Label-Free LC-MS/MS Proteomic Analysis of Cerebrospinal Fluid Identifies Protein/Pathway Alterations and Candidate Biomarkers for Amyotrophic Lateral Sclerosis

Abstract

Analysis of the cerebrospinal fluid (CSF) proteome has proven valuable to the study of neurodegenerative disorders. To identify new protein/pathway alterations and candidate biomarkers for amyotrophic lateral sclerosis (ALS), we performed comparative proteomic profiling of CSF from sporadic ALS (sALS), healthy control (HC), and other neurological disease (OND) subjects using label-free liquid chromatography-tandem mass spectrometry (LC-MS/MS). A total of 1,712 CSF proteins were detected and relatively quantified by spectral counting. Levels of several proteins with diverse biological functions were significantly altered in sALS samples. Enrichment analysis was used to link these alterations to biological pathways, which were predominantly related to inflammation, neuronal activity, and extracellular matrix regulation. We then used our CSF proteomic profiles to create a support vector machines classifier capable of discriminating training set ALS from non-ALS (HC and OND) samples. Four classifier proteins, WD repeat-containing protein 63, amyloid-like protein 1, SPARC-like protein 1, and cell adhesion molecule 3 were identified by feature selection and externally validated. The resultant classifier distinguished ALS from non-ALS samples with 83% sensitivity and 100% specificity in an independent test set. Collectively, our results illustrate the utility of CSF proteomic profiling for identifying ALS protein/pathway alterations and candidate disease biomarkers.

Graphical Abstract

INTRODUCTION

Amyotrophic lateral sclerosis (ALS) is a progressive, fatal neurodegenerative disorder marked by the loss of motor neurons of the corticospinal pathway. ALS is the most common form of adult-onset motor neuron disease, with an overall incidence of approximately 2–3 per 100,000^1–4. The disease is relentlessly progressive, with most patients dying from respiratory failure within 3–5 years of diagnosis^4–8.

At present, diagnosis of ALS is made clinically by use of the symptom-based El Escorial diagnostic criteria. While clinically stringent, the method does not take into account the underlying etiology of the disease. Numerous mechanisms have been proposed to cause motor neuron death in ALS, including oxidative stress, mitochondrial dysfunction, impaired RNA metabolism, protein aggregation, and proteasomal/autophagic dysfunction, among others^9–12. These phenomena occur in a heterogeneous fashion in sporadic ALS cases, as do differences in site of onset and rate of disease progression. This has led to the suggestion that ALS is actually a constellation of disorders of diverse cause united by the death of motor neurons and common symptoms^13,14. Thus, a more accurate understanding of the underlying protein/pathway alterations accompanying sporadic ALS could enhance diagnostic accuracy, identify novel therapeutic targets, provide relevant indices of disease progression/therapeutic response, and aid in stratification of the patient population for research studies and clinical trials.

Proteomic profiling of cerebrospinal fluid (CSF) offers considerable promise towards these ends. In the context of ALS, CSF is a proximal fluid to the site of disease¹⁵, which gives it several advantages as a biomarker source over more distal fluids, such as plasma. One such advantage is its less expansive dynamic range of protein concentrations compared to plasma. By reducing levels of highly abundant proteins such as albumin, sensitive detection and relative quantification of low abundance, disease-associated proteins can be achieved^16,17. CSF and other proximal fluids are also hypothesized to contain a higher concentration of disease-associated proteins as a result of proximity to the diseased tissue microenvironment, from which such proteins may be secreted or released. For example, CSF TDP-43, an RNA binding protein known to form inclusions in neurons in ALS and FTLD-TDP, has been suggested as a possible biomarker for these disorders^18–20. Thus, pathological intracellular changes resulting from disease can be detected externally through proximal fluids, such as CSF. CSF has proven useful for identifying biomarkers for other neurodegenerative diseases, including Alzheimer’s disease²¹, Parkinson’s disease²², and multiple sclerosis²³. CSF-based ALS biomarkers could likewise be used for several purposes, including identifying novel disease mechanisms, diagnostic testing²⁴, and evaluating drug efficacy²⁵.

The ability to detect and quantify the CSF proteome has been aided considerably by the development of label-free mass spectrometric comparative approaches, such as ion abundance and spectral counting^26–28. Several studies have shown that the values produced by these label-free methods reliably correlate with overall protein abundance across orders of magnitude in complex samples. Following appropriate processing (including filtering and normalization²⁷), robust statistical methods can be applied to identify proteins of differential abundance between samples, as well as associated pathways/processes from databases such as Gene Ontology, Reactome, or KEGG^29,30. Moreover, when combined with an appropriate feature selection method³¹, machine-learning approaches can be applied to such datasets to build parsimonious classifiers that, e.g., distinguish healthy from diseased samples based on protein profiles^31–35. Such methods have been used to define the normal human CSF proteome. A recent study identified 2,630 proteins from a group of healthy individuals using immunoaffinity depletion of abundant proteins, multiple liquid chromatography separations, and label-free quantitation³⁶. Moreover, 56% of these proteins were CSF-specific. This study and others suggest that the CSF proteome is relatively stable across time and that inter-subject differences are considerably greater than intra-subject longitudinal differences^36,37, highlighting the utility of CSF for biomarker discovery.

Several previous studies have attempted to discover CSF-based biomarkers of ALS using mass-spectrometric methods^38–42. A limitation of these studies is the small overall number of CSF samples and, consequently, proteins, peptides, or M/Z peaks detected, which can limit downstream data analysis and the detection of low-abundance, disease-associated proteins. The present study, by contrast, benefitted from an extensive patient and control population that dramatically enhanced our ability to identify CSF proteins. Subjects were segmented into groups based on age at symptom onset (< 40, 40–60, or > 60 years old) and, for ALS subjects, site of symptom onset (limb or bulbar). This grouping strategy was selected to reduce non-contributing inter-subject variability, while preserving disease-associated differences. Moreover, pooling of samples was used to maximize the sensitivity of protein detection.

The goals of this study, thus, were two-fold. First, we sought to use LC-MS/MS label-free methods to characterize the CSF proteome in ALS, healthy controls, and other neurologic disease subjects to identify disease-associated alterations in proteins and biological pathways. Second, we sought to identify CSF proteins that could distinguish sALS from both healthy control and other neurological disease subjects by applying feature selection and machine learning methods to our CSF proteomic profiles.

EXPERIMENTAL SECTION

Subjects and CSF Collection

In the first phase of the study, CSF samples from 90 sporadic amyotrophic lateral sclerosis (sALS), 80 healthy control (HC), 20 multiple sclerosis (MS), 20 Alzheimer’s disease (AD), 10 lower motor neuron disease (LMND), 10 upper motor neuron disease (UMND), and 15 familial ALS (fALS) subjects were collected. These samples were pooled into 9 sALS, 8 HC, 2 MS, 2 AD, 1 LMND, 1 UMND, and 2 fALS samples for discovery profiling (training set). Subjects were assigned to a given pool based on age and disease status, as indicated in Table 1, and all pools were gender matched. Each pool contained 200 µL of CSF from each subject in that pool. The MS, AD, LMND, and UMND samples were collectively grouped as “other neurological diseases” (OND, n=6) for subsequent statistical analyses. A separate set of 9 sALS, 7 fALS, and 4 HC CSF samples were individually processed for LC-MS/MS and used in the test set for validation of the classifier generated from our training set of pooled samples described above. The demographics of these individual samples are also described in Table 1.

