ALS blood expression profiling identifies new biomarkers, patient subgroups, and evidence for neutrophilia and hypoxia

The study was performed with two cohorts (GSE112676 and GSE112680) and gene expression evaluated using two microarray platforms (Illumina HumanHT-12 V3.0 and HumanHT-12 V4.0 expression beadchip arrays) [31]. The V3.0 platform was used to profile expression in the GSE112676 cohort (n = 233 ALS and 508 CTL samples), and the V4.0 platform was used to profile expression in the GSE112680 cohort (n = 164 ALS, 137 CTL and 75 MIM samples) (Additional file 2A). Cohort demographics have been described previously [31]. In brief, ALS, CTL and MIM groups each included a higher percentage of males (≥ 55.3%; Additional file 2B) with average ages of 63.6, 62.6 and 57.9, respectively (Additional file 2C). Most patients (> 60%) had spinal- rather than bulbar-onset ALS (Additional file 2D). The GSE112680 cohort had a larger percentage of patients with C9orf72 repeat expansions (12.8% vs. 5.2%, Additional file 2G). Survival was defined as the time interval between disease onset to death, tracheostomy or noninvasive ventilation [31]. Given this definition, median survival was 2.44 years with 50% of patients surviving 1.59 to 3.87 years (Additional file 2F). The 75 MIM patients had been diagnosed with diverse ALS-like conditions, but the most common diagnoses were benign fasciculations (n = 9), spinal muscular atrophy (n = 8) and myelopathy (n = 8) (Additional file 2G).

An original analysis of these data had identified “batch effects” as a complicating factor [31], and we further expected that microarray platform-specific effects would be present [45]. To avoid confounding batch and platform-specific effects, we followed a “late stage” data integration strategy [46], by first analyzing data from each platform individually (V3 and V4) and correcting for batch effects as appropriate. This proved to be preferable to combining V3 and V4 data and attempting to correct batch and platform-specific effects simultaneously. Differential expression analyses were thus performed using data from each platform separately, and differentially expressed genes (DEGs) were identified by integrating summary statistics (fold-change and standard errors) via a random effects meta-analysis model. For other analyses (aside from differential expression testing), we adopted an “early stage” integration approach [46], in which it was necessary to combine expression values from both datasets for corresponding genes. In such cases, batch-corrected log₂-scaled expression values from each platform were Z-score normalized, respectively, to remove any platform-specific effects and maximize comparability of expression values.

Gene expression estimates for 741 samples were generated using the Illumina HumanHT-12 V3.0 expression beadchip platform (233 ALS and 508 CTL samples) [31]. Raw expression intensities and detection p-values were downloaded in June 2018 (GSE112676_HT12_V3_preQC_nonnormalized.txt). Detection p-values for 520 samples were negatively correlated with expression intensities as expected (r_s < − 0.90), but detection p-values for 221 samples were positively correlated with expression intensities (r_s > 0.90). Given this pattern, it was likely that detection p-values for the latter group did not correspond to actual p-values (P) but rather corresponded to 1 − P. Detection p-values for these 221 samples were thus subtracted from unity and after correction there was a strong negative correlation between detection p-values and expression intensity for all 741 samples as expected (r_s < − 0.90). The 741 samples varied with regard to median intensity, intensity IQR, and the number of protein-coding genes with detectable expression at a threshold of P < 0.05 (Additional file 3A–C). Overall, 8847 protein-coding genes were detected per sample on average (range: 5035–11,392; Additional file 3C). Sample index plots for these parameters were suggestive of a batch effect, with a higher signal IQR and number of detected genes for the first set of 448 samples (GSM3076582-GSM3077650) as compared to the second set of 293 samples (GSM3077652-GSM3078510) (Additional file 3D–F).

Background correction was performed using the normal–exponential convolution model, in which intensities are assumed to be the sum of two components, including a normally distributed background noise component and an exponentially distributed signal component (R package: limma; function: backgroundCorrect) [47]. Quantile normalization was applied after background correction to normalize expression intensities across the 741 samples (R package: limma; function: normalizeBetweenArrays) [47]. This generated normalized intensity estimates for 48,803 probes, although some of these were unannotated or quantified expression for the same gene (R annotation package: illuminaHumanv3.db). For each human gene, therefore, a single representative probe was selected, yielding a reduced set of 19,236 probes each corresponding to a unique human gene. For genes associated with multiple probes, the single probe with highest average expression across the 741 microarray samples was chosen as a representative. A final filtering step was performed to include 18,035 probes assaying the expression of protein-coding genes, i.e., those genes with an “NM_” and “NP_” prefix in their Reference Sequence (RefSeq) database identifiers.

These preprocessing steps generated a normalized expression matrix for 741 samples and 18,035 probes associated with a non-redundant set of protein-coding genes. The 741 samples were plotted with respect to the first 2 and 3 principal component (PC) axes, which again demonstrated a clear batch effect unrelated to disease status or sex (Additional file 3G, H). Similar to the pattern described above, the first PC score (PC1) differed according to sample index, with one batch corresponding to 448 samples (GSM3076582-GSM3077650) and a second batch corresponding to 293 samples (GSM3077652-GSM3078510) (Additional file 3I). To correct this effect, we first applied surrogate variable analysis (R package: sva; function: sva), as implemented previously for these data [31], with data adjusted using 1 surrogate variable chosen based upon asymptotic conditional singular value decomposition (R package: sva; function: num.sv) [48]. This improved the batch effect in some analyses (Additional file 3J, K), but did not resolve the relationship between PC1 scores and the sample index (Additional file 3L). We therefore applied the ComBat algorithm as an alternative strategy (R package: sva; function: ComBat) [35]. This approach removed any apparent batch effect in PC plots (Additional file 3M, N) and also resolved the relationship between sample index and PC1 score (Additional file 3O).

