Ischemic Stroke with Comorbid Cancer Has Specific miRNA-mRNA Networks in Blood That Vary by Ischemic Stroke Mechanism

Abstract

Objective: Approximately half of ischemic strokes (IS) in cancer patients are cryptogenic, with many presumed cardioembolic. We evaluated whether there were specific miRNA and mRNA transcriptome architectures in peripheral blood of IS patients with and without comorbid cancer, and between cardioembolic versus noncardioembolic IS etiologies in comorbid cancer.

Methods: We studied patients with cancer and IS (CS; n = 42), stroke only (SO; n = 41), and cancer only (n = 28), and vascular risk factor-matched controls (n = 30). mRNA-Seq and miRNA-Seq data, analyzed with linear regression models, identified differentially expressed genes in CS versus SO and in cardioembolic versus noncardioembolic CS, and miRNA-mRNA regulatory pairs. Network-level analyses identified stroke etiology-specific responses in CS.

Results: A total of 2,085 mRNAs and 31 miRNAs were differentially expressed between CS and SO. In CS, 122 and 35 miRNA-mRNA regulatory pairs, and 5 and 3 coexpressed gene modules, were associated with cardioembolic and noncardioembolic CS, respectively. Complement, growth factor, and immune/inflammatory pathways showed differences between IS etiologies in CS. A 15-gene biomarker panel assembled from a derivation cohort (n = 50) correctly classified 81% of CS and 71% of SO participants in a validation cohort (n = 33). Another 15-gene panel correctly identified etiologies for 13 of 13 CS-cardioembolic and 11 of 11 CS-noncardioembolic participants upon cross-validation; 11 of 16 CS-cryptogenic participants were predicted cardioembolic.

Interpretation: We discovered unique mRNA and miRNA transcriptome architecture in CS and SO, and in CS with different IS etiologies. Cardioembolic and noncardioembolic etiologies in CS showed unique coexpression networks and potential master regulators. These may help distinguish CS from SO and identify IS etiology in cryptogenic CS patients. ANN NEUROL 2024;96:565-581.

FIGURE 1:

Flowchart of the overall study design. Flowchart describing an overview of the study design. The 141 total subjects had their peripheral blood collected within 120 h of stroke onset. mRNA and miRNA were sequenced, and differential expression (DE) analyses were conducted to identify cancer-stroke (CS) and stroke only (SO) specific responses. DE and weighted gene coexpression network analysis (WGCNA) analyses were also conducted to identify responses specific to cancer patients with strokes of cardioembolic (CS-CE) and noncardioembolic (CS-nonCE) etiologies. RNA biomarker models were derived to distinguish (1) CS from SO and (2) CS-CE from CS-nonCE. The later model was applied to cryptogenic CS patients to identify their likely stroke etiology. CO, cancer only; PCA, principal components analysis; VRFC, vascular risk factor control.

FIGURE 2:

Generation of differentially expressed gene lists. (A, B) Response to stroke in participants with cancer with stroke (CS) and with stroke only (SO) in mRNA (A) and miRNA (B). The numbers in the Venn diagrams represent the numbers of differentially expressed RNAs for each of the contrasts and their overlaps. (C, D) Response to stroke of cardioembolic (CE) and noncardioembolic (nonCE) etiology in participants with comorbid cancer in mRNA (C) and miRNA (D). In both C and D, removal of cancer only (CO) versus vascular risk factor control (VRFC) genes from CS-CE versus VRFC and CS-nonCE versus VRFC genes is presented. Additionally, generation of 5 per-gene lists each for mRNA and miRNA is presented. These 5 mRNA and 5 miRNA lists were used for subsequent biological analyses. “Up” indicates upregulated in CS etiology versus VRFC; “down” indicates downregulated in CS etiology versus VRFC. FC, fold change; FDR, false discovery rate.

FIGURE 2:

Generation of differentially expressed gene lists. (A, B) Response to stroke in participants with cancer with stroke (CS) and with stroke only (SO) in mRNA (A) and miRNA (B). The numbers in the Venn diagrams represent the numbers of differentially expressed RNAs for each of the contrasts and their overlaps. (C, D) Response to stroke of cardioembolic (CE) and noncardioembolic (nonCE) etiology in participants with comorbid cancer in mRNA (C) and miRNA (D). In both C and D, removal of cancer only (CO) versus vascular risk factor control (VRFC) genes from CS-CE versus VRFC and CS-nonCE versus VRFC genes is presented. Additionally, generation of 5 per-gene lists each for mRNA and miRNA is presented. These 5 mRNA and 5 miRNA lists were used for subsequent biological analyses. “Up” indicates upregulated in CS etiology versus VRFC; “down” indicates downregulated in CS etiology versus VRFC. FC, fold change; FDR, false discovery rate.