Table 1.

Subject Demographics. Patient group, diagnosis, class, and age range (years) are indicated. For ALS samples, the site of disease onset and use of riluzole are indicated. The set column indicates which classifier evaluation set a sample was assigned to. Samples are ordered by class then age.

POOLED SAMPLES
Group	Diagnosis	Class	Age Range	Site of Onset	Riluzole	Set
1	sALS	ALS	< 40	Limb	No	Train
2	sALS	ALS	< 40	Limb	No	Train
3	sALS	ALS	40–60	Limb	No	Train
4	sALS	ALS	40–60	Limb	Yes	Train
5	sALS	ALS	40–60	Limb	Yes	Train
6	sALS	ALS	40–60	Bulbar	No	Train
7	sALS	ALS	> 60	Limb	No	Train
8	sALS	ALS	> 60	Limb	Yes	Train
9	sALS	ALS	> 60	Bulbar	No	Train
10	HC	HC	< 40	n/a	n/a	Train
11	HC	HC	< 40	n/a	n/a	Train
12	HC	HC	< 40	n/a	n/a	Train
13	HC	HC	30–60	n/a	n/a	Train
14	HC	HC	40–60	n/a	n/a	Train
15	HC	HC	40–60	n/a	n/a	Train
16	HC	HC	> 60	n/a	n/a	Train
17	HC	HC	> 60	n/a	n/a	Train
18	MS	OND	< 40	n/a	n/a	Test
19	MS	OND	40–70	n/a	n/a	Train
20	LMND	OND	40–65	n/a	n/a	Test
21	UMND	OND	40–65	n/a	n/a	Train
22	AD	OND	> 60	n/a	n/a	Test
23	AD	OND	> 60	n/a	n/a	Train
24	fALS	ALS	30–60	Limb	No	Test
25	fALS	ALS	> 40	Limb	No	Test

INDIVIDUAL SAMPLES
Subject	Diagnosis	Class	Age	Site of Onset	Riluzole	Set
1	sALS	ALS	42	Limb	Yes	Test
2	sALS	ALS	44	Limb	Yes	Test
3	sALS	ALS	44	Limb	Yes	Test
4	sALS	ALS	53	Limb	Yes	Test
5	sALS	ALS	54	Limb	Yes	Test
6	sALS	ALS	55	Limb	Yes	Test
7	sALS	ALS	75	Limb	Yes	Test
8	sALS	ALS	76	Limb	Yes	Test
9	sALS	ALS	77	Limb	Yes	Test
10	fALS	ALS	50	Limb	No	Test
11	fALS	ALS	52	Limb	Yes	Test
12	fALS	ALS	52	Limb	Yes	Test
13	fALS	ALS	53	Limb	No	Test
14	fALS	ALS	60	Limb	No	Test
15	fALS	ALS	62	Limb	No	Test
16	fALS	ALS	62	Limb	Yes	Test
17	HC	HC	33	n/a	n/a	Test
18	HC	HC	50	n/a	n/a	Test
19	HC	HC	51	n/a	n/a	Test
20	HC	HC	65	n/a	n/a	Test

HC = healthy control, MS = multiple sclerosis, LMND = lower motor neuron disease, UMND = upper motor neuron disease, AD = Alzheimer’s disease, OND = other neurological disease.

Revised El Escorial criteria were used to diagnose all ALS subjects, with 18% diagnosed as definite ALS, 33% probable ALS, 24% probable/lab supported ALS, and 25% possible ALS. CSF samples were obtained by lumbar puncture from subjects at either the University of Pittsburgh Medical Center (UPMC) or Massachusetts General Hospital (MGH) upon informed patient consent. The study was approved by both institutional review boards.

Following collection, all samples were spun at 3000 rpm at 4 °C for 10 minutes to remove any cells and debris, mixed by inversion, aliquoted, and stored in protein low bind polypropylene tubes at −80 °C within 2 hours of harvesting. Only CSF samples without visible blood were processed by centrifugation. Hemoglobin levels in all final CSF samples were measured by ELISA to eliminate those with evidence of significant levels of hemoglobin, denoting blood contamination⁴³.

CSF Preparation and Digestion

To enhance the detection of low-abundance CSF proteins, abundant proteins were depleted using the Multi-Affinity Removal System spin cartridge (Agilent; Santa Clara, CA, USA) that removes the 6 most abundant human CSF proteins (albumin, IgG, IgA, haptoglobin, transferrin, and α-1-antitrypsin) according to the manufacturer’s protocol. Depleted samples were buffer exchanged into 50 mM ammonium bicarbonate (NH₄HCO₃) by centrifugation using Amicon Ultra3K columns (Millipore; Darmstadt, Germany) to a final volume of 300 µL.

The samples were reduced with addition of 10 mM DTT at 56 °C for 30 minutes. Samples were then alkylated in 55 mM iodoaceteamide in the dark at room temperature for 30 minutes. Next, 3 µL of 1% ProteaseMAX and Trypsin Gold (Promega; Madison, WI, USA) were added to the samples at a 1:20 ratio and digested at 37 °C for 9 hours. All samples were then de-salted using Pepclean C-18 spin columns (Pierce; Rockford, IL, USA). Peptide digests were eluted with 20 µL 0.1% TFA and 60% ACN by spinning at 1500 g for 1 minute twice. Finally, samples were dried by vacuum centrifugation.

Liquid Chromatography-Tandem Mass Spectrometry

Peptide digests were re-suspended in 0.1% TFA and analyzed in triplicate by nanoflow reversed-phase liquid chromatography tandem mass spectrometry (LC-MS/MS) using a Dionex Ultimate 3000 LC system (Dionex Corporation; Sunnyvale, CA, USA) coupled online to a linear ion trap (LIT) mass spectrometer (LTQ, ThermoFisher Scientific; San Jose, CA, USA). Run orders for each experiment were randomly determined. Separations were performed using 75 µm i.d. × 360 o.d. × 10 cm long fused silica capillary columns (Polymicro Technologies; Phoenix, AZ ,USA) that were slurry packed in house with 5 µm, 300 A pore size C-18 silica-bonded stationary phase (Jupiter, Phenomenex; Torrance, CA, USA). Approximately 1 µg of total peptide digest, as determined by BCA assay, was injected onto a C-18 trap column (Dionex, ThermoFisher Scientific; San Jose, CA, USA), the column was washed for 3 min with mobile phase A (2% acetonitrile, 0.1% formic acid in water) at a flow rate of 30 µL/min. Peptides were eluted using a linear gradient of 0.3% mobile phase B (0.1% formic acid in acetonitrile)/min for 130 minutes, then to 95% B in an additional 10 min, all at a constant flow rate of 0.20 µL/min. Column washing was performed at 95% B for 20 minutes, after which the column was re-equilibrated in mobile phase A prior to subsequent injections. The LIT-MS was operated in a data dependent MS/MS mode in which each full MS scan was followed by five MS/MS scans where the five most abundant peptide molecular ions are selected for collision-induced dissociation (CID), using a normalized collision energy of 30%. Data were collected over a broad mass to charge (m/z) precursor ion selection scan range of m/z 375–1800 with an isolation window of 3 m/z. Dynamic exclusion was used to minimize redundant selection of peptides previously selected for CID with the following settings: repeat count = 1, repeat duration = 30 s, exclusion list size = 500, exclusion duration = 90 s and expiration S/N threshold = 3.

