Plasma isomiRs as Candidate Biomarkers for Amyotrophic Lateral Sclerosis

Abstract

Background and Objectives

There are no FDA-approved diagnostic biomarkers for amyotrophic lateral sclerosis (ALS). TDP-43 is a known cofactor in the cleavage of long premature microRNAs (miRNAs) into their short, mature products. isomiRs are miRNA variants that differ in their 5′ and 3′ end points and regulate distinct mRNA targets. In this study, we tested the hypotheses that circulating isomiR profiles differ in the context of TAR DNA-binding Protein pathology and that isomiRs are superior to miRNAs for classification of ALS.

Methods

We obtained RNA from plasma samples of 14 patients with ALS and 14 age-matched and sex-matched controls for sequencing on a NextSeq 2000. Data were processed using Unique Molecular Identifier tools and a custom pipeline designed to match miRNA variant sequences without mismatches. Differential expression (DE) was identified using DEseq2 at FDR ≤ 0.1. XGBoost classifiers were built using a subset of (Model 1) isomiRs or (Model 2) miRNAs that were present above a median threshold in all sequencing batches. Parameters were tuned using grid search and 10-fold cross-validation while training to distinguish ALS samples from controls among a single large public data set. Models were then validated on in-house samples and 1 publicly available holdout data set.

Results

Fourteen (0.2%), 355 (2.7%), and 14 (0.7%) isomiRs were differentially expressed in in-house plasma, public ALS plasma, and public ALS serum, respectively. One (0.1%), 94 (5.5%), and 13 (2.4%) miRNAs were differentially expressed, respectively. Model 1 accurately classified in-house ALS plasma and public ALS serum (area under the curve [AUC] = 0.87) and did not distinguish 40 of 41 Alzheimer disease samples from control plasma (GSE215789; AUC = 0.47) or 60 of 77 Parkinson disease samples from control whole blood (GSE180193; AUC = 0.55). In comparison, Model 2 using miRNAs performed worse on in-house plasma (AUC = 0.49).

Discussion

Analyzing individual isomiRs may improve the performance of circulating noncoding RNAs as diagnostic biomarkers of ALS.

Introduction

Despite the exploration of several promising candidates from blood and CSF, no reliable diagnostic or prognostic biomarkers for amyotrophic lateral sclerosis (ALS) are available in the clinical setting.¹ MicroRNAs (miRNAs) are short, noncoding RNAs that mediate post-transcriptional regulation via targeted degradation of messenger RNA (mRNA) transcripts or the inhibition of translation.² Specific miRNAs have been linked to the pathogenesis of ALS. This includes transactive DNA-response binding protein 43 (TDP-43)–independent mechanisms, for example, as mediators of oxidative stress responses³ or cofactors in responses to nerve injury.⁴ This also includes TDP-43-dependent mechanisms, for example, as endogenous regulators that are suppressed during disease pathogenesis⁵ or as binding partners for mislocalized TDP-43.⁶

Given that miRNAs are inherently stable in circulation,⁷ these have been explored as circulating prognostic⁸ biomarkers and as diagnostic biomarkers in both cell-free compartments^9-11 and in extracellular vesicles.¹² However, the success of miRNA-based prognostic biomarkers for ALS in clinical trials, for example, remains mixed.¹³

After they are transcribed, miRNAs are folded into a hairpin structure (termed the pri-miRNA). They are first cleaved by the endonuclease Drosha into a pre-miRNA hairpin and then exported from the nucleus into the cytoplasm. Then, they are cleaved by the endonuclease Dicer into a mature noncoding RNA.¹⁴ Mature miRNAs are usually 21–22 nucleotides in length and target mRNAs for post-transcriptional regulation based on their “seed” sequence. The seed is an approximately 7-nucleotide sequence starting from the 2nd base at the 5′ end, which is the reverse complement of sequences in target mRNAs. The 2-step cleavage process of miRNA maturation is imperfect and can lead to the production of miRNA variants that have different 5′ and 3′ end points (termed “isomiRs”).¹⁵ isomiRs by nature can have distinct seed sequences (given that they may start at different 5′ end points) and, therefore, different mRNA targets.¹⁶ isomiR production is conserved, nonrandom, and often stimulus-induced.¹⁷ Hairpin sequence¹⁸ and RNA-binding cofactors¹⁹ both appear to influence isomiR production. While TAR DNA-binding Protein (TDP) is a known cofactor in miRNA biogenesis,²⁰ it is unknown whether mislocalization of nuclear TDP to the cytoplasm affects the alternative cleavage of isomiRs. Furthermore, it is essentially unknown whether isomiRs may serve as reliable biomarkers of neurodegenerative disease.

