linked genes⁷. While more than 50 potential causative or disease-modifying genes have been identified, pathogenic variants in SOD1, C9ORF72, FUS, and TARDBP are most frequent, whereas variants in other genes are relatively uncommon²⁸. However, in cases of sporadic ALS, diagnostic progress has only elucidated a fraction of the cases, with the etiology remaining unexplained in over 90% of patients²⁹. It is widely accepted that genetic risk factors play a significant role in sporadic ALS, with heritability estimated at approximately 60% based on twin studies³⁰. However, despite extensive genetic association studies, identifying heritable genetic risk factors in sporadic ALS remains elusive.

Despite decades of research, ALS's underlying causative pathogenic mechanisms remain uncertain, particularly in sporadic cases. The development and progression of the disease are likely influenced by multiple factors rather than a single initiating event³¹. Moreover, genetic and phenotypic variations among patients pose challenges in comprehending and drawing conclusions about the general pathogenic mechanisms of ALS. Given the extensive number of genes and cellular processes implicated in ALS, numerous disease mechanisms have been proposed, including disruptions in RNA metabolism³², compromised protein homeostasis³³, defects in nucleocytoplasmic transport³⁴, impaired DNA repair³⁵, mitochondrial dysfunction³⁶, oxidative stress³⁷, disturbances in axonal transport³⁸, and oligodendrocyte dysfunction³⁹. Further clarification is required to determine the timing and extent to which each of these mechanisms contributes to the pathogenesis of ALS.

As an attempt to see whether ALS clinical phenotypes could be differentiated by miRNAs, we constructed new cohorts of ALS patients from the same GSE168714 dataset by their clinical phenotypes: bulbar-onset or non-bulbar-onset, with 83 or 170 patients, respectively (Supplementary File 10), and these new cohorts of patients were subjected to de novo analysis in machine learning. However, the performance of the newly-established model was poor, with a recall rate of 75.9% and 24.6% for the cohorts, respectively. Thus, machine learning could not differentiate the cohorts of ALS, at least under the present condition. We further analyzed whether the identified miRNAs express differentially in the cohorts of ALS (Supplementary File 11). We found that miR-206 and miR-205 express differentially in the cohorts of ALS, but not significantly. We suspect that the sample size is the bottleneck to uncover the nature of the ALS cohorts. We also summarized other publicly-available miRNA datasets (Supplementary File 12). However, their sample size is insufficient for an independent machine learning study, and none of the datasets meets the requirement for validation of the current model that all eight miRNAs used in the machine learning model must be available.

miR-206

miR-206 participates in various stages of muscle differentiation, encompassing alternative splicing, DNA synthesis, and cell apoptosis⁴⁰. During development, miR-206 hinders the activity of Pax7 and Pax3, effectively restricting the proliferative potential of satellite cells while promoting their differentiation into myogenic progenitor cells. Conversely, reducing miR-206 leads to the overexpression of Pax7 and Pax3, which consequently inhibits the differentiation of myoblasts. As Pax7 and Pax3 are known pro-survival factors, the down-regulation of miR-206 can induce apoptosis⁴¹. Thus, miR-206 plays a protective role and facilitates the regeneration of neuromuscular junctions following acute nerve injury, particularly in the context of ALS⁴². Knock-out of miR-206 delays and mutilates muscle reinnervation in ALS mouse models of SOD mutant⁴³. Recent findings indicate elevated levels of miR-206 in the plasma of ALS patients and could indicate of disease progression^44,45.

miR-205

miR-205 exhibits significant expression levels in various human epithelial tissues, including the breast, prostate, skin, eye, and thymus. Its primary role in these tissues is crucial to tissue morphogenesis and homeostasis. Specifically, it upholds the epithelial phenotype by directly targeting two transcription factors: zinc finger E-box-binding homeobox 1 (ZEB1) and ZEB2, which repress E-cadherin and other genes associated with polarity⁴⁶. During the early stages of embryonic development, miR-205 is expressed in trophoblasts, where it regulates placental development by suppressing the Mediator of RNA polymerase II transcription subunit 1 (MED1)⁴⁷. Moreover, in embryonic development, miR-205 governs the differentiation of extraembryonic endoderm and spermatogenesis by influencing cell migration and adhesion genes⁴⁸. In the mammary gland, miR-205 displays high expression levels in the basal stem cells. Overexpression of miR-205 has been shown to induce the expansion of the progenitor cell population while reducing cell size and promoting cellular proliferation. These effects are achieved by repressing PTEN⁴⁹. In this context, miR-205 regulates the production of the basement membrane protein complex laminin-332 and its receptor integrin-β4, thereby ensuring proper tissue polarity and morphogenesis⁵⁰. In the skin epidermis and stratified epithelia of the esophagus and tongue, miR-205 has been found to play a significant role in expanding the stem cell population through its regulation of PI3K signaling⁵¹. Additionally, by influencing the same signaling pathways, miR-205 enhances the migration of human epidermal and corneal epithelial keratinocytes, thereby contributing to wound healing and corneal development⁵².