Spectral Counting

Tandem mass spectra were searched against a combined UniProt human protein database from the European Bioinformatics Institute (http://www.ebi.ac.uk/integr8, downloaded 10-05-2010, 58,769 sequences) using the SEQUEST algorithm in BioworksBrowser (v3.31, ThermoFisher Scientific; Waltham, MA, USA). For a fully tryptic peptide to be considered legitimately identified, it had to achieve charge state and proteolytic cleavage-dependent cross correlation (Xcorr) scores of 1.9 for [M+H]1+, 2.2 for [M+2H]2+ and 3.5 for [M+3H]3+, and a minimum delta correlation ΔCn of 0.08. Additionally, peptides were searched for methionine oxidation with a mass addition of 15.9949 and serine, threonine and tyrosine phosphorylation with a mass addition of 79.9663. A false peptide discovery rate less than 2% was determined by searching the primary tandem MS data using the same criteria against a decoy database wherein the protein sequences are reversed⁴⁴. Results were further filtered using software developed in-house, and differences in protein abundance between the samples were derived by summing the total CID events that resulted in a positively identified peptide for a given protein in accession across all samples (spectral counting)^26–28. We further filtered our data by specifying that for a protein to be considered for subsequent statistical analysis, at least 2 unique peptides of the protein had to be detected in at least one pooled sample, in keeping with recent guidelines designed to improve reproducibility in spectral count data analyses⁴⁵.

Statistical Analysis of Relative Protein Abundance

We used the beta binomial test for spectral count data⁴⁶ together with normalization to total spectral counts²⁷ to assess relative levels of identified proteins between groups. To control the false discovery rate (FDR)⁴⁷ of the test, we imputed obtained p values into the R package “Q Value” to generate q values from the associated p values. Bootstrap-based resampling of p values^48–52 was used to correct for multiple testing, with the FDR set at 0.05. To provide further insight into the relative differences in protein abundance between groups, we computed log2 fold difference values for each 2 group comparison (sALS/HC, sALS/OND, OND/HC). A value of “1” was added to each protein’s spectral count value to account for proteins with a zero spectral count before computing the fold difference⁵³. To further control false positives in our analysis, we used a joint q value and fold change criteria to assign statistical significance to a given protein identified in our analysis. Previous work has shown that proteins with high spectral count values tend to produce low fold differences and vice versa^27,45. Accordingly, we used a minimum log2 fold difference of 0.26 for proteins with an average spectral count of 100 or more, a 0.58 log2 fold difference for proteins with an average spectral count of 1 to 100, and a 0.77 log2 fold difference for proteins with an average count of less than 1 to filter our list of proteins with a q value of less than 0.05.

Ontological Enrichment Analysis

After filtering our dataset, we used the program STRAP (version 1.5)^54,55 to characterize the CSF proteome based on the annotations of each identified protein in the Gene Ontology (GO) domains “Biological Process”, “Cellular Complex”, and “Molecular Function”. We next visualized our list of statistically significant proteins in a network layout using Cytoscape⁵⁶. Proteins were visualized first by log2 fold difference values using a red (decreased in x for the x/y fold difference) to green (increased in x for the x/y fold difference) gradient above each protein’s name.

To identify biological pathways associated with our list of differentially abundant proteins, we performed enrichment analysis using ClueGO⁵⁷. The hypergeometric test (with Benjamini-Hochberg multiple testing correction⁴⁷) was used to assess enrichment of categories in the GO domain “Biological Process”⁵⁸. Consistent with recent guidelines for the selection of a reference set^59,60, we used our list of all identified proteins across groups as the reference set for enrichment analysis. In the resultant graph, the proportion of shared proteins associated with a pair of given GO terms was evaluated using the kappa statistic. Pairs of terms (nodes) with a kappa value of at least 0.4 were connected with edges in the network, with the edge thickness reflecting the kappa score value.

Following the identification of over-represented GO Biological Processes, we used Cluepedia⁶¹ to enrich the networks created from these terms. First, we visualized leading terms (GO Biological Process terms with the highest number of statistically significant proteins) together with the proteins associated with that term and the respective fold difference. In the second phase of enrichment, we sought to provide a functional context to the alterations in relative protein levels we identified. To that end, we selected an over-represented GO Biological Process (“Regulation of Extracellular Matrix” [ECM]) and visualized known STRING^62,63 actions (activation, inhibition, expression, and post-translational modification) for statistically significant ECM proteins, other ECM proteins detected in our CSF samples, and several ECM proteins not found in our CSF samples. STRING action scores were used to connect nodes (proteins) in the network. All final figures were made in Adobe Illustrator CS5 (Adobe Systems, Mountain View, CA, USA).

Feature Selection, Classifier Construction, and Validation

In the second phase of the study, we used our CSF proteomic profiles to construct a classifier capable of separating ALS CSF samples from HC and OND CSF samples. The data mining software package Weka⁶⁴ was used for this purpose. The classification task was the binary separation of ALS samples (ALS) from HC and OND (collectively, NON) samples. The training set included all of our pooled sALS and HC CSF proteomic profiles and 3 of the 6 OND proteomic profiles. An independent test set comprised of the remaining 3 OND samples, 9 individual sALS samples, 2 pooled fALS samples, 7 individual fALS samples, and 4 individual HC samples was used for subsequent validation of the classifier. The assignment of samples to the training or test set is indicated in Table 1.

To develop a model resilient to overfitting of the training set, we used a support vector machines algorithm for classification and performed filtering and feature selection on our list of proteins in the training set samples. Filtering was performed by removing any protein from our list if the mean spectral count of any of the classes was zero. Golub’s index (GI), a feature selection method previously used for both disease classification⁶⁵ and feature selection⁶⁶ of spectral count data, was used to select proteins for the classifier. The GI for a protein, i, was calculated as follows:

$${\text{GI}}_i=({\mathrm{\mu }}_i^{\mathit{\text{ALS}}}-{\mathrm{\mu }}_i^{\mathit{\text{NON}}})/({\mathrm{\sigma }}_i^{\mathit{\text{ALS}}}+{\mathrm{\sigma }}_i^{\mathit{\text{NON}}})$$

where ${\mathrm{\mu }}_i^{\mathit{\text{CLASS}}}$ is the mean normalized spectral count for protein i for the indicated class and ${\mathrm{\sigma }}_i^{\mathit{\text{CLASS}}}$ is the standard deviation for protein i for the indicated class. Proteins with a GI ≥ 1.5 were included in the classifier. Weka’s linear support vector machine (SVM) learning algorithm was used to build the classifier. Consistent with prior studies^66,67, we used a complexity parameter, C, of 100 for classifier training and test set evaluation.