Following ComBat correction, one outlier sample (GSM3077426) was identified with respect to PCs 1 and 3 (Additional file 3M–O). This sample was removed prior to further analyses, but otherwise no outliers were visually evident from PC plots (Additional file 3M–O). Following removal of GSM3077426, Grubb’s test [49] for univariate outliers was non-significant with respect to each of the first 2 PC axes (P ≥ 0.74; R package: outliers; function: grubbs.test). The removal of 1 sample outlier (GSM3077426) was appropriate in our judgement and may be considered a less aggressive approach to outlier exclusion. For comparison, 67 outlier samples (34 ALS, 33 CTL) were identified and excluded in the initial GSE112676 analysis reported by van Rheenen et al. [31].

Gene expression estimates for 376 samples were generated using the Illumina HumanHT-12 V4.0 expression beadchip platform. Raw expression intensities and detection p-values were downloaded in June 2018 (GSE112680_HT12_V4_preQC_nonnormalized.txt). The 376 samples included 164 from ALS patients, 137 from control subjects, and 75 from MIM patients (Additional file 2G). Detection rate p-values were negatively correlated with signal intensity estimates for all samples as expected (r_s < − 0.99). Samples with a higher GSM identifier index tended to have increased median intensity and signal IQR but a lower number of detected genes (Additional file 4A–F). An average of 9152 protein-coding genes was detected among the 376 samples (range: 6939–11,061). Background correction and quantile normalization was performed as described above (R package: limma; functions: backgroundCorrect and normalizeBetweenArrays) [47]. This yielded normalized expression intensities for 47,323 probes and 376 samples. To limit redundancy in subsequent analyses, multiple probes associated with the same gene were filtered and a single representative probe was selected (R annotation package: illuminaHumanv4.db). This generated a filtered expression matrix consisting of 20,937 probes each representing a unique human gene. These probes were further filtered to include only protein-coding human genes (with “NM_” and “NP_” prefixes in RefSeq identifiers), leaving 18,490 probes upon which further analyses were based.

The 376 samples were plotted with respect to the first 3 PC axes, which suggested a weak batch effect involving only a small number of samples (Additional file 4G, H). Inspection of PC1 scores suggested that this effect was again (as above) related to sample ordering (GSM indices), with higher scores for the last 82 samples (GSM3080099–GSM3080180) as compared to the first 294 (GSM3079737–GSM3080098) (Additional file 4I). We first attempted to remove this effect using surrogate variable analysis (R package: sva; function: sva), with data adjusted using 2 surrogate variables (R package: sva; function: num.sv) [48]. As in the other cohort, this improved the batch effect for some analyses (Additional file 4J, K), but not the association between PC1 score and sample index (Additional file 4L). The alternative correction using ComBat was therefore applied (R package: sva; function: ComBat) [35], which did succeed in removing the batch effect (Additional file 4M, N) and the association between PC1 scores and sample index (Additional file 4O). There was no strong visual evidence for outliers and accordingly Grubb’s test for univariate outliers was non-significant with respect to the first 2 PC axes (P = 1.00) (R package: outliers; function: grubbs.test). The initial analysis of van Rheenen et al. [31] had identified and excluded 20 samples as outliers (13 ALS, 7 CTL) for the GSE112680 dataset, but this did not appear necessary following the normalization, batch correction and filtering procedures applied in the current analysis.

The above processing steps yielded an expression matrix for 18,035 protein-coding genes and 740 samples (232 ALS and 508 CTL samples). Differential expression analyses were performed using a subset of 11,210 genes with detectable expression in at least 20% (> 148/740) of samples. Effects of sex were removed by fitting a linear regression model for each gene, with expression as the response variable and sex (male or female) coded as a 0–1 categorical predictor variable. Residuals from the regression fit were used as sex-corrected expression values in subsequent analyses. Differential expression was evaluated using limma generalized least square linear models with moderated t-statistics (R package: limma; functions: lmFit and eBayes) [50]. To control the false discovery rate (FDR), raw p-values were corrected using the Benjamini–Hochberg method [51]. In all analyses, we define DEGs as genes with at least 10% expression change (FC > 1.10 or FC < 0.909) and FDR less than 0.10. Given these thresholds, we identified 2381 ALS-increased DEGs and 2646 ALS-decreased DEGs using only GSE112676 samples (Additional file 5). Up- and down-regulated FC estimates were symmetrically distributed and not associated with mRNA abundance (see Volcano and MA plots; Additional file 5).

The above processing steps yielded an expression matrix for 18,490 protein-coding genes and 376 samples (164 ALS, 137 CTL and 75 MIM samples). Differential expression analyses were performed for two comparisons (1: ALS vs. CTL, 2: MIM vs. CTL). For the ALS vs. CTL comparison, differential expression analyses were performed using 10,670 genes with detectable expression in 20% (61/301) of ALS and CTL samples. For the MIM vs. CTL comparison, differential expression analyses were performed using 10,679 genes with detectable expression in 20% (43/212) of MIM and CTL samples. Raw expression values were adjusted using residual analysis to remove effects of sex as described above. Differential expression testing was performed using linear models and moderated t-statistics (R package: limma; functions: lmFit and eBayes) [50] with correction of raw p-values using the Benjamini–Hochberg method [51]. These steps led to the identification of 1380 ALS-increased DEGs and 1186 ALS-decreased DEGs based upon the GSE112680 dataset (FDR < 0.10, FC > 1.10 or FC < 0.909; Additional file 6). Volcano and MA plots demonstrated that up- and down-regulated FC estimates were symmetrically distributed and not associated with mRNA abundance (Additional file 6).

A “late stage” meta-analysis data integration framework [46] was used to identify DEGs through integration of summary statistics generated from each cohort, respectively (GSE112676 and GSE112680). 9822 protein-coding genes were assayed in both datasets with detectable expression in at least 20% of GSE112676 samples (> 148/740) and at least 20% of GSE112680 samples (> 61/301). For these 9822 genes, log₂-adjusted FC estimates and their standard errors from both cohorts were combined using a random effects meta-analysis model, with pooling of FC estimates based upon inverse variance weighting (R package: meta; function: metagen). The mean expression difference between log₂-normalized expression intensities in ALS and CTL patients was used as the meta-analysis summary measure (equivalent to log₂-scaled FC estimates). This generated a meta-FC estimate for each of the 9822 genes, with a pooled meta-p-value providing a test for consistent differential expression in both cohorts. Meta-p-values were corrected for multiple hypothesis testing using the Benjamini–Hochberg method [51]. Up- and down-regulated meta-FC estimates were symmetrically distributed and not associated with mRNA abundance based upon inspection of volcano and MA plots (Additional file 7).