FIGURE 3:

Principal components analysis (PCA) and biological relevance of cancer–stroke genes. For cancer with stroke (CS) versus vascular risk factor control (VRFC; A) and CS versus stroke only (SO; B), differentially expressed (false discovery rate p < 0.05, |fold change| > 2) miRNA (top left of each panel) and mRNA (bottom left of each panel) separate groups on PCA are shown. The top 20 relevant pathways significantly enriched ( p < 0.05, −log₁₀[p] > 1.3) in the miRNA-mRNA pairs for each contrast are presented along the right of each panel. The significance threshold −log₁₀( p) = 1.3 (corresponds to p value of 0.05) is depicted by a vertical black line. Activity pattern prediction is depicted by bar shading (blue for suppression/negative Z score and orange for activation/positive Z score), where darker color represents larger |Z score|. *Statistically significant activity pattern prediction (|Z| ≥ 2). ^‡After removing cancer-only genes. (C, D) Biofunctions coagulation (C; p = 2.58 × 10⁻⁶, Z = 3.15) and activation of blood platelets (D; p = 1.48 × 10⁻⁵, Z = 2.39) were significantly associated with CS versus SO genes and were predicted activated. Fold change was colored for a range of −5 to 5. EMT = epithelial mesenchymal transition; GF = growth factor; NK = natural killer cell; PRR = pattern recognition receptors.

FIGURE 3:

Principal components analysis (PCA) and biological relevance of cancer–stroke genes. For cancer with stroke (CS) versus vascular risk factor control (VRFC; A) and CS versus stroke only (SO; B), differentially expressed (false discovery rate p < 0.05, |fold change| > 2) miRNA (top left of each panel) and mRNA (bottom left of each panel) separate groups on PCA are shown. The top 20 relevant pathways significantly enriched ( p < 0.05, −log₁₀[p] > 1.3) in the miRNA-mRNA pairs for each contrast are presented along the right of each panel. The significance threshold −log₁₀( p) = 1.3 (corresponds to p value of 0.05) is depicted by a vertical black line. Activity pattern prediction is depicted by bar shading (blue for suppression/negative Z score and orange for activation/positive Z score), where darker color represents larger |Z score|. *Statistically significant activity pattern prediction (|Z| ≥ 2). ^‡After removing cancer-only genes. (C, D) Biofunctions coagulation (C; p = 2.58 × 10⁻⁶, Z = 3.15) and activation of blood platelets (D; p = 1.48 × 10⁻⁵, Z = 2.39) were significantly associated with CS versus SO genes and were predicted activated. Fold change was colored for a range of −5 to 5. EMT = epithelial mesenchymal transition; GF = growth factor; NK = natural killer cell; PRR = pattern recognition receptors.

FIGURE 4:

mRNA pathway enrichment and miRNA-mRNA pairs in cardioembolic stroke with cancer (CS-CE) and noncardioembolic stroke with cancer (CS-nonCE). For CS-CE unique (A) and CS-nonCE unique (B) genes, the top 20 relevant pathways significantly enriched ( p < 0.05, −log ₁₀[p] > 1.3) in mRNA are presented (left side of each panel) with cancer-only genes removed. The significance threshold −log₁₀( p) = 1.3 (corresponds to p value of 0.05) is depicted by a vertical black line. Activity pattern prediction is depicted by bar shading (blue for suppression/negative Z score and orange for activation/positive Z score), where darker color represents larger |Z score|. *Statistically significant activity pattern prediction (|Z| ≥ 2). miRNA-mRNA regulatory pairs are presented on the right of each panel. Both gene lists have cancer-only genes removed. miRNA-mRNA regulatory pairs are connected, and genes involved in one of the relevant significant canonical pathways are connected to that pathway. miRNA and mRNA are colored by their fold change (green for negative and red for positive). Node positions and edge lengths are arbitrarily assigned. AD = Alzheimer disease; NER = nucleotide excision repair.

FIGURE 5:

Cell-type enrichment of cardioembolic stroke with cancer (CS-CE) and noncardioembolic stroke with cancer (CS-nonCE) lists. Cell-type specific gene list enrichment of the per-gene lists (left panel) and weighted gene coexpression network analysis (WGCNA) modules (right panel) are shown. Purple shading represents −log ₁₀( p) where 1.3 corresponds to a p value of 0.05; higher −log₁₀( p) corresponds to lower (more significant) p value. Nonsignificant hypergeometric probabilities are displayed as white cells. On the right, blue and red shading represents the beta coefficient for diagnosis in a linear regression on the module eigengene (red representing upregulated in ischemic stroke [IS], blue downregulated in IS). Enrichment of hub gene lists in cell-type specific lists is presented at the bottom. All cell type-specific lists were taken from Watkins et al other than those marked with an asterisk (*). These cell-type specific lists were taken from Chtanova et al for more comprehensive coverage of T cell-specific genes. Granulocytes, of which neutrophils are the largest percentage, are listed as neutrophils. VRFC, vascular risk factor control.