To evaluate the performance of the classifier on the training set, we used stratified 10-fold cross validation. We then validated our classifier on the test set described above and in Table 1. To evaluate the performance of the classifier on the test set, we used common measures of predictive performance (ROC curve, sensitivity, specificity, ROC AUC, and F-measure).

Validation of Selected Proteins

To validate our LC-MS/MS results, we used several complementary techniques. Proteins selected for our SVM classifier were validated using Western blotting (WB). Polyacrylamide gel electrophoresis (PAGE), electrophoretic transfer to PVDF membrane, and WB were performed as previously described⁶⁸. Briefly, equal volumes of CSF were loaded in each SDS-PAGE gel lane and total protein by reversible PVDF membrane stain (G-Biosciences; St. Louis, MO, USA) was used as a loading control. For WB experiments, the following antibodies were used: rabbit-anti-WDR63 (Abcam; Cambridge, MA, USA), rabbit-anti-APLP1 (Proteintech; Chicago, IL, USA), rabbit-anti-SPARCL1 (Lifespan Biosciences; Seattle, WA, USA), rabbit-anti-CADM3 (Sigma Aldrich; St. Louis, MO, USA), and rabbit-anti-secretogranin I (Proteintech; Chicago, IL, USA). To permit blot to blot comparisons, Western blot data were normalized using sum total normalization⁶⁹.

ELISAs for complement C3 and cystatin C were performed as previously described^70,71. Immunohistochemistry and light microscopy of spinal cord tissues for tenascin R were performed as previously described⁷². Immunohistochemistry, immunofluorescence, light microscopy, and confocal microscopy of spinal cord tissues for eIF 4e-transporter (4e-T) were performed as previously described⁷².

RESULTS

Global CSF Analysis

Our experimental work flow is illustrated in Figure 1. The demographics of each of our pooled samples are shown in Table 1. Within each of these pooled CSF samples, we identified an average of 6,137 peptides, corresponding to an average of 1,234 unique proteins per sample. Collectively, we identified 1,712 unique proteins across all groups after filtering. (Table S-1). To further understand the composition of the obtained CSF proteome, we grouped these proteins based on their annotations in the Gene Ontology (GO) domains “Biological Process”, “Cellular Complex”, and “Molecular Function”. Figure 2 shows the proportional representation of the various sub-categories in each GO domain with each value expressed as a percentage of the total. Sub-category proportions did not differ substantially between sporadic ALS (sALS), healthy control (HC), and other neurological disease (OND) groups (data not shown). Our results indicate that proteins associated with diverse biological processes, cellular complexes, and molecular functions are found in the CSF proteome. Notably, we detect proteins from numerous intracellular organelles, including the nucleus, mitochondria, ribosomes, and ER in the CSF, consistent with previous work^36,37.

Figure 1.

Experimental Workflow. The process of sample preparation, data acquisition, and data analysis are shown in a flowchart. sALS = sporadic amyotrophic lateral sclerosis, HC = healthy control, OND = other neurological disease.

Figure 2.

Gene Ontology (GO) domain overview of all identified proteins. All identified proteins were input into the three GO domains – Biological Process, Cellular Complex, and Molecular Function – and the resultant terms and percentage of proteins associated with these terms are visualized as pie charts. Term names and percentages are located next to their position on the chart. Percentages correspond to all 1,712 proteins identified across all classes (sALS, HC, and OND).

Statistical Analysis of Relative Protein Abundance

To identify proteins of differential abundance in the CSF between sALS, HC, and OND groups, we performed univariate statistical analysis. Filtering the list of statistically significant proteins from this analysis using a fold difference criteria (see Experimental Section) predominantly had the effect of removing several proteins for which the class average spectral count was less than 1. In total, we identified 123 proteins that met our FDR and fold difference significance criteria (Table S-2). The top 20 proteins of increased and decreased abundance are shown in Table 2. Several proteins with documented associations with ALS were identified in the top 20 protein list, including complement C3, cystatin C, neurofilament medium polypeptide, ephrin type-A receptor 4, and secretogranin 2. Another consistent theme in the top 20 protein list was a decrease in relative levels of extracellular matrix (ECM)-associated proteins in sALS, including tenascin R, semaphorins 7A and 3G, cell adhesion molecule 3, neurexin-3-alpha, agrin, and oligodendrocyte-myelin glycoprotein (Table 2).

Table 2.

The top 20 proteins with statistically significant increased or decreased relative abundance in sALS samples are shown by q value rank (lowest to highest). Protein name, gene name, and Uniprot accession number is shown and protein isoform is indicated in the accession number. The log2 fold difference (FD) for each two group comparison is shown.

INCREASED ABUNDANCE
Protein	Gene Name	Uniprot Accession	q	FD sALS-HC	FD sALS-OND	FD OND-HC
Neurofilament medium polypeptide	NEFM	P07197	<0.0001	2.3	2.4	−0.1
WD repeat-containing protein 63	WDR63	Q8IWG1-1	<0.0001	1.8	2.3	−0.5
Complement C3	C3	P01024	<0.0001	0.3	0.1	0.2
Very large A-kinase anchor protein	CRYBG3	Q68DQ2	0.0004	1.5	1.3	0.2
Acid phosphatase-like protein 2	ACPL2	Q8TE99-1	0.0010	0.9	2.0	−1.1
Serine/threonine-protein kinase PLK1	PLK1	P53350	0.0010	1.2	1.4	−0.2
Protein phosphatase 1J	PPM1J	Q5JR12-1	0.0013	1.4	1.3	0.1
AF4/FMR2 family member 4	AFF4	Q9UHB7-1	0.0041	0.6	1.5	−0.8
PDZ domain-containing RING finger protein 4	PDZRN4	Q6ZMN7-1	0.0044	2.1	1.6	0.5
Actin	ACTB	P60709	0.0044	1.5	0.4	1.1
Disks large-associated protein 1	DLGAP1	O14490-1	0.0050	1.3	0.3	0.9
Tripartite motif-containing protein 43B	TRIM43B	A6NCK2	0.0050	1.2	0.8	0.5
ALK tyrosine kinase receptor	ALK	Q9UM73	0.0056	1.4	1.2	0.2
Protein Daple	CCDC88C	Q9P219-1	0.0056	1.1	0.7	0.5
SMG5	SMG5	Q9UPR3	0.0056	1.0	0.8	0.2
V-set and immunoglobulin domain-containing protein 4	VSIG4	Q9Y279-1	0.0066	1.2	0.7	0.5
Tafazzin	TAZ	Q16635-1	0.0066	0.9	0.7	0.2
Novel protein XXbac-BCX360G3.2-001	BCX360G3.2	B0V386	0.0088	1.1	0.9	0.2
OTU domain-containing protein 7A	OTUD7A	Q8TE49-1	0.0105	0.8	1.2	−0.3
Hyaluronan-binding protein 2	HABP2	Q14520	0.0107	1.1	0.9	0.2