Riluzole (RZE) is the first-line ALS treatment [52], and since patients in our analyses would likely have taken this drug, some genes with ALS-altered expression may represent RZE responses. To address this possibility, we analyzed microarray data from a previous study of MDA-MB-231 cells (GSE96653) that compared expression between vehicle (DMSO)-treated cells (n = 3) and cells treated with 25 µM RZE for 24 h (n = 3) [53]. MDA-MB-231 is a breast adenocarcinoma cell line commonly used in cancer research [54]. Although this cell type is not directly related to ALS pathogenesis, we reasoned that RZE expression responses in the MDA-MB-231 cell line may be representative of those in blood-derived cells. Furthermore, expression responses in MDA-MB-231 cells had been profiled using the same Illumina HumanHT-12 V4.0 platform used to measure gene expression in the GSE112680 cohort (see above). This limited platform-based variation and it was possible to apply the same methods for background correction, quantile normalization, probe annotation, and probe filtering. The 6 array samples were mutually similar in terms of median intensity, IQR, and number of genes with detectable expression (P < 0.05, 11,362 on average), and there was no evidence for sample outliers (Additional file 8). Differential expression analyses were performed for 11,868 protein-coding genes with detectable expression in 2 of the 6 samples (R package: limma; functions: lmFit and eBayes) [50], leading to the identification of 255 RZE-increased DEGs (FC > 1.10) and 192 RZE-decreased DEGs (FC < 0.909) at an FDR threshold of 0.10 (Additional file 8).

The study of van Rheenen et al. [31] included a supplemental data file listing 2593 genes identified as differentially expressed (ALS vs. CTL) in the original analysis of the GSE112676 and GSE112680 datasets (FDR < 0.05 with FC > 1.50 or FC < 0.67). This list was used to assess overlap between ALS DEGs identified in the current study to those identified in the prior analysis [31]. Otherwise, among prior ALS patient blood gene expression studies (Additional file 1), one analysis considered whole blood [32] such that results should be comparable to those from the current analysis. This study had evaluated expression in smaller ALS and CTL patient cohorts (n = 30 per group) using the Illumina Sentrix HumanRef-8 Expression BeadChip microarray platform [32]. Raw data from this study has not been submitted to a public database, but a list of genes differentially expressed between ALS and CTL patients has been provided (based upon results from a two-sample t-test with FDR threshold of 0.05; see Additional file 1 from [32]). The supplemental file from this study was filtered to remove redundant or unannotated probes, leading to a filtered set of 2163 genes with differing expression between ALS and CTL patients (1089 ALS-increased and 1074 ALS-decreased). 1584 of these 2163 genes were included in our meta-analysis with detectable expression in both GSE112676 and GSE112680 cohorts (793 ALS-increased and 791 ALS-decreased). Fisher’s exact test was used to evaluate the overlap of these 1584 genes with those identified as altered in ALS patients from the current study. Additionally, using differential expression statistics reported previously [32], we extend the meta-analysis approach described above (R package: meta; function: metagen) to integrate FC estimates and standard errors with those obtained in our analysis (GSE112676 and GSE112680), leading to a “high confidence” set of genes differentially expressed in ALS blood. Results from this extended meta-analysis are included as supplemental material, although in this manuscript we focus on genes identified from the GSE112676 and GSE112680 cohort meta-analysis (for which all raw data are available).

Functional properties of ALS DEGs were evaluated by testing for enrichment of Gene Ontology [36] and Kyoto Encyclopedia of Genes and Genomes (KEGG) [37] annotations using a conditional hypergeometric test (R package: GOstats; function: hyperGTest) [55]. Non-conditional hypergeometric tests were performed to assess DEG enrichment for Reactome [56] and Disease Ontology (DO) [57] annotations (R packages: ReactomePA and DOSE; functions: enrichPathway and enrichDO) [58, 59]. ALS-associated genes were identified based upon a previous analysis of 9 database resources linking genes to specific diseases [60]. A gene was considered to be ALS-associated if it was linked to ALS based upon at least 2 of the 9 databases included in the analysis [60]. ALS DEGs were additionally assessed for enrichment of annotations included in the Pathway Commons database (R package: paxtoolsr) [61]. ALS DEGs were also evaluated to assess for overlap with gene sets included in the MSigDB database (R package: msigdbr) [62]. The MSigDB database is a collection of annotated gene sets developed to be used for gene set enrichment analysis (GSEA). Our analysis considered the “C7” MSigDB database collection, which includes 4872 gene sets derived from microarray studies of immune cells [62]. In addition to the MSigDB database, ALS expression changes were further compared with those observed in a previously compiled set of 462 gene expression signatures, where each signature was established from comparisons among human PBMC microarray samples available in Gene Expression Omnibus (GEO) [63]. Genetic loci associated with ALS were identified using the NHGRI GWAS catalog [40].

The expression of exosome-associated genes [39, 64] was evaluated to determine if corresponding mRNAs were disproportionately increased or decreased in ALS patients. We evaluated 91 proteins with altered abundance in serum exosomes from ALS patients compared to control subjects (n = 3 per group; 41 ALS-increased, 50 ALS-decreased), which had previously been identified using mass spectrometry (see Table S4 from Tomlinson et al. [65]). We also evaluated 83 exosome-associated mRNAs from the ExoCarta database [66] that had been identified in exosomes from at least 4 separate experiments (download file: EXOCARTA_PROTEIN_MRNA_DETAILS_5.txt). We further evaluated a list of the 100 mRNAs most frequently detected in human exosomes compiled by the EVpedia database [67]. Expression of ALS DEGs in blood-derived exosomes was quantitatively compared to that of non-DEGs using TPM-normalized (Transcript Per Million) expression values downloaded from exoRBase [68] (filename: Normal_mRNA_TPM.txt), which had been generated by mapping of mRNA-seq reads (hg38 reference genome) generated by Li et al. [69] (GSE100206). TPM values were averaged across samples from 32 normal subjects [69].