miR-376a and miR-412

The physiological roles of miR-376a and miR-412 are not fully understood yet, despite some studies reported its participation in cancer and neurological disorders. For example, increased levels of miR-376a have been observed in the T cells of patients with multiple sclerosis (MS)⁵³, and miR-412 may inhibit clear cell renal cell carcinoma progression⁵⁴. Conversely, in the late-onset form of Alzheimer's disease (LOAD), miR-376a has been identified as down-regulated in the brain⁵⁵. Meanwhile, the expression change of miR-412 has been mentioned in the brain of alcohol use disorder⁵⁶, and also in Alzheimer's disease⁵⁷.

BCL2

BCL2 is targeted by miR-205⁵⁸ and controls caspase activation and the initiation of programmed cell death⁵⁹ and thus regulates neuronal development and neurodegeneration⁶⁰. Several lines of evidence show that BCL2 probably involves in ALS pathological progression. Epidemiologic studies found altered expression of BCL2 in ALS spinal cord motor neurons⁶¹ and post‐central gyrus⁶². In vivo studies showed that overexpression of BCL2 prolongs the survival of the ALS mouse model⁶³ and improves neuromuscular function⁶⁴. In vitro studies revealed that ALS-associated mutant SOD1 aggregates BCL2⁶⁵ and advocates BCL2 conformational changes⁶⁶.

NEFH

Neurofilament heavy polypeptide (NEFH) is one of the three intermediate filament proteins forming neurofilaments⁶⁷. NEFH is targeted by miR-205⁶⁸ and could be phosphorylated by GSK3β⁶⁹ and regulate the Akt-β-catenin pathway⁷⁰. Moreover, epidemiological study showed that NEFH mutation^71,72 and expression⁷³ is associated with ALS. Besides, NEFH mutation or expression is associated with other disorders of central or peripheral neural system, including schizophrenia⁷⁴, alcoholics⁷⁵, and Charcot-Marie-Tooth neuropathy⁷⁶.

OPTN

OPTN, also known as optineurin, is a highly conserved protein in various species⁷⁷. OPTN is target by miR-205⁶⁸ as well. It plays diverse roles in vesicular trafficking, NFKB/NF-κB signaling, and autophagy. Specifically, OPTN has been identified as an autophagy receptor that facilitates the connection between ubiquitinated autophagy substrates and MAP1LC3/LC3-positive phagophore membranes⁷⁸. Furthermore, mounting evidence suggests that OPTN acts as an inducer of autophagy, initiating the autophagic process^79,80. Moreover, studies indicate that OPTN's involvement in autophagic initiation can commence as early as the formation of autophagosomal membranes^81,82. These groundbreaking findings underscore the multifunctional role of OPTN as a potential autophagy receptor throughout the autophagic process, expanding beyond its traditional perception as a receptor operating solely at a single stage of autophagy. OPTN is gathering attention in ALS research, since variants of this gene are associated with ALS^83–85. Moreover, OPTN mutant induces neuronal cell death by mediating mitophagy⁸⁶, autophagy and ER stress⁸⁷. Interestingly, OPTN mutation might be the common cause of ALS and corticobasal syndrome (CBS)⁸⁸.

PI3K in ALS

The PI3K-askt signaling pathway governs metabolism, cell survival, motility, transcription, and cell-cycle progression. In recent years, studies have revealed the involvement of the PI3K-Akt signaling pathway in neurodegenerative diseases. For instance, butylphthalide has been shown to activate the PI3K-Akt/GSK-3β signaling pathway in an ischemic cerebral infarction model, reducing nerve function damage and protecting local nerve cells⁸⁹. Therefore, therapeutic strategies for ALS targeting the PI3K-Akt pathway has been shown to increase anti-apoptotic protein expression levels, reduce pro-apoptotic protein expression levels, and improve cell survival rate and mitochondrial function in ALS⁹⁰. Moreover, studies by Xiang and colleagues have found that AEG-1 can regulate the PI3K-Akt pathway⁹¹, and the absence of AEG-1 in ALS motor neurons inhibits the PI3K-Akt pathway and increases cell apoptosis⁹¹. Thus, dysregulated miRNAs may promote ALS pathology by mediating PI3K-Akt signaling pathway.