DECREASED ABUNDANCE
Protein	Gene Name	Uniprot Accession	q	FD sALS-HC	FD sALS-OND	FD OND-HC
Multiple epidermal growth factor-like domains protein 8	MEGF8	Q7Z7M0-1	0.0004	−1.1	−0.3	−0.8
Cystatin-C	CST3	P01034	0.0004	−0.1	−0.3	0.1
Golgi membrane protein 1	GOLM1	Q8NBJ4-1	0.0005	−1.4	−0.1	−1.2
Cell adhesion molecule 3	CADM3	Q8N126-1	0.0013	−0.9	−0.9	0.0
Semaphorin-7A	SEMA7A	O75326	0.0038	−0.8	−0.5	−0.3
Semaphorin-3G	SEMA3G	Q9NS98	0.0040	−0.6	1.0	−1.6
Secretogranin-2	SCG2	P13521	0.0040	−0.6	−0.2	−0.4
Neurexin-3-alpha	NRXN3	Q9Y4C0-1	0.0044	−0.7	0.2	−1.0
Putative PRAME family member 24	PRAMEF24	A6NMC2	0.0050	−1.5	−1.1	−0.4
Cartilage acidic protein 1	CRTAC1	Q9NQ79-1	0.0050	−0.3	0.6	−0.9
Oligodendrocyte-myelin glycoprotein	OMG	P23515	0.0058	−0.9	−0.8	−0.2
Disks large homolog 5	DLG5	Q8TDM6-1	0.0064	−1.4	−0.8	−0.6
Testican-2	SPOCK2	Q92563	0.0114	−0.9	0.2	−1.2
Cathepsin D	CTSD	P07339	0.0142	−0.2	−0.7	0.5
Agrin	AGRN	O00468	0.0158	−0.7	−0.4	−0.3
Gelsolin	GSN	P06396-1	0.0165	−0.1	−0.2	0.1
Follistatin-related protein 4	FSTL4	Q6MZW2-1	0.0176	−0.6	0.9	−1.5
Di-N-acetylchitobiase	CTBS	Q01459	0.0183	−1.3	−1.2	−0.1
Delta and Notch-like epidermal growth factor-related receptor	DNER	Q8NFT8	0.0194	−1.4	−0.6	−0.8
Ephrin type-A receptor 4	EPHA4	P54764	0.0202	−0.6	−0.1	−0.5

Ontological Enrichment Analysis

To identify biological pathways associated with our list of differentially abundant CSF proteins, we performed enrichment analysis in the Gene Ontology (GO) Biological Process domain. We found 27 enriched GO terms that were over-represented in our list of statistically significant proteins. These results are shown in Figure 3. Several themes emerged from our analysis of overrepresented biological processes. We identified many neuron-specific processes, including “synapse organization”, “regulation of axonogenesis”, “regulation of synaptic plasticity”, and “fasciculation of sensory neuron axon”. Of note, all proteins comprising the term “fasciculation of sensory neuron axon” (ephrin type A receptors 3 and 4 and multiple epidermal growth factor-like domains protein) were significantly altered in sALS samples compared to healthy control samples. Other major processes identified from the analysis included “regulation of extracellular matrix”, “acute inflammatory response”, and “glial cell differentiation”, each of which has clear relevance to ALS.

Figure 3.

GO Biological Process enrichment analysis. Proteins whose levels were significantly different between sALS, HC, and OND groups were used to identify over-represented GO Biological Process terms. The results are shown in a network view where node size corresponds to the term’s p value and edge thickness corresponds to kappa score value as indicated in the legend at right. Nodes are positioned for ease of interpretation and edge length is arbitrary. Leading terms, GO terms with the highest number of included proteins, are shown in a larger, colored font.

Taking those biological processes with the highest number of associated proteins, we next visualized these terms and their associated proteins together in a network view (Figure 4). A protein’s association with a term is indicated by an edge connecting the two. The relative contributions of proteins of increased and decreased abundance in sALS can be seen in this view. For example, the identification of the terms “acute inflammatory response” and “regulation of inflammatory response” is mainly due to increased levels of proteins associated with these terms in sALS CSF samples, particularly those proteins of the complement pathway. Likewise, the enrichment of the terms “synapse organization” and “extracellular matrix organization” stems primarily from the decreased levels of proteins associated with these terms in sALS samples. Collectively, this network view underscores the notion that the ALS disease process is associated with a concomitant decrease in synaptic and ECM proteins and increase in inflammation-related proteins within the CSF. We also note that proteins with a high degree of connectivity in the network of Figure 4 (i.e., those associated with several terms) tended to be ECM-associated proteins, such as tenascin R and laminin subunit alpha 2 (LAMA2), providing further evidence that ECM alterations are a preponderant pathological phenomenon in sALS (Figure 4).

Figure 4.

Leading terms from Figure 3 are shown with their associated proteins. Each two-group log2 fold difference (sALS/HC, sALS/OND, OND/HC) is visualized using a red to green gradient. Proteins with a log2 fold difference > 0.58 for sALS relative to both HC and OND groups are emphasized as indicated, as are externally validated proteins.

To provide functional context and a pathway-level understanding of the observed ECM alterations, we linked ECM proteins using STRING action scores to create a network of the “regulation of extracellular matrix” GO term. We note that not all proteins from the term are shown to enhance the interpretability of the figure. The final network is shown in Figure 5. Several findings emerge from this analysis. First, we provide clear evidence of the power of LC-MS/MS CSF proteomic profiling to identify ECM-associated proteins. Of the 102 proteins shown in Figure 5, 85 could be detected in at least one of our pooled CSF samples (those that could not are bracketed with parentheses in Figure 5). Second, proteins whose levels are significantly altered in sALS CSF converge on few common targets, suggesting that pathological ECM protein and signaling alterations are widespread in ALS. Lastly, we again note that the majority of altered ECM proteins are decreased in sALS samples. The network view thus emphasizes that these decrements may have widespread signaling effects, particularly for proteins such as gelsolin (GSN) that act on numerous ECM proteins.

Figure 5.

Enrichment of “Regulation of Extracellular Matrix” GO Term. Statistically significant proteins from the GO Biological Process term “Regulation of Extracellular Matrix” were visualized as circular nodes, with log2 fold difference (on a red to green gradient) shown in the three squares above each node. Proteins with a log2 fold difference > 0.58 for sALS relative to both HC and OND groups are emphasized as indicated, as are externally validated proteins. The network was enriched using STRING action scores (activation, inhibition, expression, and post-translational modification; colored as indicated) to add associated proteins to the network. Statistically significant protein names from the LC-MS/MS data are shown in red font for emphasis. Names of proteins not detected in our LC-MS/MS analysis are bracketed by parentheses. Edge width reflects the strength of the STRING action score as indicated in the legend and edge length is arbitrary.