Given the possibility that mislocalization of TDP affects isomiR production, we hypothesized that loss of nuclear TDP in ALS may (1) destabilize Drosha-pri-miRNA interactions, resulting in differential cleavage of 3′-isomiRs, or (2) increase cytoplasmic pre-miRNA binding and result in differential production of 5′-isomiRs. Therefore, we hypothesized that exploration of isomiR profiles might (1) reveal differences in the diversity of isomiR content in the context of TDP mislocalization and (2) improve the diagnostic capacity of miRNA-based biomarkers. To test these hypotheses, we analyzed isomiR profiles in plasma from patients with ALS compared with age-matched and sex-matched controls. We further explored the composition of isomiR profiles in these samples and compared them with public ALS, Alzheimer disease (AD), and Parkinson disease (PD) data sets.

Methods

In-House Patient Selection

We identified 14 individuals who underwent electromyography and were diagnosed with ALS by board-certified neuromuscular neurologists according to Gold Coast criteria. All individuals had either upper or lower motor neuron limb onset. We further identified 14 age-matched and sex-matched healthy controls who scored above reference norms on the Mini-Mental Status Examination, had no significant cognitive or psychiatric history, and had an unremarkable neurologic examination. Table 1 summarizes basic demographic and clinical features.

Table 1.

Demographic Characteristics of In-House Samples

Characteristic	ALS N = 14	Control N = 14	p Value
Age at sampling	59 (11)	64 (4)	0.1
Sex, (%)			0.4
Female	6 (43)	8 (57)
Male	8 (57)	6 (43)
Disease duration at sampling (Years, Mean [SD])	6.4 (5.1)
Years from onset to diagnosis (Years, Mean [SD])	2.50 (4.29)
Symptom onset site, (%)
Cervical LMN	5 (36)
Lumbosacral LMN	4 (29)
Lumbosacral UMN	5 (36)
Age at diagnosis	54 (12)
Age at onset	52 (12)
History of military service? (%)	2 (15)
Height	66.6 (4.0)
Weight	178 (53)
BMI	28.0 (6.9)
Seated FVC percentile	78 (21)
Total ALS-FRS score	31 (8)
Penn Upper Motor Neuron score (max 32)	12 (11)

Abbreviations: ALS = amyotrophic lateral sclerosis; ALS-FRS = ALS Functional Rating Scale.

Sex and age at sampling for ALS and control samples are listed and compared using chi-square and Wilcoxon rank-sum testing, respectively. Several ALS clinical measures are listed separately for patients with ALS only, including Functional Rating Scale (FRS), Penn Upper Motor Neuron score, seated forced vital capacity, and characteristics around symptom onset and diagnosis.

Whole blood was collected through a standard venipuncture procedure and processed using standard operating procedures, including plasma fractionation and storage at −80C.

Comparison With Public Data Sets

Public RNA-seq data sets for comparison were downloaded from Gene Expression Omnibus according to the following accessions.

GSE168714⁸—a set of Illumina miRNA-seq data from plasma collected from 113 patients with ALS and 103 controls. Most samples were from female participants (n = 115, 53%). Samples were divided into 25 batches.

GSE148097¹⁰—a set of Illumina miRNA-seq data from serum collected from 13 patients with ALS and 6 controls. Available demographic data were limited.

GSE215789²¹—a set of Illumina miRNA-seq data from plasma collected from 40 patients with AD and 41 controls. Most samples were from male participants (n = 45, 56%).

GSE180193²²—a set of Illumina miRNA-seq data from whole blood collected from 60 patients with PD and 77 controls. Available demographic data were limited.

FASTQ files from experiments with reads containing Unique Molecular Identifiers (UMIs) were processed using the pipeline described further. FASTQ files from experiments without UMIs were first quality and adapter trimmed using cutadapt v5.1.²³ Trimmed reads were then passed through the pipeline described further. Available demographic data were limited.

Samples from a sequencing batch that included only control or only patients with disease were excluded. Samples with fewer than 100,000 raw reads attributable to miRNAs were also excluded. Batch data were unavailable for 3 of 4 public data sets; those samples were treated as a single sequencing batch.

