### Ethics and study design
This study was a cross-sectional analysis of plasma proteomic data to identify a biomarker panel for ALS. Samples were obtained from participants from the United States and Italy. recruitment sites were located at the University of Turin in Turin, Italy; the NIH in Bethesda and Baltimore, Maryland; and Johns Hopkins University in Baltimore, Maryland.
Written consent was obtained from all individuals enrolled in this study. The institutional review boards of the National Institute on Aging (NIA) (protocol numbers 03-AG-0325 and 03-AG-N329); the National Institute of Neurological Disorders and Stroke (01-N-0206, 13-N-0188 and 17-N-0131); the National Institute of Allergy and Infectious Diseases (NIAID) (09-I-0032); Johns Hopkins University (00173663); and the University of Turin (004462) approved the study.
### Participants
From September 2008 through Febuary 2023, 281 patients with ALS and 258 healthy individuals were enrolled in the study at the University of Turin and the NIH. The Italian samples consisted of neurologically healthy individuals (n = 196) and patients diagnosed with ALS (n = 236) living in Northern Italy and recruited in a population-based study known as the Piedmont and valle d’Aosta Registry for ALS (PARALS; established 1 January 1995)9. The registry’s near-complete case ascertainment of ALS among its catchment population of almost 4.5 million inhabitants ensures the applicability of our findings9. The US plasma samples comprised patients diagnosed with ALS (n = 45) evaluated at the NIH Clinical Center in Bethesda, Maryland, as part of a natural history study (NCT03225144)47.### Sample collection
The plasma samples were collected via phlebotomy of the upper limb. The patients were not fasting. The italian samples (n = 427) were mainly collected using heparin tubes, and the remaining samples (n = 306) were collected using EDTA tubes. Blood cells were removed by centrifugation within 2 hours of collection, and the supernatant, consisting of the plasma, was carefully removed without disturbing the cell pellet. All plasma samples were aliquoted and stored at −80 °C.The number of freeze-thaw cycles between aliquoting and running the proteomic assays was minimized, typically one but no more than three.
The CSF samples (n = 14) were collected via lumbar puncture with the patient sitting.
Materials and Methods: Data Acquisition and Processing
Proteomic Data Acquisition & Quality Control
Proteomic profiling was performed using the Olink Proteomics platform. Replicate samples (n=3 per plate) were included as bridging samples to calculate intra- and inter-plate coefficients of variation. These bridging samples were independent of the inter-plate sample controls.Coefficients of variation were calculated using the formula: (s.d./mean of the replicate measurements) × 100. Samples with median NPX values (calculated from all proteins in each sample) deviating by more than 3 standard deviations (s.d.) from the mean of the median NPX values of all samples, or exceeding 3 s.d. from the mean of the interquartile range, were identified as outliers and excluded from the analysis. Olink data from the UK Biobank was accessed through the UK Biobank Research Analysis Platform.
Case and Control Definition
ALS cases were defined as individuals with a diagnosis of ICD-10 code G12.2, confirmed at death, and with blood collected up to 1 year before symptom onset. Control samples consisted of individuals without a recorded diagnosis of myopathy (ICD-10 codes G70-G73) or neuropathy (ICD-10 codes G60-G64).
Genetic Data Generation
Whole-genome sequencing was conducted on a HiSeq X10 sequencer (Illumina) for 476 participants (150-bp paired-end reads, 35× coverage). Genotyping was performed on infinium GDA-8+ NeuroBooster BeadChips (version 1.0, Illumina) for 245 participants, following manufacturer’s instructions. Standard sample-level and variant-level quality control procedures were applied to the genetic datasets. Variants were extracted for principal component analysis, performed using ‘flashPCA’ (version 2.0). Principal components 1-10 were reduced to two dimensions using UMAP version 0.2.10, and these UMAP values were used to correct for population stratification in the statistical analysis of individual protein analytes.## Statistical Analysis
In our analysis, a linear regression model was employed to identify differentially abundant proteins between ALS cases and controls. This model was implemented using the ‘limma’ package (version 3.58.1) in R. All covariates were included in the model simultaneously. Consistent with our primary analysis, the regression model for patients carrying the C9orf72 expansion was adjusted for sex, age at sample collection, the type of plasma collection tube and the frist two dimensions of UMAP based on genetic data.
Multiple comparisons were controlled using the FDR procedure defined by Benjamini and Hochberg, with two-sided FDR-adjusted P values reported in Table 1, Fig. 2 and Extended Data Table 2. The significance threshold was set at an FDR P value of 0.05.The Discovery Cohort, comprising 183 cases and 309 controls, had 80% power to detect a differentially abundant protein with an effect size (Cohen’s f statistic) of 0.242, assuming a significance threshold P value of 1.73 × 10−5. All analyses were conducted using R (version 4.3.2).We used z-scores to compare the effect sizes of differentially abundant proteins in the Discovery and Replication cohorts. z-scores were chosen as they provide an accurate statistical depiction of the difference in effect size between cases and controls by accounting for the s.e. of the effect size58. Without standardization with z-scores, the slope of the relationship between protein abundance and disease status could be artificially inflated or deflated, resulting in misleading correlation statistics.### Comparison of Olink and ELISA for differentially abundant proteins
Quantification was performed on plasma samples (n = 16 ALS cases and n = 16 healthy controls) using commercially available colorimetric ELISA or fluorescence-based ProQuantum high-sensitivity immunoassay kits according to the manufacturer’s protocol. Immunoassay kits were available for 30 of the 33 differentially abundant proteins (Extended Data Fig. 1). The optimal dilution factor for measuring each protein was steadfast empirically. Each plasma sample and protein standards were assayed in duplicates.A standard curve for each protein was generated by plotting the absorbance.### Comparison of protein in CSF and plasma
CSF proteomic measurements were generated from ALS cases (n = 14) and healthy control individuals (n = 89) using the SomaScan 7K platform. Most of the ALS cases (n = 13) and none of the healthy control individuals used to generate the SomaScan data were not included in the Discovery Cohort or the Replication Cohort used to create the plasma proteomic data. Individual proteins were assessed using generalized linear modeling, adjusted for age at collection and sex.
### Pathway analysis
The proteins that were differentially abundant in ALS were analyzed for pathway enrichment using ‘clusterProfiler’ (version 4.10.1) using the following databases: (1) GO11(2) KEGG12 and (3) Reactome13.
Supervised Machine Learning
We used supervised machine learning to find a molecular signature of ALS. We randomly chose 80% of the samples for the Training Set (n = 183 ALS cases versus n = 172 healthy controls plus n = 137 other neurological diseases). The remaining samples were the Testing Set (n = 48 ALS cases versus n = 42 healthy controls plus n = 33 other neurological diseases; see fig. 1 for a flowchart).The samples in the training Set and Testing set were the same as those used for the protein abundance analysis.
We used ‘caret’ (short for ‘classification and regression training’, version 6.0-94), a free, open-source package that simplifies building predictive models. We entered 33 proteins that showed differences, along with clinical data like age, sex, and the type of blood tube used for collection (total of 36 features) into the model.
We limited the features to 33 proteins to make the model more stable and prevent overfitting, which can happen when analyzing complex data65. Using fewer features also makes the model more practical for use in clinics, where cost and testing complexity are vital. For example, while UMAP components are helpful for analyzing population differences, we left them out of the model. Including them would require genetic testing for all patients, which would be expensive and complex. We focused on proteins that can be directly measured in blood plasma to keep the model practical.
65 Guyon, I. & Elisseeff, A. An introduction to variable and feature selection. J. Mach. Learn. Res. 3, 1157–1182 (2003).