Validation of Known ALS Biomarkers

We observed significant differences in a number of proteins previously identified as candidate ALS CSF biomarkers by LC-MS/MS. To further evaluate the utility of these proteins as CSF biomarkers, we performed ELISAs and Western blots on individual CSF samples from subjects comprising our ALS, HC, and OND pooled samples. The results of these experiments are shown in Figure 6. Complement C3 levels were measured by ELISA in 23 sALS, 26 HC, and 12 OND CSF samples. Statistically significant increases in CSF C3 levels were observed when comparing sALS or OND samples to HC samples (mean sALS = 3.42 ± 0.29 µg/mL, mean OND = 3.40 ± 0.37 µg/mL, and mean HC = 2.22 ± 0.12 µg/mL, p < 0.001 for the sALS/HC comparison, p < 0.01 for the OND/HC comparison). We measured cystatin C levels by ELISA in 20 sALS samples, 8 HC samples, and 11 OND samples. A statistically significant decrease in cystatin c levels was observed in sALS CSF relative to HC (means = 3.27 ± 0.34 µg/mL and 4.98 ± 0.36 µg/mL, respectively, p < 0.01). A non-significant decrease in cystatin C levels was seen when comparing sALS samples to OND samples (means = 3.27 ± 0.34 µg/mL and 4.26 ± 0.35 µg/mL, respectively, p > 0.05). Levels of secretogranin I were measured by Western blotting of 14 sALS, 10 HC, and 10 OND CSF samples. Secretogranin I was significantly decreased in sALS compared to HC (p < 0.01) and OND (p < 0.01) samples.

Figure 6.

Validation of Known ALS Biomarkers. A. Complement C3 levels in CSF from sALS, HC, and OND subjects were measured by ELISA in triplicate. The concentration of Complement C3 (µg/mL ± SEM) is indicated on the y axis. B. Cystatin C levels in CSF from sALS, HC, and OND subjects were measured by ELISA. The concentration of cystatin C (µg/mL ± SEM) is indicated on the y axis. C. Relative levels of secretogranin I in CSF from sALS, HC, and OND subjects were measured by Western blot. The mean normalized integrated density value ± SEM is indicated on the y axis. For A–C. * = p < 0.001 for the indicated comparison, # = p < 0.01 for the indicated comparison.

Validation of Protein Alterations in ALS Spinal Cord Tissue

We previously demonstrated that altered levels of CSF proteins in sALS can be used to identify pathological alterations in ALS motor neurons⁷². To extend this approach to the current study, we validated select proteins from our list of differentially abundant proteins using immunohistochemistry and immmunofluorescence. Because decreases in ECM proteins were a consistent theme in our results, we stained 4 sALS and 4 HC lumbar spinal cord tissue sections with anti-tenascin R antibody. Tenascin R is a component of perineuronal nets, the densely organized extracellular matrix surrounding neurons. We observed a clear loss of tenascin R immunoreactivity surrounding motor neurons in sALS lumbar spinal cord that was not seen in HC subjects (Figure S-1). Thus, the ECM of the ALS-afflicted motor neurons shows perturbations consistent with the results of LC-MS/MS proteomic profiling of CSF.

LC-MS/MS also identified a significant increase in levels of the eIF 4e transporter (4e-T) in the CSF of sALS patients when compared to HC. 4e-T is a component of stress granules⁷⁴ and we therefore performed tissue staining to determine if 4e-T is a component of inclusions in sALS spinal cord motor neurons, as has been observed for other stress granule-associated proteins⁷⁵. We observed filamentous and granular nuclear 4e-T staining and cytoplasmic 4e-T-positive inclusions in lumbar spinal cord motor neurons in the 4 sALS cases examined (Figure S-2). By contrast, the 4e-T staining in motor neurons of all 4 HC subjects was diffuse, non-filamentous, and nuclear. Immunofluorescence microscopy was then used to determine if 4e-T inclusions were p62 positive. We observed considerable accumulation of autofluorescent cytoplasmic lipofuscin in motor neurons (Figure S-2; arrowheads) that obfuscated the detection of cytoplasmic 4e-T. This accumulation was not present in the nucleus, however, and numerous nuclear, p62-positive 4e-T granules in sALS spinal cord motor neurons not seen in motor neurons of HC subjects are evident (Figure S-2; arrows).

Classifier Construction and Machine Learning

In the final phase of the study, we used our pooled sample CSF proteomic profiles from the training set to build a support vector machine (SVM) classifier capable of distinguishing ALS samples from non-ALS samples in an independent test set. We first used filtering to remove proteins for which any class mean spectral count was 0, as such proteins could lead to an overfitted classifier and, consequently, poor validation performance on separate test sets. Moreover, low CSF peptide counts may reflect low protein levels that can lead to an inability to validate protein levels via other methodologies.

Feature selection was used to identify proteins that exhibited large alterations between ALS and HC/OND (collectively, NON) samples. Four proteins were selected for training of the classifier: WD repeat-containing protein 63 (WDR63), amyloid-like protein 1 (APLP1), SPARC-like protein 1 (SPARCL1), and cell adhesion molecule 3 (CADM3). WDR63 was increased in the CSF of ALS samples, while the remaining three proteins were decreased in sALS samples relative to NON samples.

These four proteins were used as features for building a linear SVM classifier using the training set samples. For the initial evaluation of the classifier, we used stratified 10-fold cross validation. All training set samples were correctly classified in the cross validation. An independent test set consisting of individual CSF samples analyzed by LC-MS/MS was used to further validate the classifier. Several familial ALS (fALS) samples were included to increase the size of the test set, as indicated in Table 1. While etiologically distinct from sALS subjects, we postulated that motor neuron death may nevertheless produce changes in CSF protein levels that are common to both sALS and fALS. The performance metrics of the test set evaluation and an ROC curve are shown in Figure 7. The 4 protein classifier achieved 83% sensitivity and 100% specificity with the test set, misclassifying one sALS sample and two fALS samples. Inspection of the misclassified samples showed that relative levels of WDR63 were lower and relative levels of APLP1 were higher in these samples than in the other ALS test set samples.

Figure 7.

Classifier Performance. A. ROC curve showing the performance of the support vector machines classifier on an independent test set. The various levels of the decision threshold are shown as diamonds along the curve. B. Confusion matrix, classifier errors, and the indicated performance metrics are shown for the application of the classifier to an independent test set.

We then validated the levels of our four classifier proteins by Western blot (WB) in select individual CSF samples from the training set samples. The results of these experiments are shown in Figure 8. WB experiments were performed on sALS (n=15) samples, HC samples (n=12), and OND (n=10) samples. Relative levels of WDR63 were increased in sALS CSF compared to HC and OND samples (p < 0.001 for both comparisons). Two adjacent bands for WDR63 were observed and the results reflect levels of total WDR63 in the CSF. Relative levels of APLP1 were decreased in the CSF of sALS patients relative to HC (p < 0.001) and OND (p < 0.01) samples. Relative levels of SPARCL1 were decreased in the CSF of sALS patients relative to HC (p < 0.01) and OND (p < 0.05) samples. Relative levels of CADM3 were decreased in the CSF of sALS patients relative to HC and OND samples (p < 0.001 for both comparisons).

Figure 8.

Validation of Classifier Proteins. Relative levels of classifier proteins in CSF from training set sALS, HC, and OND subjects were measured by Western blot. The mean normalized integrated density value ± SEM is indicated on the y axis. A. WDR63. B. APLP1. C. SPARCL1. D. CADM3. For A–D, * = p < 0.001 for the indicated comparison, # = p < 0.01 for the indicated comparison, and § = p < 0.05 for the indicated comparison.