Total RNA Extraction and Small RNA Sequencing

Plasma samples were thawed to 37°C in a water bath, after which total RNA was extracted from 200 μL per patient using the QIAgen kit #217184, according to the manufacturer's instructions (QIAgen, Germantown, MD). Experimenters were blinded to the sample identity. Samples were then quantified using the Qubit Kit #Q32880, according to the manufacturer's instructions. Concentrations ranged between 0.3 and 4.4 ng per microliter (eTable 1).

Sequencing libraries were then prepared in a single batch with the QIAseq Kit #331502 using 5 μL of eluent as the starting material, according to the manufacturer's instructions (QIAgen, Germantown, MD). Libraries were indexed using the QIAseq kit #331595 (QIAgen, Germantown, MD), which contains UMIs that help identify artifacts such as qPCR duplicates.²⁴ Libraries were investigated using the Bioanalyzer kit #NC1758992 (ThermoFisher Scientific, Waltham, MA), and all were found to have appropriate peaks at approximately 180bp. Libraries were subsequently multiplexed at 4 nM per sample and sequenced on a NextSeq 2000 (Illumina, San Diego, CA) on a single flow cell as 75-bp single-end reads with an 8-bp index.

FASTQ files were de-multiplexed on BaseSpace using bclconvert and subsequently analyzed using a custom pipeline. The pipeline included UMItools²⁴ for adapter trimming, quality filtering, and UMI counting. Reads were quality filtered above a phred33 cutoff of 15. The standard Illumina 3′ adapter sequence AAC​TGT​AGG​CAC​CAT​CAA​T was identified immediately 2 bases 3′ of the UMI and removed via adapter trimming. Reads were counted by UMI, and the dominant sequence was retained in all cases.

Pre-miRNA hairpins from the most recent miRBase version 22.1²⁵ were processed into unique isomiRs starting or ending up to 3 bases upstream or downstream from the 5′ and 3′ end points of the canonical miRNA. isomiRs were named according to the following convention: hsa-miR-21-5p|0|0|1, where the first component names the parent miRNA, the second component names the 5′ modification, the third component names the 3′ modification, and the fourth component numbers the pri-miRNA from which the sequence derives. Only reads that perfectly matched these sequences were retained in a custom annotator written using C++23 and compiled using GNU Compiler Collection 13 on a MacBook Pro with M1 architecture (Apple, Cupertino, CA). Finally, reads between 18 and 26 nucleotides were retained for downstream analysis.

Whole-Genome Sequencing to Identify Pathogenic Variants

DNA was extracted from peripheral blood using the QuickGene DNA whole-blood kit (Autogen, Holliston, MA). Pathogenic variants were screened using whole-genome sequencing analyzed through Geneticist Assistant software (Soft Genetics, State College, PA),²⁶ and pathogenic variants of >30 hexanucleotide repeats in C9orf72 or >22 repeats in ATXN2 were screened using modified repeat-primed Polymerase Chain Reaction.²⁷

Batch Correction and Differential Expression

Technical preparation has been shown to influence isomiR content.²⁸ GSE168714 was sequenced in 25 batches, ranging in size from 6 to 20 samples. In-house data were sequenced in a single batch. Batch data were not available for the remaining 3 data sets. To account for technical bias as much as possible, raw reads were first normalized to reads per million (RPM) sequenced reads and then batch-corrected by processing the entire collected data sets using the removeBatchEffect function from limma.²⁹ All analyses were run using R version 4.4.3.

Reads were collected at the level of the individual isomiR or at the level of the parent miRNA. Differential expression (DE) was assessed per data set using DEseq2 and raw read counts, including condition (control vs ALS samples) and sequencing batch (where appropriate) as design factors.³⁰ Individual RNAs with Wald test FDR ≤0.1 and |log₂ fold change| ≥ 1.0 were considered differentially expressed.

Classification Models

XGBoost³¹ classifiers were trained on all ALS and control samples from GSE168714 and included features of isomiRs (Model 1) and miRNAs (Model 2) with median read per million (RPM) count greater than or equal to 5 in all sequencing batches. Grid search (per the scikitlearn³² package) with 10-fold cross validation was used to identify optimal model parameters.

Models were then tested on the 4 remaining holdout data sets. A third model using only the 7 isomiRs with greatest feature importance from Model 1 was also built and validated by the abovementioned approach. All XGBoost analyses were run on Python v3.13 using XGBoost v.2.1.4. False-positive rate, true-positive rate (TPR), and area under the curve (AUC) were calculated using scikitlearn.

