Skip to main content
Fig. 8 | Journal of Translational Medicine

Fig. 8

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

Fig. 8

Blood gene expression signature to predict ALS patient survival. a Gene selection approach. 11,480 protein-coding genes expressed in at least 20% of ALS samples were filtered to include 2828 genes associated with survival in covariate-adjusted Cox PH models (P < 0.15; covariates: age of onset + site of onset + sex + batch). These genes were clustered into 211 groups with one representative gene chosen per group. Forward and backward variable selection was then used to obtain the final set of 61 genes. b, c Correlation matrices. Heatmaps show expression correlations among the filtered sets of 2828 and 61 genes. d–f Expression heatmaps and hazard ratios. Blue-yellow heatmaps show Z-score normalized expression (rows: genes; columns: patients; clustering: hierarchical based upon Euclidean distance) with hazard ratios (right) (d: 2828 gene set; e: 61 gene set). f Gene importance (likelihood ratio tests). P-values were calculated by comparing the full model log-likelihood (61 genes + covariates) with that of reduced models (60 genes + covariates) obtained by dropping each gene from the full model. The 28 genes with lowest likelihood ratio test p-values are shown (vertical axis: -log10-transformed p-values; red font: HR > 1.00; blue font: HR < 1.00). g Gene symbol cloud. The 61 signature genes are shown with size inversely proportional to likelihood ratio test p-values (red: HR > 1.00; blue: HR < 1.00). Gene symbol colors are proportional to hazard ratio estimates among from monogenic Cox PH models relating ALS patient survival to gene expression and covariates (brown: HR < 0, turquoise: HR > 0). h Cross-validation analysis of survival prediction accuracy (Cox PH model with 61 genes + covariates; training set: 296 ALS patients; testing set: 100 ALS patients). Concordance index distributions are shown for the base model (clinical covariates only) and full model (clinical covariates + 61 signature genes). Boxes (top) outline the middle 50% of outcomes (middle line: median; whiskers: 10th to 90th percentiles). i Predicted survivorship with favorable and risk gene expression signatures. The Cox PH model was fit and predicted 50% survival times were obtained for each patient (n = 396; clinical covariates + 61 gene signature). The favorable signature was obtained from the patient with predicted 50% survival time nearest to the 80th percentile, and the risk signature was obtained from the patient with predicted 50% survival time nearest to the 20th percentile. Estimated median survivorship for each group is shown (dashed lines) with the between-group ratio (upper-right). j Predicted survival times across the observed range of gene expression signatures. Patient percentiles were calculated from predicted survival times and boxes outline the middle 50% of survivorship times for each percentile (25th to 75th percentile; middle line: median)

Back to article page