Since gene expression in unfractionated blood is partly determined by the abundance of constituent immune cells, in silico approaches have been developed to quantify inter-sample variation in cell type abundance based upon gene expression in blood or other whole tissues [70‐73]. We assembled a gene expression database to identify genes with cell type-specific expression in 12 immune cell types (neutrophils (NP), monocytes (MC), dendritic cells (DC), macrophages (MP), platelets (PL), red blood cells (RBC), eosinophils (ES), CD4 T cells (CD4), CD8 T cells (CD8), gamma-delta T cells (GDT), B cells (B) and NK cells (NK)). We here define a “cell type-specific expression pattern” as one in which a gene’s expression is quantitatively higher in one cell type as compared to other cell types, even though expression of a gene may be qualitatively detectable in multiple cell types. In this sense, cell type-specific expression is not a binary concept (expressed vs. not expressed) but varies along a continuum. To identify genes exhibiting such a pattern, we assembled a database of samples deposited in the GEO database [74] that had been generated using the Affymetrix Human Genome U133 Plus 2.0 platform. Database samples were identical to those described in a previous analysis [41] for 8 cell types (NP, n = 492; MC, n = 560; DC, n = 491; MP, n = 450; CD4, n = 574; CD8, n = 161; B, n = 435; NK, n = 160). For the other 4 cell types, additional samples were added to the database (PL, n = 48; RBC, n = 41; ES, n = 7; GDT, n = 17), but procedures followed for normalization and identification of cell type-specific genes were otherwise consistent with those previously described [41]. We refer to the RBC group as “RBC/erythroid lineage cells”, since experiments used for gene expression analysis evaluated RBC precursors such as erythroblasts and reticulocytes (see GEO accessions GSE17639, GSE18679, GSE22552, GSE24849, GSE65577).

Samples for each cell type were normalized using Robust Multichip Average (RMA) [75], and if more than 50 samples were available for a given cell type, the 50 samples with lowest average Euclidean distance to all other samples were selected as representatives [41]. Following normalization and sample selection, cell type-specific genes were identified by two-group comparisons between each cell type and all samples associated with the other 11 cell types (R package: limma; functions: lmFit and eBayes) [41]. To select signature genes for a given cell type, we first identified the set of 150 genes with elevated expression in that cell type with lowest p-values from the two-group comparison. These 150 genes were then sorted based upon FC (target cell type/other cell types), and the 100 genes with highest FC were selected as signature genes for the target cell type. Given these thresholds, the final set of 100 signature genes chosen for each cell type exhibited higher expression in samples from the target cell type when compared to the pooled set of samples from the 11 other cell types, with no less than 2.48-fold elevated expression compared to the 11 other cell types (P ≤ 1.64e−07 in all comparisons). In every case, median expression of the 100 signature genes was at least 60% greater in the target cell type than median expression in every other cell type individually (Additional file 9).

To obtain signature scores for each blood sample, we calculated the (weighted) average of Z-score normalized expression estimates for the 100 signature genes identified for each cell type. This average was calculated with greater weight assigned to genes with a stronger cell type-specific expression pattern (R function: weighted.mean). Preliminary weights assigned to each gene were equal to w^1/2, with w = 100 for the top-ranked gene, w = 99 for the second-ranked gene, w = 98 for the third ranked gene, and so on. Preliminary weights were then scaled to the [0, 1] interval by dividing preliminary weights by the maximum value, i.e., (100)^1/2, yielding final weights used for the weighted average signature score calculation [41].

These same procedures were followed to calculate M1 and M2 macrophage signature scores based upon a previously published dataset (GSE5099) [76]. M1 signature genes were identified from the two-group comparison between M1 macrophages (GSM115055–GSM115057, GSM115070–GSM115072) and the combined set of M2 and non-polarized macrophages (GSM115052–GSM115054, GSM115058–GSM115060). Similarly, M2 signature genes were identified by comparing expression between M2 macrophages (GSM115058–GSM115060, GSM115073–GSM115075) and the combined set of M1 and non-polarized macrophages (GSM115052–GSM115057, GSM115067–GSM115072). Normalization and differential expression analyses were performed as described above using RMA and linear models (R package: limma; functions: lmFit and eBayes). Experiments were performed using two early-generation Affymetrix microarray platforms (U133A and U133B). Differential expression analyses were thus performed for each platform separately, followed by late-stage integration of differential expression summary statistics [46], excluding any genes from the U133B analysis that were already included from analysis of the more comprehensive U133A platform. As above, signature genes were identified by first selecting the 150 genes with lowest p-values in two-group comparisons (M1 vs. other, M2 vs. other), and of these the 100 genes with high FC were chosen as signature genes (M1/other or M2/other). M1 and M2 signature scores were then calculated for each blood sample as described above, with weighted averaging of Z-score normalized expression of signature genes, with increased weight assigned to those genes having expression most specific to M1 or M2 macrophages.

Alternative methods have been proposed for cell type deconvolution, which utilize different algorithms and cell type marker genes chosen from compiled databases [70‐73]. Results obtained using the above approach were thus compared with those obtained using the ImSig algorithm, which calculates sample-specific cell type scores based on a different algorithm with independently identified cell type marker genes (R package: ImSig) [70].

Diagnostic accuracy was defined as the ability of blood gene expression measures to discriminate ALS from non-ALS patients (MIM and CTL). Area under the curve (AUC) statistics were calculated to assess diagnostic accuracy based upon the expression of individual genes (R library: pROC, function: ci.auc) [77]. Cross-validation analyses were performed using randomly chosen training (296 ALS patients vs. 296 CTL/MIM subjects) and testing sets (100 ALS patients vs. 100 CTL/MIM subjects) with 10,000 simulation trials. For individual genes, classification models were generated from training data using logistic regression (R function: glm). For multigene models, classification models were generated from training data using the random forest algorithm (R package: randomForest; R function: randomForest) [78], logistic regression [79] (R function: glm; binomial error distribution), and support vector machines [80] (R package: e1071; function: svm). In each simulation trial, accuracy, sensitivity and specificity were calculated (R library: caret; function: confusionMatrix) and McNemar’s Chi squared test was used to test the hypothesis of confusion matrix marginal homogeneity (R function: mcnemar.test) [81]. We then calculated the proportion of significant McNemar tests and average accuracy, sensitivity and specificity across the 10,000 simulation trials.