Alternative Cleavage Metrics

To test whether TDP-related disease activity modified the total pool of available isomiRs, i.e., produced a measurable shift in the alternative cleavage of miRNA precursors, we compared the number of unique isomiRs in each sample using the Simpson diversity¹⁹ and Shannon entropy indices.³³ Simpson diversity index was calculated using the formula $\mathrm{D}=1-{\displaystyle \sum }\frac{n\left(n-1\right)}{N\left(N-1\right)}$, where n represents the read count attributable to an individual isomiR. The index is bounded by 0 and 1, where 1 represents the absence of diversity within a population. Shannon entropy was calculated using the formula $\mathrm{H}=-1\hspace{0.17em}*\sum p_i*\ln \left(p_i\right)$, where $p_i$ represents the proportion of reads attributable to an individual isomiR. Shannon entropy is not bounded, and higher values indicate greater diversity. Indices were calculated at the level of the sample.

Given the anticipated effect of sequencing depth on the number of unique RNAs identified, linear models using base R were calculated with the formula $number\hspace{0.17em}of\hspace{0.17em}unique\hspace{0.17em}RNAs\hspace{0.17em}\sim \log \left(read\hspace{0.17em}depth\right)+Condition:\log \left(read\hspace{0.17em}depth\right)+Batch.$ Linear models using base R were then run to compare Simpson and Shannon indices across data sets, using the formula $index\hspace{0.17em}\sim \hspace{0.17em}Condition\hspace{0.17em}\left[Control\hspace{0.17em}vs.\hspace{0.17em}Disease\right]+Batch$.

Standard Protocol Approvals, Registrations, and Patient Consents

All clinical data were obtained from the Penn Integrated Neurodegenerative Disease Database.²⁶ All individuals provided written informed consent, including collection of clinical data and plasma, and the study was approved by an institutional review board convened at the University of Pennsylvania.

Data Availability

Raw sequencing data (.fastq files) and analytical code will be available to qualified investigators from the corresponding author on reasonable request.

Results

Demographic Characteristics

Living patients selected for study included 14 with ALS and 14 healthy controls (Table 1). Age at sampling (mean ± SD; ALS 59 ± 11 years and control 64 ± 4 years; Welch p = 0.14) and sex (6/14 vs 8/14 female; Fisher p = 0.63) were comparable across groups. Age at symptom onset was 51 ± 12 (mean ± SD) years. The ALS Functional Rating Scale score obtained around the time of plasma sampling was 32 ± 7 (mean ± SD). The percentage seated forced vital capacity with respect to an age-matched reference was 78 ± 21 (mean ± SD), and the University of Pennsylvania Upper Motor Neuron score³⁴ was 12 ± 11 (mean ± SD) out of 32 points total. Six of the 14 patients went through with whole-genome sequencing, none of whom had pathogenic variants in ALS-associated genes (e.g., SOD1 and FUS). None of the 14 patients had pathogenic C9orf72 repeats. One patient had 27 repeats in a single copy of ATXN2.

All samples (control and ALS) were banked between 2009 and 2023 and selected for study during March 2025. Although there was a trend toward older age among ALS samples, there was no significant difference in the age of samples at the time of processing (median (Q1, Q3) ALS 10.9 (10.2, 12.3) vs control 7.2 (6.0, 11.5); Wilcoxon rank-sum exact p = 0.05).

Public data sets as summarized in Table 1 included the following: 116 ALS and 103 control plasma samples (GSE168714), 13 ALS and 6 control serum samples (GSE148097), 40 AD and 41 control plasma samples (GSE215789), and 60 PD and 77 control whole-blood samples (GSE180193).

Sequencing Statistics

To validate our re-analysis, we compared raw read counts obtained from public ALS plasma using our pipeline with those that were posted in GEO data sets. Raw read counts were strongly correlated (Pearson ρ = 0.8, p < 2.2e-16), with differences likely reflecting selective retention/rejection of transcripts across pipelines due to tolerance of mismatches. Across individual samples, reads attributable to specific miRNAs were consistently nearly perfectly correlated (Pearson ρ ranged from 0.981 to 0.999).

Raw counts of 18–26 nucleotide sequences that passed quality filtering for in-house samples were comparable between control and ALS samples (mean ± SD of million reads; ALS 6.8 ± 2.5 and control 8.3 ± 10.3; t test p = 0.59). Read counts in public data were comparable and similarly not significantly different between groups, except in PD whole blood: ALS plasma (ALS 6.1 ± 3.7 and control 6.1 ± 3.1; t test p = 0.96), ALS serum (ALS 8.5 ± 2.5 and control 10.0 ± 3.0; t test p = 0.33), AD plasma (AD 2.4 ± 1.3 and control 2.4 ± 1.2; t test p = 0.99), and PD whole blood (PD 10.7 ± 3.4 and control 12.2 ± 5.4; t test p = 0.05) (eFigure 1A).