FIGURE 6:

Networks and pathway enrichment of cardioembolic stroke with cancer (CS-CE) and noncardioembolic stroke with cancer (CS-nonCE) modules. Network diagrams (left) and top 20 relevant pathways significantly enriched (right) for select modules CS-CE-DarkOrange (A) and CS-nonCE-Tan (B) are shown. On the left networks, nodes represent genes within the modules and edges represent connections between the genes (nodes). Weak connections and nodes with few connections have been removed for legibility. Hub genes are depicted with larger nodes and represent potential master regulators. Genes are shaded grey by default and only colored if they are cell type specific. Nodes with purple outlines represent genes involved in NETosis. Hubs, cell type-specific genes, and NETosis genes are labeled. Genes colored as T cell specific are from Chtanova et al T cell or T cell receptor and signaling lists or Watkins et al T helper or T cytotoxic lists. ^, For the right pathway enrichment, −log₁₀( p) of 1.3 (corresponds to p value of 0.05; significance threshold) is depicted by a vertical black line. Activity pattern prediction is depicted by bar shading (blue for suppression/negative Z score and orange for activation/positive Z score), where darker color represents larger |Z score|. *Statistically significant activity pattern prediction (|Z| ≥ 2). Embr. = embryonic; Mamm. = mammalian; Plur. = pluripotency; Rec. = receptor.

FIGURE 6:

Networks and pathway enrichment of cardioembolic stroke with cancer (CS-CE) and noncardioembolic stroke with cancer (CS-nonCE) modules. Network diagrams (left) and top 20 relevant pathways significantly enriched (right) for select modules CS-CE-DarkOrange (A) and CS-nonCE-Tan (B) are shown. On the left networks, nodes represent genes within the modules and edges represent connections between the genes (nodes). Weak connections and nodes with few connections have been removed for legibility. Hub genes are depicted with larger nodes and represent potential master regulators. Genes are shaded grey by default and only colored if they are cell type specific. Nodes with purple outlines represent genes involved in NETosis. Hubs, cell type-specific genes, and NETosis genes are labeled. Genes colored as T cell specific are from Chtanova et al T cell or T cell receptor and signaling lists or Watkins et al T helper or T cytotoxic lists. ^, For the right pathway enrichment, −log₁₀( p) of 1.3 (corresponds to p value of 0.05; significance threshold) is depicted by a vertical black line. Activity pattern prediction is depicted by bar shading (blue for suppression/negative Z score and orange for activation/positive Z score), where darker color represents larger |Z score|. *Statistically significant activity pattern prediction (|Z| ≥ 2). Embr. = embryonic; Mamm. = mammalian; Plur. = pluripotency; Rec. = receptor.

FIGURE 7:

Support vector machine (SVM) model prediction probabilities. (A) Validation cohort probabilities of the 15 mRNA SVM model to predict cancer with stroke (CS) versus stroke only (SO). Patients are distributed along the x axis and grouped into columns by their true diagnosis (CS, SO). Predicted probabilities of each stroke patient being CS (orange square) or SO (blue diamond) are presented on the y axis. Using a threshold of ≥0.5 (black dashed line), the model correctly predicted 13 of 16 (81%) CS patients and 12 of 17 (71%) SO patients. The black arrow indicates a validation cohort participant who was originally enrolled as a SO patient and was diagnosed with cancer 1 month after their stroke and thus likely had cancer at the time of stroke onset. Notably, this participant was correctly predicted as CS with high probability. (B) Predicted probability of each CS patient having cardioembolic (CE; green circles) or noncardioembolic (nonCE; purple triangles) etiology. Patients are distributed along the x axis and grouped into columns by their true diagnosis (CS-CE, CS-nonCE). Predicted probabilities of being CS-CE or CS-nonCE are presented on the y axis. For true CS-CE and true CS-nonCE patients, displayed probabilities are the result of a 10-fold cross-validation on the entire set for the best performing model using the 15 combined mRNA and miRNA predictors. (C) Predicted probability of each CS-cryptogenic patient having CE or nonCE etiology. Displayed probabilities are the result of deploying the best performing model on this group. In B and C, using a threshold of ≥0.5 (black dashed line), 13 of 13 (100%) CS-CE participants were correctly predicted and 11 of 11 (100%) CS-nonCE participants were correctly predicted. This cutoff predicted 11 of the CS-cryptogenic strokes to be CE (69%) and 5 to be nonCE (31%). Using a more conservative threshold of ≥0.7 (yellow dashed line), 12 of 13 (92%; 1 uncertain, 0 incorrect) CS-CE participants were correctly predicted and 8 of 11 (73%; 3 uncertain, 0 incorrect) CS-nonCE cases were correctly predicted. This cutoff predicted 8 of the CS-cryptogenic strokes to be CE (50%) and 3 to be nonCE (19%), and 5 were uncertain (31%).