Cox proportional hazard (PH) models were used to evaluate the significance of gene expression-survival associations and estimate corresponding hazard ratios in covariate-adjusted models (R library: survival; function: coxph). Analyses were performed for 11,480 protein-coding genes with detectable expression in at least 20% of ALS patients (> 80/396). All Cox PH models included age, sex, site of onset, and cohort as covariates, with additional variables corresponding to the expression of one or more genes. A multigene Cox PH model was developed using stepwise variable selection with p-value of 0.15 required for variables to enter or remain within the regression model (R library: My.stepwise; function: My.stepwise.coxph). To assess survival prediction accuracy using cross-validation, 10,000 simulation trials were performed, with patients randomly assigned to a training (n = 296) or test set (n = 100) in each trial. The training set was used to estimate coefficients for a Cox PH model, which was then applied to the testing set, yielding linear predictor scores used to stratify patients in terms of predicted survival (function: predict.coxph). Correspondence between linear predictors and test set survival times was evaluated using the AUC concordance index proposed by Heagerty and Zheng (R library: risksetROC; function: risksetAUC) [82]. For any two randomly chosen test set patients, this index estimates the probability that Cox PH model outputs successfully determine which patient survives longer [82, 83]. A total of 10,000 simulation trials were performed and the average concordance index was determined. The average concordance was compared between a base model (clinical covariates only) and a full model (clinical covariates + gene expression) to assess the contribution of expression variables to survival prediction accuracy.

Meta-analysis identified 752 ALS-increased DEGs (FDR < 0.10 with FC > 1.10) with a consistent differential expression pattern in both cohorts (GSE112676 and GSE112680) (Additional file 10A). Genes most strongly elevated in ALS blood included ribosomal protein L9 (RPL9), ribosomal L24 domain containing 1 (RSL24D1), vanin 2 (VNN2) and mitochondrial amidoxime reducing component 1 (MARC1) (Fig. 1). Several top-ranked ALS-increased DEGs had previously been associated with ALS, such as matrix metallopeptidase 9 (MMP9) [84], ATP binding cassette subfamily G member 1 (ABCG1) [85], and selectin L (SELL/CD62L) [86] (Fig. 1a, b). There was no significant overlap between ALS-increased DEGs and genes up-regulated in MDA-MB-231 cells following riluzole treatment (25 µM for 24 h; P = 0.66; Fig. 1d) [53]. ALS-increased DEGs were most strongly enriched with respect to ribosome-associated genes (Additional file 11) and were additionally associated with translation and neutrophil activation (Additional file 11A). Comparison to MSigDB gene sets [62] showed that ALS-increased DEGs overlapped significantly with genes elevated in neutrophils compared to B or T cells (Additional file 11G). ALS-increased DEGs were also significantly associated with immune cell activation, viral transcription, and RAGE receptor binding (Additional file 11).

The 752 ALS-increased DEGs were compared to those identified in the original analysis of van Rheenen et al. [31]. Of the 752 ALS-increased DEGs, 208 (28%) had been identified as differentially expressed in the original analysis (P = 5.6e−26; Fisher’s exact test). We next compared the 752 DEGs we identified to DEGs from the prior study of Saris et al. [32], which compared whole blood gene expression in ALS patients and controls using a different set of samples (n = 30 per group) and different microarray platform. Of the 752 ALS-increased DEGs, 225 (29.9%) had been reported as differentially expressed by Saris et al. [32], and nearly all of these (224/225) were ALS-increased as observed in the current study (P = 2.1e−76, Fisher’s exact test; Additional file 12A). Conversely, the complete set of ALS-increased DEGs reported by Saris et al. [32] was biased towards increased expression in ALS blood samples from the current study (P = 1.51e−183, Wilcoxon rank sum test; Additional file 12B). These results demonstrate good directional consistency between our results and those reported by Saris et al. [32]. A ranked list of “high confidence” ALS-increased genes, generated from meta-analysis of results from the current study and results from Saris et al. [32], is included in supplemental materials with gene set enrichment analysis findings (Additional file 12C–E).

Meta-analysis identified 764 ALS-decreased DEGs (FDR < 0.10 with FC < 0.909) (Additional file 10B). Genes most strongly decreased in ALS blood included prostate and testis expressed 2 (PATE2), BCR, RhoGEF and GTPase activating protein (BCR), host cell factor C1 (HCFC1), and leukocyte immunoglobulin like receptor B1 (LILRB1) (Fig. 2). At least one top-ranked ALS-decreased gene had previously been associated with ALS (i.e., heat shock protein family B small member 1, HSPB1; Fig. 2a) [87]. Overlap between ALS-decreased DEGs and genes down-regulated in MDA-MB-231 cells following riluzole treatment (25 µM for 24 h) was non-significant (P = 0.11; Fig. 2d) [53]. Genes decreased in ALS patient blood were associated with TGF-beta response, Z disc, antigen processing/presentation, and platelet/clotting disorders (e.g., blood platelet disease, hemorrhagic thrombocythemia, congenital hemolytic anemia) (Additional file 13). The strongest association was significant overlap between ALS-decreased genes and genes up-regulated in PBMC from infants with acute respiratory syncytial virus (RSV) infection (as compared to infants with influenza; Additional file 13G).

The 764 ALS-decreased DEGs were compared to those identified in the original analysis of van Rheenen et al. [31]. 236 of the 764 (30.9%) had been identified as differentially expressed in the original analysis (P = 1.55e−37; Fisher’s exact test). The 764 DEGs were next compared to those identified in the prior microarray study performed by Saris et al. using a different sample set (n = 30 per group) [32]. Of the 764 ALS-decreased DEGs, 186 (24.3%) had been identified as differentially expressed by Saris et al. [32], and the majority of these (174/186) were ALS-decreased as in the current analysis (P = 1e−43, Fisher’s exact test; Additional file 14A). The complete set of ALS-decreased genes identified by Saris et al. [32] was biased towards ALS-decreased expression in the current study (P = 7.7e−140, Wilcoxon rank sum test; Additional file 14B). High confidence ALS-decreased genes identified from meta-analysis of results from the current analysis and prior study of Saris et al. [32] are listed with further analysis in supplemental materials (Additional file 14C–E).