The percentage of raw reads attributable to miRNAs was higher in ALS samples in in-house data (mean ± SD; ALS 53.8 ± 7.8 and control 41.7 ± 9.5; t test p = 0.001), lower in public plasma (ALS 22.4 ± 15.8 and control 27.9 ± 16.7; t test p = 0.01), and comparable in public serum (ALS 9.9 ± 7.9 and control 14.1 ± 6; t test p = 0.22) (eFigure 1B). AD (AD, 17.1 ± 10.4 and control 16.7 ± 9; t test p = 0.87) and PD (PD, 27.8 ± 8.7 and control 26.8 ± 13.4; t test p = 0.97) data sets were similarly comparable.

Diversity of Unique isomiR Profiles

Sequencing depth affects the number of unique isomiRs recovered per data set. In a linear model including all samples and accounting for read depth and sequencing batch, the interaction between ALS and read depth revealed a significant decrease in unique isomiRs identified compared with control (β = −83 [-160, −6], p = 0.04). Interaction terms for AD (β = 196 [-52, 444], p = 0.12) and PD (β = 24 [-56, 105], p = 0.6) were not significant.

Given the expected influence of read depth on miRNA coverage, we calculated diversity metrics that have been previously published in work comparing isomiR diversity in different contexts, namely Simpson diversity¹⁹ and Shannon entropy indices.³³ Diversity indices differed significantly across data sets (Analysis of Variance p < 2.2e-16 for both Shannon and Simpson indices; pairwise Wilcoxon rank-sum tests are illustrated in eFigure 2, where * represents a Bonferroni-corrected p value <0.05/√20). Comparison of diversity metrics among ALS and control patients within each separate project did not show significant differences (Figure 1, A and B). In a linear model accounting for sequencing batch, Shannon (ALS β = −0.13 [−0.2, −0.02], p = 0.02) and Simpson (ALS β = 0.02 [0.01, 0.03], p = 0.005) indices revealed a greater diversity of isomiR content in ALS samples compared with controls and no significant difference in either metric for AD (Simpson β = 0 [-0.02, 0.02], p > 0.9; Shannon β = −0.01 [−0.2, 0.2], p > 0.9) or PD (Simpson β = −0.01 [−0.03, 0.01], p = 0.3; Shannon β = 0.06 [−0.08, 0.2], p > 0.9.

Figure 1. Diversity Indices in Disease Samples Compared With Controls.

Boxplots demonstrate median and IQR of the index, either (A) Shannon entropy or (B) Simpson diversity index, while points demonstrate values for individual samples. Points are colored by disease status. The results of Wilcoxon rank-sum tests comparing disease samples with controls in each individual project are plotted above each corresponding pair of boxplots, where * denotes a Bonferroni-corrected p-value <0.05/√10 and ns denotes a value above this threshold. In a linear regression accounting for sequencing batch and including all samples, indices showed greater diversity among ALS samples compared with controls, but not among PD or AD samples compared with controls. AD = Alzheimer disease; ALS = amyotrophic lateral sclerosis; PD = Parkinson disease.

DE of isomiRs and miRNAs

Fourteen (0.2%), 14 (0.7%), and 355 (2.7%) unique isomiRs were differentially expressed in ALS samples compared with control samples from in-house plasma, public ALS serum, and public ALS plasma (Figure 2, A–C; eTable 2). One (0.1%), 13 (2.4%), and 94 (5.5%) of miRNAs were differentially expressed, respectively (Figure 2, D–F; eTable 2). Of these, 3 isomiRs were differentially expressed in both in-house and public ALS plasma, namely hsa-miR-206|0|0|1, hsa-miR-206|0|-2|1, and hsa-mir-30a-5p|0|1|1.

Figure 2. Distinct isomiRs and miRNAs Are Differentially Expressed.

The authors used DEseq2 to identify differentially expressed (DE) isomiRs and miRNAs from in-house samples and from the 2 publicly available ALS data sets. Volcano plots demonstrate the negative log₁₀ FDR (y-axis) and log₂ fold change (x-axis) as calculated using DESeq2 on raw data and accounting for batch in the column design where able. Points are colored according to whether the FDR surpassed a chosen threshold of 10% (red = no, blue = yes). Data are plotted separately for DE calculations from miRNAs (A–C) and isomiRs (D–F). Similarly, projects are plotted separately for in-house (A and D), publicly available serum (B and E), and publicly available plasma (C and F). ALS = amyotrophic lateral sclerosis; DE = differential expression.