DISCUSSION

The goals of this study were two-fold. First, we sought to use LC-MS/MS to construct comprehensive CSF proteomic profiles of sALS, HC, and OND samples to identify ALS protein/pathway alterations. Second, we sought to build a classifier capable of distinguishing ALS from HC and OND samples using our CSF proteomic profiles. With regard to our first aim, we identified 1,712 proteins in the CSF across all groups. This number is consistent with recent studies characterizing the normal CSF proteome^36,37. Similarly, our data sets show strong concordance in the proportion of all identified proteins assigned to the various Biological Process, Cellular Complex, and Molecular Function Gene Ontology domains, suggesting that the overall composition of the CSF proteome is relatively stable, despite multiple daily turnovers of the total CSF volume^{76, 77}. This, together with CSF’s proximity to the CNS tissue microenvironment, makes it an ideal biomarker source for neurological disorders such as ALS. The results also illustrate the considerable sensitivity that is achieved by pooling samples and using label-free relative quantitation, a finding that has been shown in other studies characterizing the CSF proteome^36,37.

The sensitivity of our LC-MS/MS method enabled the detection of differences in relative levels of many proteins between sALS and non-sALS samples. Several of these proteins have documented associations with ALS. For example, we detected and validated alterations in the previously described candidate ALS CSF biomarkers complement C3⁷⁰, cystatin C^63,78, and secretogranin I⁷⁹. We also demonstrate decreases in the known ALS disease modifier⁸⁰, ephrin type A receptor 4 (EphA4) in the CSF of ALS subjects. Reducing EphA4 expression or signaling prolongs survival in ALS model systems⁸⁰, so the decrease we observe in ALS CSF may reflect a compensatory response by the CNS to the ALS disease process.

We used enrichment analysis to identify biological processes and pathways altered in ALS. The majority of over-represented processes were related to inflammation, synaptic activity, cell growth, or extracellular matrix (ECM) regulation. The latter three processes are similar to those recently identified using microarray analysis of gene expression in oculomotor and spinal motor neurons⁸¹, while elevations in inflammation-associated proteins, specifically those of the complement pathway, have been previously associated with human ALS^{70, 82–84} and ALS model systems^{85, 86}. We detect elevations in several proteins of the complement pathway, confirming previously-described elevations of CSF C3^70,82 and providing new evidence for an increase in C5 (which is upregulated in animal models of ALS^{85, 87}) and C2, as well as decreased complement factor I. Our results support a role of elevated complement pathway activation as a contributing factor to motor neuron death and a potential therapeutic target. These complement pathway proteins may also be useful biomarkers to evaluate the effectiveness of drugs that target this pathway.

Aberrant synaptic changes are also well-described in human ALS and ALS model systems. Axonal dying-back is recognized both as a pathologic feature and hypothesized causal mechanism in ALS^88,89 and degenerative structural changes are observed at synapses of motor neurons and neuromuscular junctions (NMJ) in ALS spinal cord tissue^90,91. We now identify candidate synaptic proteins that contribute to these pathological changes. For example, decreases in semaphorins 7A and 3G were observed in sALS CSF. The semaphorins act as axonal guidance molecules and have previously been linked to ALS^92,93. They are expressed in a variety of cell types, including neurons, glia, regulatory T-cells, and vascular epithelial cells. Probing these cell types for altered levels of semaphorins and other synaptic proteins in ALS tissues or model systems may provide insights into how cell type-specific altered synaptic protein expression leads to changes in synapse structure and function that lead to or contribute to motor neuron degeneration in ALS.

Many of the synaptic proteins that were altered in sALS samples are also components of the extracellular matrix (ECM). Pathological ECM alterations have been described for ALS⁹⁴ and ALS model systems⁹⁵. Our results showed consistent decreases in levels of ECM proteins, including tenascin R, agrin, and cell adhesion molecule 3. These observations point to a loss of integrity of perineuronal nets (PNN), the highly specialized ECM surrounding neurons. The functions of PNNs are diverse and include physically surrounding neurons, protecting against harmful external agents⁹⁶, influencing synaptic transmission⁹⁷, and buffering against oxidative stress⁹⁸. Validation of the decrease in tenascin R seen in CSF by immunohistochemistry of spinal cord tissues further establishes altered morphology and protein composition of PNNs as a pathological phenomenon in ALS, consistent with findings on TNR in ALS animal models⁹⁵. As the alterations we observed were localized to motor neurons, we propose that pathological PNN alterations contribute to selective MN vulnerability in ALS.

An implicit and overarching assumption of CSF proteomic profiling is the idea that protein alterations detected in CSF can provide evidence of intracellular changes resulting from the ALS disease process⁴². We showed previously that RBM45, an RNA binding protein whose levels are increased in the CSF of ALS patients⁷², forms cytoplasmic inclusions in motor neurons and glia of ALS patients. Similarly, we now show that increased levels of the stress-granule associated⁷⁴ eIF4E transporter protein (4e-T) in the CSF of sALS patients correlate with its presence in p62-positive nuclear granules and cytoplasmic inclusions (Figure S-2). Intranuclear inclusions are a pathological feature of sALS^99,100 and we demonstrate the presence of a transport protein, 4e-T, in these inclusions. Moreover, while p62-positive intranuclear inclusions have been described in c9ORF72-linked familial ALS tissue¹⁰¹, we demonstrate the presence of such inclusions for the first time in sporadic ALS patients.

Our second major objective was to build a discriminant classifier capable of distinguishing ALS patients from healthy controls and other neurodegenerative disease patients on the basis of the CSF proteomic profiles of each group. We⁷⁰ and others¹⁰² have shown that the CSF protein levels and machine learning can be used to distinguish sALS and healthy control subjects. Whether this approach is also feasible for the simultaneous separation of sALS from HC and OND samples was unclear. As is typical for high-dimensional –omics data sets, we were able to build a classifier capable of correctly classifying all training set samples. Validation on an independent test set composed of sALS, fALS, HC, and OND samples demonstrated that the classifier generalized well, achieving 83% sensitivity and 100% specificity.

Given the considerable clinical and pathological heterogeneity of ALS^9–12, no biomarker (or panel of biomarkers) is likely to achieve 100% sensitivity and specificity when separating ALS from both HC and OND samples. Because ALS is a relatively rare disorder, the specificity of a classifier is of paramount importance. In this regard, our set of classifier proteins shows considerable promise, correctly classifying all HC and OND samples. The misclassified ALS samples had lower WDR63 and higher APLP1 levels than the other ALS samples profiled. Determining the relationship between WDR63 and APLP1 CSF levels and ALS disease mechanisms and clinical parameters, thus, is an important area of future research. Similar to other LC-MS/MS studies⁶⁶, the sample size of both the training and test sets is small relative to the number of proteins identified. Future studies are thus needed to confirm the predictive ability of this biomarker panel with a larger set of samples.