Annotation-based enrichment analyses showed that ALS-increased DEGs were linked to neutrophil-associated terms (Additional file 11A), whereas ALS-decreased DEGs were linked to anemia and bleeding disorders (Additional file 13F). To dissect this further, DEGs were compared to lists of genes specifically expressed in 12 peripheral blood cell types [41]. Among the 12 cell types, the 752 ALS-increased DEGs were most significantly enriched for genes with neutrophil-specific expression (Fig. 3a, b), whereas ALS-increased DEGs included very few genes with RBC lineage-specific expression (Fig. 3a, c). An opposite pattern was observed among ALS-decreased DEGs, which were enriched for RBC lineage-specific genes but few neutrophil-specific genes (Fig. 3d, e). Overall, 32.6% of ALS-increased DEGs (238/731) had higher expression in neutrophils than any other cell type (Fig. 3g), while 13.8% of ALS-decreased DEGs (101/734) had higher expression in RBC lineage cells than any other cell type (Fig. 3h). ALS-increased DEGs most highly expressed in neutrophils included VNN2, BCL6 and MME (Fig. 3i), and ALS-decreased DEGs most highly expressed in RBC lineage cells included DHFR, CRCP, and SHCBP1 (Fig. 3j).

The analysis was repeated using the ImSig database [70], which provides signature gene sets for 8 cell types and 2 biological processes (translation and proliferation) with calculation of scores based upon signature gene expression and co-expression. This again showed that neutrophil signature scores were most strongly elevated in ALS patients, with significant elevation of translation, plasma cell and macrophage scores as well (Additional file 15A). In contrast, NK cell signature scores were reduced in ALS patients (Additional file 15A), in agreement with over-abundance of NK cell-specific genes among ALS-decreased DEGs (Fig. 3h). 13 of the ImSig neutrophil signature genes were ALS-increased DEGs, whereas none were ALS-decreased DEGs (Additional file 15D, E). Similarly, 29 translation signature genes were ALS-increased DEGs and none were ALS-decreased DEGs (Additional file 15F, G).

Enlargement of blood-derived exosomes was recently demonstrated in ALS patients [38], and proteins with altered abundance in blood exosomes from ALS patients have been identified [65]. We compared mRNA FC estimates in the current study to those obtained previously for exosome proteins altered in ALS patients compared to CTL subjects [65], which revealed a weak positive FC correlation (r_s = 0.16, P = 0.25) and 5 mRNA-protein pairs with consistent changes in abundance (ALS-increased: THBS1; ALS-decreased: DPYSL5, SLC4A1, TTN, TLN1) (Additional file 16A). We next identified exosome-associated mRNAs from the ExoCarta [66] and EVpedia [67] databases and showed that these were biased towards ALS-increased expression (P ≤ 6.3e−3; Additional file 16B, C). Genes with higher expression in blood-derived exosomes from normal subjects (n = 32) [68, 69] were also more likely to be elevated in ALS blood (Additional file 16D). Conversely, genes most strongly elevated in ALS blood tended to have high blood exosome expression (Additional file 16E), while genes most strongly decreased in ALS blood had lower blood exosome expression (Additional file 16F). ALS-increased DEGs most highly expressed in blood-derived exosomes encoded ribosomal subunits and other translation-associated proteins (e.g., EEF1A1, RPL13A, RPS6; Additional file 16G, H).

Gene expression shifts in ALS blood were compared to 462 signatures derived from human PBMC gene expression datasets [63]. This was a hypothesis-generating analysis for which the goal was to determine if ALS-like expression patterns could be discerned from existing datasets, which may help to explain patterns observed in our study. There was significant correspondence to PBMC signatures associated with old age, active tuberculosis and influenza vaccination (Fig. 4a). However, the strongest match was obtained with respect to a gene expression signature derived from the comparison of PBMC from subjects at sea level to those rapidly transported to high altitude (GSE46480; P = 4.89e−17; Wilcoxon rank sum test; Fig. 4a). Consistent with this, there was a significant genome-wide correlation between ALS/CTL and high altitude/sea level FC estimates (r_s = 0.52; Fig. 4b). The 752 ALS-increased DEGs were significantly enriched for genes elevated with altitude stress (Fig. 4c, and likewise, the 764 ALS-decreased DEGs were enriched for genes decreased by altitude stress (Fig. 4d). Among 714 ALS-increased DEGs assayed in both experiments, 567 (79.4%) were elevated in high altitude subjects (Fig. 4e). Likewise, among 714 ALS-decreased DEGs assayed in both experiments, 403 (56.4%) were decreased in high altitude subjects (Fig. 4e). ALS-increased genes most strongly elevated in high altitude subjects included COMM domain containing 6 (COMMD6), C-type lectin domain family 2 member B (CLEC2B), and ribosomal protein L31 (RPL31) (Fig. 4f) and as a group these genes were associated with protein targeting to membrane, mRNA decay and interferon-beta synthesis (Fig. 4h). ALS-decreased genes most strongly decreased in high altitude subjects included stromal interaction molecule 1 (STIM1), zinc finger protein 652 (ZNF652) and RNA binding motif protein 14 (RBM14), and collectively such genes were associated with modulation by virus of host, anion transmembrane transport, and the MAPK cascade (Fig. 4i).