Among in-house samples, isomiRs hsa-mir-30e-3p|0|-2|1, hsa-mir-141-3p|0|0|1, hsa-mir-10b-5p|0|0|1, hsa-mir-15a-5p|0|0|1, hsa-mir-1277-5p|-3|0|1, hsa-mir-30a-5p|0|1|1, hsa-mir-4433b-5p|2|0|1, and hsa-mir-148b-3p|0|-3|1 were decreased in ALS samples compared with controls, while isomiRs hsa-mir-4286|0|1|1, hsa-mir-324-5p|0|-2|1, hsa-mir-339-5p|0|-3|1, hsa-mir-206|0|-2|1, hsa-mir-206|0|0|1, hsa-mir-133a-3p|-1|1|1, and hsa-mir-133a-3p|-1|1|2 were increased in ALS samples compared with controls (eTable 2).

In public serum data, isomiRs hsa-mir-181a-5p|0|0|2, hsa-mir-181a-5p|0|0|1, hsa-mir-423-3p|1|0|1, hsa-mir-222-3p|0|3|1, hsa-mir-423-3p|0|0|1, hsa-mir-1843|1|1|1, hsa-mir-30e-3p|0|0|1, hsa-mir-151a-3p|2|0|1, hsa-mir-21-5p|0|1|1, hsa-let-7e-5p|0|0|1, hsa-let-7i-5p|0|0|1, hsa-mir-30d-5p|0|2|1, hsa-mir-26a-5p|0|0|1, hsa-mir-26a-5p|0|0|2, and hsa-mir-4433b-5p|1|0|1 were decreased in ALS samples compared with controls. An isomiR hsa-mir-10b-5p|0|1|1 was increased in ALS samples compared with controls (eTable 2).

Given that several of the differentially expressed isomiRs in our in-house plasma samples derive from miRNAs that have been described as important to myoblast proliferation, we tested whether there were differences in the levels of the 14 differentially expressed isomiRs between upper and lower motor neuron–predominant ALS patients in our cohort. We dichotomized the 14 samples according to whether their Penn Upper Motor Neuron total was less than or equal to 10 and then used the Wilcoxon rank-sum exact tests to query whether the RPM counts of these isomiRs differed between the 2 groups. None of the comparisons were significant (eTable 3).

Sample Classification Using XGBoost

We subselected 103 isomiRs and 70 miRNAs that had a median RPM count of over 5.0 in all sequencing batches, to ensure that classifier models did not identify batch-specific signal representative of technical artifacts. Zero, 8, and 52 of these isomiRs were differentially expressed in in-house samples, GSE168714, and GSE148097, respectively. Zero, 11, and 28 of these miRNAs were differentially expressed in in-house samples, GSE168714, and GSE148097, respectively.

To train XGBoost models, we first used the limma R package to correct isomiR and miRNA counts according to the sequencing batch. More details are available in the corresponding section in Methods. We then trained XGBoost models on isomiRs or miRNAs separately. Grid search found optimal parameters for max_depth, learning_rate, gamma, and n_estimators, which were [9, 0.1, 0, and 1,000] and [7, 0.2, 0, and 1,000]. The isomiR model accurately classified in-house plasma and public serum samples (AUC = 0.87, positive predictive value [PPV] = 0.74, negative predictive value [NPV] = 1.0) (Figure 3A) and demonstrated the specificity of this isomiR signal in ALS, given poor accuracy in performance to classify AD (AUC = 0.47, PPV = 0.48, NPV = 0.4) or PD (AUC = 0.55, PPV = 0.6, NPV = 0.5) (Figure 3B). By contrast, the miRNA model achieved relatively worse performance on in-house plasma (AUC = 0.49, PPV = 0.5, NPV = 0.77).

Figure 3. XGBoost Accurately Classifies ALS Samples From Holdout Test Sets.

Scatter plots show the true-positive (TPR) and false-positive (FPR) rates at each cutpoint tested for classification of both in-house plasma and publicly available serum ALS and control samples from GSE148097, using 2 separate XGBoost models trained on publicly available plasma samples with (Model 1) values from isomiRs in GSE168714 data (black) and (Model 2) values from miRNAs in GSE168714 data (magenta) (A). Similarly, scatter plots show the true-positive (TPR) and false-positive (FPR) rates at each cutpoint tested for classifying control from AD samples among publicly available plasma in GSE215789 data (green) and control from Parkinson disease (PD) samples from publicly available whole blood in GSE180193 data (purple) using Model 1 that was trained on publicly available plasma from ALS and control samples. ALS = amyotrophic lateral sclerosis.