Previously identified CSF protein biomarkers for ALS have often been of increased abundance in ALS, such as complement C3 and neurofilament proteins⁷⁰. Markers of inflammation or neuronal injury will correctly separate ALS from HC samples. They are, however, less likely to correctly distinguish ALS from OND samples (Figure 6a), which is also important for biomarker-based disease classification. To account for this, we grouped HC and OND samples together as NON (i.e., non-ALS) samples prior to feature selection. Proteins identified by this approach may include more specific markers of ALS-associated loss of motor neuron synaptic integrity or degeneration. The roles of APLP1 and CADM3 in neuromuscular junction (NMJ) function discussed below provide support for this notion.

Classification based on decreased protein abundance creates the potential for misclassification in subsequent studies due to protein instability or differences in analytical sensitivity. For this reason, future studies are required to evaluate the longitudinal stability of the decreases we observed in our classifier proteins. Likewise, evaluating the clinical stage at which these alterations become apparent has considerable implications for the diagnostic/prognostic value of these markers. From an analytical perspective, our ability to detect all proteins of the classifier in all of our individual samples profiled by Western blot suggests that these markers are readily detectable in CSF. They are thus likely amenable to future measurement by quantitative approaches, such as ELISA, which may further enhance measurement sensitivity and classification accuracy.

The four proteins used in our classifier have plausible connections to motor neuron degeneration. Three of the proteins used in the classifier, amyloid-like protein 1 (APLP1), SPARC-like protein 1 (SPARCL1), and cell adhesion molecule 3 (CADM3) were decreased in the CSF of ALS patients. Each protein has documented associations with neurons and synaptic activity, making them promising candidate biomarkers for ALS. APLP1 associates with NMDA receptors and regulates surface expression of this family of glutamatergic receptors¹⁰³. This regulation is crucial for maintaining cellular homeostasis and synaptic activity. Knockout of APLP1 and APLP2 (also significantly reduced in our ALS samples) results in reduced pre- and post-synaptic compartment size at the NMJ¹⁰⁴. Reduced levels of APLP1 in ALS, therefore, may contribute to muscle de-innervation and axonal die-back by altering the integrity of the NMJ.

SPARCL1 belongs to the BL-40 family of ECM proteins¹⁰⁵. The protein is secreted by astrocytes and promotes synapse formation¹⁰⁶. Knockout of the SPARCL1 gene in mice resulted in a decrease in the number of excitatory synapses in the superior colliculus¹⁰⁷. In addition to synapse formation, the protein may also be essential for synaptic maintenance. If so, reduced levels of SPARCL1 could result in decreases in synaptic activity that ultimately contributes to motor neuron degeneration. SPARCL1 also binds extracellular calcium¹⁰⁸. Reduced astrocytic or neuronal SPARCL1 may then also promote calcium dyshomeostasis, which could likewise contribute to motor neuron vulnerability.

CADM3 (or nectin-like molecule 1 or synCAM3) is a cell junction protein that localizes to synapses¹¹⁰. The protein is an immunoglobulin-like molecule that is enriched in the nervous system and regulates cell-cell contacts and synapse formation¹⁰⁹. In keeping with this role, developmental patterns of CADM3 expression are observed in the nervous system of multiple organisms¹⁰⁹. The protein also has a role in the formation of functional NMJs¹¹⁰. Decreased levels of CADM3, thus, are predicted to impair synaptic maintenance, produce neuromuscular junction impairments, and reduce ECM integrity. Cumulatively, CADM3-induced alteration of any or all of these processes could make motor neurons susceptible to degeneration.

At present, studies characterizing the function of WDR63, the last of our classifier proteins, have not been performed. Unlike APLP1, SPARCL1, and CADM3, levels of WDR63 were elevated in ALS samples relative to HC and OND. This specificity for ALS makes it a promising biomarker though a hypothesized role in motor neuron degeneration or ALS more generally is difficult to predict. The WD-repeat is a common structural motif and many diverse functions, including signal transduction, mRNA synthesis, and cytoskeletal assembly, among others, can be assigned to proteins containing WD repeats^111,112. Thus, further research is necessary to determine a functional connection between WDR63 and ALS.

In conclusion, we constructed LC-MS/MS proteomic profiles of sALS, HC, and OND CSF samples. We used these profiles to identify proteins whose levels are significantly altered in sALS samples relative to HC and OND samples. In doing so, we identified several proteins with documented associations with ALS, as well as several new candidate biomarkers with clear biological relevance to motor neuron degeneration. Using ontological analysis, we described several biological pathways that are altered in ALS. Lastly, we used an SVM learning algorithm to build a classifier capable of separating ALS samples from HC and OND samples.

Collectively, our results illustrate the utility of label-free LC-MS/MS proteomic methods, the promise of CSF as a biomarker source, and the applicability of machine learning methods to classifying samples based on mass spectrometric-based proteomic profiles.

Supplementary Material

Supporting Information

Figure S-1. Tenascin R Validation Data. Lumbar spinal cord sections from sALS and HC patients were stained using a polyclonal anti-tenascin R antibody. Top Panel. Motor neurons from HC subjects show strong tenascin R immunoreactivity around the cell body. Bottom panel. Motor neurons from sALS cases show an absence of this staining pattern. All images were acquired with a 40× objective and the scale bar = 10 µm.

Figure S-2. eIF 4e Transporter (4e-T) Validation Data. Lumbar spinal cord sections from sALS and HC patients were stained using a polyclonal 4e-T antibody. Top panels. Light microscopy was used to visualize 4e-T staining in the lumbar spinal cord of sALS and HC subjects. Staining in HC subjects is diffuse throughout the nucleus. In contrast, sALS cases show nuclear and cytoplasmic (arrows) 4e-T inclusions and a loss of diffuse staining. All images were acquired with a 40× objective and the scale bar = 10 µm. Bottom panels. Immunofluorescence was used to further characterize 4e-T alterations in sALS. sALS 4e-T nuclear inclusion staining overlaps with p62 staining (arrows). Nuclear 4e-T staining in HC is diffuse and no nuclear p62 is observed (arrows). The presence of cytoplasmic autofluorescent lipofuscin is prominent in both cases and is indicated by arrowheads. All images were acquired with a 63× objective and the scale bar = 10 µm.

Table S-1

Table S-1. Global Results. The raw, unfiltered data set comprising all spectral counts for all groups prior to normalization is shown. Protein name and Uniprot accession number is shown and protein isoform is indicated in the accession number. HC = healthy control, OND = other neurological disease, MS = multiple sclerosis, UMND = upper motor neuron disease, LMND = lower motor neuron disease, AD = Alzheimer’s disease.

Table S-2

Table S-2. All Differentially Abundant Proteins. All proteins of significantly increased or decreased relative abundance in sALS samples are shown by q value rank (lowest to highest). Protein name, gene name, and Uniprot accession number is shown and protein isoform is indicated in the accession number. Log2 fold difference (FD) for each two group comparison is shown.

ABBREVIATIONS

ALS

amyotrophic lateral sclerosis

sALS

sporadic amyotrophic lateral sclerosis

fALS

familial amyotrophic lateral sclerosis

HC

healthy control

OND

other neurological disease

MS

multiple sclerosis

LMND

lower motor neuron disease

UMND

upper motor neuron disease

AD

Alzheimer’s disease

ROC AUC

Area Under Receiver Operating Characteristic Curve