It has been unclear whether genes with altered expression in ALS blood are simply responses to disease progression or instead involved with disease-causing pathogenetic mechanisms [88]. To address this, we evaluated overlap between the top 500 ALS increased/decreased DEGs and genes near ALS GWAS susceptibility loci [40, 89]. ALS-increased DEGs overlapped significantly with genes near susceptibility loci (distance ≤ 62 kb; P < 0.05, Fisher’s exact test), but did not overlap significantly with genes distant from susceptibility loci (distance > 62 kb) (Fig. 5a). For instance, among the top 500 ALS-increased DEGs, 6 of 500 (1.2%) overlapped significantly with genes less than 9 kb from ALS GWAS loci, and this fraction was significantly greater than observed for non-DEGs, i.e., only 49 of 8790 (0.56%) non-DEGs were within 9 kb of a GWAS locus (P = 0.028, Fisher’s exact test). Consistent with this, the average distance between each ALS-increased DEG and its nearest susceptibility locus was significantly reduced compared to randomly sampled genes (P = 0.004; Fig. 5b). In contrast, ALS-decreased DEGs were not significantly more likely to be near a GWAS locus, although similar trends were noted in both analyses (Fig. 5a, c). ALS-increased DEGs nearest to susceptibility loci included interferon related developmental regulator 1 (IFRD1), TANK binding kinase 1 (TBK1), cAMP responsive element binding protein 5 (CREB5), ATP binding cassette subfamily G member 1 (ABCG1), selectin L (SELL) and annexin A3 (ANXA3) (Fig. 5d, f–j). ALS-decreased DEGs nearest to susceptibility loci included methyltransferase like 21A (METTL21A), T cell lymphoma invasion and metastasis 1 (TIAM1) and erythrocyte membrane protein band 4.1 (EPB41) (Fig. 5e, j, k).

Prior studies have suggested that diagnostic blood indicators for ALS may include increased abundance of phosphorylated neurofilament heavy chain (pNFH) and neurofilament light chain (NFL) proteins [6, 90]. We evaluated whether expression of corresponding genes (NEFH and NEFL) could discriminate ALS and non-ALS (CTL + MIM) subjects, but cross-validation analyses showed that expression of these genes could predict diagnosis with only 49–53% accuracy (logistic regression; Additional file 17A–C). We next searched for single genes for which low or high expression might be diagnostic, focusing on ALS-increased DEGs with low MIM/CTL FC estimates (e.g., MCTP2, MGAM, CREB5; Additional file 17D) and ALS-decreased DEGs with high MIM/CTL FC estimates (e.g., LDHB, METTL16, FAM102A; Additional file 17E). All such genes had significant AUC statistics for discrimination between ALS and non-ALS subjects (Additional file 17F, G). The highest AUCs (≥ 0.66) were calculated for multiple C2 and transmembrane domain containing 2 (MCTP2) and RNA polymerase III subunit C (POLR3C) (Additional file 17F, G). Cross-validation testing showed that expression of these genes could predict ALS diagnosis with 62–63% accuracy (logistic regression; Additional file 17H–J).

Genes with altered expression in ALS blood were specifically expressed by neutrophils and RBC-lineage cells, and ALS-like expression shifts were observed in subjects transported to high altitudes (Figs. 3 and 4). Signature scores for immune cell types and altitude stress were calculated and several had significant AUC statistics (P < 0.05), with highest AUCs obtained for neutrophil (0.64), NK cell (0.60) and eosinophil (0.58) scores (Additional file 18A). Closer inspection of the neutrophil signature showed it was elevated in ALS patients, but not similarly increased in MIM subjects (Additional file 18B), and cross-validation analyses showed that neutrophil score could predict ALS diagnosis with 60% accuracy (logistic regression; Additional file 18C). The ratio of neutrophils to monocytes has previously been shown to predict ALS diagnosis correlate with disease progression [91]. Along these lines, combining neutrophil scores with those from other cell types in bivariate models improved AUCs slightly (+monocytes: 0.642, +macrophage: 0.655; +eosinophils: 0.660). However, among bivariate score combinations, the highest AUC was obtained for models combining neutrophil and altitude stress scores (Additional file 18D). These two scores were negatively correlated among patients (r = − 0.21; Additional file 18E) and together predicted ALS diagnosis with 64% accuracy (logistic regression; Additional file 18F).

We next evaluated the diagnostic performance of multigene classifiers using random forest [78], logistic regression [79] and support vector machines (SVMs) [80]. Random forest variable importance scores were calculated to determine which genes, may be most important to prediction accuracy within multigene models (Fig. 6a. This highlighted genes such as brain protein I3 (BRI3), ATP synthase membrane subunit e (ATP5ME), and host cell factor C1 (HCFC1) as important multigenic model predictors (Fig. 6a). Cross-validation analysis showed that random forests using 450 input genes could predict ALS diagnosis with 77% accuracy on average (sensitivity: 78%, specificity: 77%) (Fig. 6b). However, using 63 PC scores as random forest predictors, rather than 450 input genes, improved prediction accuracy to 82% (sensitivity: 81%, specificity: 83%; Fig. 6c, d). Using logistic regression models with PC predictors further improved prediction accuracy to 83% (sensitivity: 83%, specificity 83%; Fig. 6e, f). Ultimately, our best results were obtained using SVMs, which could predict ALS diagnosis with 87% accuracy (sensitivity 86%, specificity 87%; Fig. 6g, h, I).

We reviewed the literature to identify 72 pairs of sensitivity and specificity estimates from previous fluid-based biomarker studies comparing ALS patients to controls (healthy controls or patients with non-ALS neurological diseases) (Additional file 19). The 86% sensitivity and 87% specificity obtained from SVM classification was better than 21 of 25 prior blood studies (Fig. 6j). Notably, however, most of the higher sensitivity/specificity estimates had been generated from small cohort studies (e.g., n ≤ 50 per group), and none of the prior blood biomarker studies performed with larger sample sizes (n ≥ 60 per group) had achieved greater than 70% sensitivity and specificity (Fig. 6k).