As a final test, we evaluated isomiR model feature importance scores and manually subselected the top 14 isomiRs by feature importance, to yield a comparable minimal number of isomiRs as were discovered on DE. We then trained a new, third XGBoost model using only these isomiRs to determine whether expression levels of only these isomiRs could adequately classify ALS samples from controls. Performance was comparable on in-house plasma and public serum (AUC = 0.84).

Discussion

In this work, we tested whether there are measurable changes in miRNA profiles in the context of ALS. We sequenced plasma RNA from a cohort of patients with ALS and age-matched and sex-matched healthy controls and then processed data at the level of miRNAs and individual isomiRs (miRNA cleavage variants) to specifically test whether exploration at the isomiR level improves sample classification. We found that isomiR profiles were more diverse in ALS than in control samples, and that disease reference groups of AD and PD did not show similar measurable differences in diversity, demonstrating the potential specificity of our results to ALS. Furthermore, we found that using isomiRs in tree-based machine learning models improved the accuracy of sample classification over miRNAs. In comparison with a recent study that reanalyzed the same publicly available ALS data sets using a distinct but overlapping computational pipeline, we found comparable group-level differences in isomiRs as a function of disease state.³⁵

High-throughput RNA sequencing methods identify a small proportion of the input RNA. As a result, the number of identifiable unique RNA sequences increases as a function of sequencing depth,³⁶ usually up to a plateau above which unique sequences are unlikely to be identified. Across all samples that we analyzed, increasing read count associated with a higher number of identified isomiRs. TDP-43 binds both mature miRNAs⁶ and their precursors.²⁰ Given this, we hypothesized that pathologic mislocalization of TDP, specifically in the context of ALS, would lead to measurable differences in premature miRNA cleavage, thereby resulting in a distinct complement of mature isomiRs with distinct end points.¹⁵ Indeed, there was a significant interaction between read count and condition in a linear model accounting for sequencing batch, denoting that ALS samples at large contained fewer unique isomiRs at any read depth.

To account for potential technical biases related to inherent differences in read depths across data sets, we also tested established indices that allow for comparison of both richness and abundance across samples.³⁷ In a linear model accounting for batch, both indices were significantly different in only ALS samples, indicating greater diversity. miRNA maturation relies on endonucleolytic cleavage by a nuclear Drosha complex and cytoplasmic Dicer.¹⁴ TDP-43 stabilizes the Drosha complex³⁸ and interacts with Dicer.²⁰ TDP-43 mislocalization leads not only to widespread alternative mRNA splicing but also to DE of miRNAs.³⁹ Taken together, these findings suggest that the differences in species diversity observed in both in-house and public data may reflect the sequelae of pathologic TDP-43 mislocalization. Future work will test this hypothesis in vitro using human cell models.

We identified several differentially expressed isomiRs among in-house plasma samples that derive from mature miRNA sequences, which were previously discussed as diagnostic biomarkers for ALS, namely miR-206,^10,11 miR-133a-3p,^10,11 and miR-10b-5p.⁹ Among in-house samples, the following specific isomiRs were differentially expressed between ALS and control samples: hsa-mir-206|0|-2|1, hsa-mir-206|0|0|1, hsa-mir-133a-3p|-1|1|1, hsa-mir-133a-3p|-1|1|2, and hsa-mir-10b-5p|0|0|1. Of note, both the canonical miR-206 species—the isomiR denoted as |0|0|1 and an isomiR lacking 2 bases at the 3′ terminus, |0|0|-2—were differentially expressed among both in-house plasma samples and those from a larger publicly available cohort. miR-206 is a muscle-specific miRNA that functions in the maintenance of endplate synapses.⁴ miR-133a arises from a bicistronic transcript with the miR-1/206 family and promotes myoblast proliferation.⁴⁰ While prior sources have identified miR-133a as differentially decreased in ALS samples compared with controls, we found that an isomiR deriving from this miRNA was differentially increased, potentially reflecting a difference in RNA turnover at the isomiR level in the context of ALS. miR-10b-5p similarly regulates myoblast proliferation.⁴¹ In classification models, an optimal panel also included isomiRs from miR-151a-3p, let-7b-5p, let-7g-5p, let-7i-5p, miR-103a-3p, miR-16-5p, miR-186-5p, miR-27b-3p, miR-30d-5p, miR-30e-5p, miR-744-5p, miR-423-5p, and miR-148a-3p. The miR-30 family, for example, has also been implicated in myoblast differentiation.⁴² Taken together, these results suggest that these isomiRs have relevance to ALS, perhaps biomarking the complex, bidirectional relationship between neuron loss and skeletal muscle turnover.⁴³