Disease course varies among ALS patients with rapid progression in some and slow progression in others [92], or in rare cases, apparent disease reversal [93]. We evaluated 11,480 protein-coding genes to determine if expression was associated with survival in covariate-adjusted Cox PH models, and found that a larger-than-expected proportion of genes (1204/11,480, 10.5%) were survival-associated (P < 0 0.05). No single gene was significant at a stringent FDR threshold (FDR < 0.10), although 6 genes were survival-associated at a less conservative threshold (i.e., FDR < 0.30; Additional file 20A, C). The most significant survival-associated gene was copper chaperone for superoxide dismutase (CCS), which was expressed at higher levels in patients with improved survival (HR = 0.77; P = 1.84e−05; FDR = 0.14) (Additional file 20C). In contrast, genes encoding neurofilament proteins were not significantly associated with survival (NEFH: HR = 1.02, P = 0.77; NEFL: HR = 1.03, P = 0.59). Survival-associated genes with high expression linked to poor survival (P < 0.01) were associated with protein metabolism, macrophage differentiation, and platelet activation (Additional file 20B). Survival-associated genes with high expression linked to improved survival (P < 0.01) were associated with the mitochondrial respiratory chain, semaphorin-plexin signaling and nucleotide phosphorylation (Additional file 20D). Cross-validation analyses were performed to determine if individual genes could predict survival, but adding the expression of single genes only slightly improved performance of Cox PH models already including clinical covariates (age, site-of-onset, sex and cohort; mean C = 0.62 vs. mean C = 0.60–0.61; Additional file 20E–G). Covariate-adjusted accelerated failure time models [94], however, showed that predicted median survival differed 20–33% depending upon whether an individual gene had low (20th percentile) or high (80th percentile) expression (Additional file 20H–J).

Gene signature scores were calculated by averaging Z-score normalized expression for the 100 genes most specifically expressed by each immune cell type (or the 100 genes most highly elevated with high altitude stress). Higher RBC and MP scores were associated with improved survival in covariate-adjusted Cox PH models (RBC: HR = 0.71, FDR = 0.018; MP: HR = 0.41; FDR = 0.018) (Additional file 21A). Consistent with this, the 2-way combination RBC + MP predicted survival better than any other combination, and RBC + PL + MP was the best 3-way combination (likelihood ratio tests; Additional file 21B, C). The combination MC + NP, previously reported to predict disease progression [91], was a less effective combination for predicting survival (Additional file 21B). Notably, altitude stress signatures were not significantly associated with survival (HR = 1.06; P = 0.47; FDR = 0.79) (Additional file 21A). Signature scores or combinations thereof only slightly improved performance of predictive models in cross-validation analyses (Additional file 21D–F), although predicted median survival improved 27% in patients with favorable RBC + PL + MP scores (Additional file 21I).

To deal with ALS patient heterogeneity, clinical trials have increasingly defined more homogeneous patient cohorts for enrollment [95]. This strategy can be biomarker-driven and is well-suited for immune-modulating drug candidates with specific targets (e.g., the anti-IL6R drug tocilzumab) [42, 43, 96]. Immune cell and high altitude stress signature scores varied among the 396 patients in our analysis (Fig. 7a). We discerned two patient groups, with one group having higher expression of genes expressed by myeloid-derived cells (NP, MC, DC, MP, PL, RBC, ES), and the other having higher expression of genes expressed by lymphoid-derived cells (CD8, CD4, GDT, B, NK) or high altitude response genes (Fig. 7a). Scores were averaged for each cell type group, yielding negatively correlated “myeloid” and “lymphoid” composite scores, respectively (Fig. 7b).

We identified 120 patients with high myeloid but low lymphoid scores, and conversely 90 patients with high lymphoid but low myeloid scores (Fig. 7b). Comparing these groups, T-cell marker gene (CD2, CD3D) and IL23A expression was increased in the lymphoid group (Fig. 7c, d, f). The myeloid group had higher expression of IL6R, IL-17 family genes (IL17RA, IL17C, IL17REL) and IL-1 family genes (IL1RN, IL1R2, IL1B) (Fig. 7d, e). We considered whether myeloid and lymphoid groups corresponded to “high inflammation” and “low inflammation” groups identified in a prior PBMC gene expression study [42]. Among genes previously shown to be elevated in “high inflammation” ALS patients, there was indeed a significant trend towards higher expression in our myeloid patient group (e.g., CD14, RXRA, MMP9; P = 0.015; Fig. 7g, h). However, consistent trends were not observed for all genes (e.g., MGAT2, PLCG1, PTGS2) including interleukin 6 (IL6) and interleukin 6 receptor (IL6R) (Fig. 7d, g).

Circulating monocytes may develop a pro-inflammatory phenotype in ALS patients with early stages of M1 macrophage polarization discernable from gene expression [97]. Signature scores were calculated based upon the expression of genes induced during macrophage polarization [76], which revealed a trend towards improved survival in patients with high M2 signatures (HR = 0.76; P = 0.39) and worse survival in patients with high M1 signatures (HR = 1.17), although neither association was significant (P = 0.53 and P = 0.39, respectively) (Additional file 22A). M1 and M2 signatures were weakly associated with myeloid and lymphoid scores (Additional file 22B, C), with both scores marginally elevated in myeloid group patients (M1: P = 0.056; M2: P = 0.161; Additional file 22D, E). Genes most strongly elevated during M1 or M2 polarization did not differ consistently between myeloid and lymphoid group patients (Additional file 22F, G).

Our analysis identified individual genes and signature scores associated with survival in Cox PH models, although these did not substantially improve prediction performance when evaluated by cross-validation (Additional file 19E–G; Additional file 20D–F). We therefore determined whether cross-validation performance could be improved by developing a more complex multigenic Cox PH model (Fig. 8). 2828 genes marginally associated with survival in single-gene Cox PH models were isolated (P < 0.15) and, as expected, expression of many of these genes was correlated (Fig. 8a, d). The 2828 genes were clustered (distance metric: |1–r|), yielding 211 non-overlapping gene groups (13.4 genes/group on average; range: 5–90 genes/group), from which we selected the 1 gene per group most strongly associated with survival in single-gene Cox PH models (Fig. 8b, e). These 211 genes were iteratively entered into a multigenic Cox PH regression model (i.e., forward and backward variable selection), leading to a final set of 61 genes (Fig. 8c, f, g; Additional file 23). Cross-validation showed that adding these 61 genes to Cox PH models with clinical covariates (age, site-of-onset, sex and cohort) substantially improved performance (mean C = 0.74 vs. mean C = 0.60; Fig. 8h). Predicted median survival differed 2-fold depending upon whether patients had a favorable or risk-associated expression pattern among the 61 predictor genes (Fig. 8i, j).