Several circulating ALS biomarkers have been proposed, for both diagnosis and guiding prognosis.¹ Blood neurofilament light chain (NfL) may be a nonspecific marker of disease, given that levels are similarly altered in other neurologic disorders.¹ Despite considerable work elucidating the mechanisms behind cryptic exon inclusion induced by pathologic TDP mislocalization, miRNAs remain the only circulating RNA-based biomarkers explored to date, given the inherent rarity and instability of cryptic exons in circulation.⁴⁴ Prior studies have demonstrated strong performance using panels of miRNA biomarkers, including a best reported AUC of 0.98 using an 8 miRNA panel.⁹ In this study, we report comparable performance, including perfect NPV in holdout data sets. Future work will test the performance of this panel in large validation cohorts and in the context of other TDP-associated diseases, for example, limbic-associated TDP-43 encephalopathy and frontotemporal lobar degeneration. A biomarker with high NPV could serve as an ideal complement to NfL.

Our study had several limitations. As in many small RNA-seq experiments, sample sizes in our validation cohorts were small. Although there was no significant difference in age among control and ALS samples, there was a trend toward slightly older age among ALS samples, raising the possibility that RNA degradation affected our results. Prior studies have found that long-term storage at −80C do not significantly affect the integrity of recoverable miRNAs.⁴⁵ Furthermore, information on sequencing batch was limited for several public data sets, limiting our ability to fully account for batch effects in downstream analysis. Future work will seek to overcome this limitation in a multicenter study pooling samples from multiple biobanks. We did not consider nontemplate additions or internal variants, given that these sequences may be more likely than end point variants to represent technical artifacts.²⁸

In summary, we analyzed several small RNA-seq data sets to test the hypotheses that exploration of miRNA profiles at the level of individual isomiRs may improve the performance of RNA-based diagnostic biomarkers. Our work was motivated by a mechanistic hypothesis related to the potential loss of function of wild-type TDP-43 and gain of function of pathologic protein:RNA interactions in the cytoplasm. Future work will explore potential mechanisms related to impaired miRNA metabolism and determine the utility of isomiRs as biomarkers of TDP-related disease activity. Future work with autopsy-confirmed samples will also determine the generalizability of our results to other TDP-43 proteinopathies.

Glossary

AD

Alzheimer disease

ALS

amyotrophic lateral sclerosis

ANOVA

Analysis of Variance

AUC

area under the curve

DE

differential expression

EMG

electromyography

GCC

GNU Compiler Collection

GEO

Gene Expression Omnibus

miRNA

microRNA

mRNA

messenger RNA

NfL

neurofilament light chain

NPV

negative predictive value

PCR

Polymerase Chain Reaction

PD

Parkinson disease

PPV

positive predictive value

RPM

reads per million

TDP

TAR DNA-binding Protein

TDP-43

transactive DNA-response binding protein 43

UMI

unique molecular identifier

UMN

Upper Motor Neuron

Author Contributions

R.G. Magee: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data; study concept or design; analysis or interpretation of data. Vivianna M. Van Deerlin: analysis or interpretation of data. C.T. Mcmillan: drafting/revision of the manuscript for content, including medical writing for content; analysis or interpretation of data EB Lee: major role in the acquisition of data. L.B. Elman: major role in the acquisition of data. D.J. Irwin: drafting/revision of the manuscript for content, including medical writing for content. Z. Mourelatos: drafting/revision of the manuscript for content, including medical writing for content.

Study Funding

This work was supported by the NIH: NINDS (2UE5NS065745, RF1-NS145263) and the National Institute on Aging (P01AG066597, P30-AG-072979). This manuscript is the result of funding in whole or in part by the NIH (NIH). It is subject to the NIH Public Access Policy. Through acceptance of this federal funding, NIH has been given a right to make this manuscript publicly available in PubMed Central upon the Official Date of Publication, as defined by NIH.

Disclosure

The authors report no relevant disclosures. Go to Neurology.org/NG for full disclosures.