Small RNA Changes in Plasma Have Potential for Early Diagnosis of Alzheimer's Disease before Symptom Onset

Abstract

Alzheimer's disease (AD), due to its multifactorial nature and complex etiology, poses challenges for research, diagnosis, and treatment, and impacts millions worldwide. To address the need for minimally invasive, repeatable measures that aid in AD diagnosis and progression monitoring, studies leveraging RNAs associated with extracellular vesicles (EVs) in human biofluids have revealed AD-associated changes. However, the validation of AD biomarkers has suffered from the collection of samples from differing points in the disease time course or a lack of confirmed AD diagnoses. Here, we integrate clinical diagnosis and postmortem pathology data to form more accurate experimental groups and use small RNA sequencing to show that EVs from plasma can serve as a potential source of RNAs that reflect disease-related changes. Importantly, we demonstrated that these changes are identifiable in the EVs of preclinical patients, years before symptom manifestation, and that machine learning models based on differentially expressed RNAs can help predict disease conversion or progression. This research offers critical insight into early disease biomarkers and underscores the significance of accounting for disease progression and pathology in human AD studies.

Keywords: Alzheimer’s disease; biomarker; extracellular vesicle; mild cognitive impairment; plasma; small RNA sequencing.

Figure 1

Plasma EVs are a rich source of small RNA compared with the EV−depleted fraction. (a) 117 EV samples passed quality sequencing metrics and were used for this study. Age ranges were evenly distributed among the experimental groups. Mean cognitive test scores decreased with disease severity. (b) For 86 samples, the small RNA was also sequenced from the EV-depleted fraction, and 46 samples fulfilled sequencing quality criteria. miRNA, tRNA, YRNA, piRNA, and protein-coding biotypes were identified in both isolations. The EV-enriched portion consistently had greater transcript diversity (unique RNAs). (c) The majority of transcripts were found in both EV-enriched and EV-depleted samples, with the highest number of unique transcripts found in EVs. (d) A display of the 10 most abundant transcripts in each isolation method, EV-enriched or EV-depleted, shows enrichment for YRNA4P10 in EVs and miR-486 in EV-depleted samples. (e) EV-enriched RNA had lower variability compared with the EV-depleted RNA, as assessed by the % of CVs of all genes with normalized expression levels > 10 counts in >50% of the samples’ most expressed genes, regardless of disease status. Statistical analysis (b,e) was carried out via t-test.

Figure 2

Postmortem diagnosis and disease progression information complicates the initial cohort sample grouping. (a) Autopsy findings for a fraction of our cohort show that initial diagnosis does not always correspond with postmortem brain pathology and that several individuals from the control and MCI groups progressed in disease status after sample collection. (b) Gene expression comparison of controls with confirmed normal vs. PART pathology is depicted as a volcano plot, where red dots represent sequences that are log(2)fc > 1 and a p-value < 0.05. (c) Samples with unknown postmortem diagnosis from the control group (blue) were analyzed using the pathology-based model (asterisk denotes the 0.5 ELNET model probability cutoff used for sample inclusion criteria). The waterfall plot across the top of the heatmap demonstrates ELNET model confidence in assigning groups (0 to 1, with 0.5 being no confidence in the assignment), with dots colored to the diagnosis group, while the sequences picked by ELNET are displayed on the right of the heatmap. The naming scheme for RNA biotypes is described in Section 2 (RNASeq data analysis). PART = primary age-related tauopathy, DLBs = dementia with Lewy bodies, FTD = frontotemporal dementia, and DX = diagnosis.

Figure 3

Differentially expressed small RNA transcripts in MCI and AD are both distinct and overlapping. Differential gene expression analysis is shown as volcano plots between control and AD (a), control and MCI (d), and MCI and AD (g) groups. Box plots of the top 3 most significantly upregulated and downregulated transcripts are shown for each comparison: control vs. AD (b), control vs. MCI (e), and MCI vs. AD (h). ELNET was used to select features from the DE gene list to construct a classification algorithm for each comparison: control vs. AD (c), control vs. MCI (f), and MCI vs. AD (i). (j) A Venn diagram of the DE sequences between AD and MCI displays the overlapping transcripts, all of which are concordant. ↑ arrows denote transcripts that are upregulated relative to the control group, while ↓ arrows mark downregulated transcripts. The p-value (unadjusted) for each comparison is displayed for each box plot (b,e,h). The waterfall plot across the top of the heatmap demonstrates ELNET model confidence in assigning groups (0 to 1, with 0.5 being no confidence in the assignment), with dots colored to the diagnosis group.

Figure 4

Changes in EV-RNA predate symptom onset in the preclinical AD cohort. (a) Differential gene expression analysis between the control non-converters and control pre-AD (C to AD) found 384 significant transcripts. (b) Representative transcripts that are significantly up/downregulated in the C to AD group compared with non-converter controls are depicted as box plots, with expression levels in control, C to AD, and AD shown. (c) A third of DE transcripts in the C to AD group are concordant and overlapping with AD DE transcripts. (d) Regression analysis by time to disease conversion with corresponding R² and p-values are displayed, where the X axis marks the years to approaching disease diagnosis (e.g., -4 denotes a participant diagnosed with AD 4 years after the blood draw); the median expression level for the transcript in control non-converters (blue) and AD (red) is provided as a reference. (e) The ELNET control vs. AD model is used to determine whether the C to AD samples can be accurately predicted, with the top bar marking years elapsed from sample collection to disease diagnosis and asterisks indicating a confirmed postmortem AD diagnosis at autopsy. In the Venn diagram, ↑ arrows denote transcripts that are upregulated relative to the control group, while ↓ arrows mark downregulated transcripts. Significance is denoted in the box plots as ** (p < 0.01), and n.s. (not significant). The waterfall plot across the top of the heatmap demonstrates ELNET model confidence in assigning groups (0 to 1, with 0.5 being no confidence in the assignment), with dots colored to reference the diagnosis group.

Figure 5

Converters to MCI express MCI-associated transcripts before diagnosis. (a) A total of 621 DE transcripts are significant when comparing the C to MCI group with the control group. (b) Box plots of representative significantly upregulated and downregulated sequences are displayed for the control, C to MCI, and MCI groups. (c) A Venn diagram comparing significant transcripts in each group relative to the control group shows 100 are overlapping, with the majority concordant. (d) Regression analysis plots by time to disease conversion with corresponding R² and p-values are displayed, where the X axis of years to disease diagnosis is shown; the median expression level for the transcript in control non-converters (blue) and MCI (yellow) is provided as a reference. (e) A third of the MCI converters were identified as MCI by the ELNET model, while another third cluster near the midpoint of the model. In the Venn diagram, ↑ arrows denote transcripts that are upregulated relative to the control group, while ↓ arrows mark downregulated transcripts. Significance is denoted in the box plots as * (p < 0.05), and n.s. (not significant). The waterfall plot across the top of the heatmap demonstrates ELNET model confidence in assigning groups (0 to 1, with 0.5 being no confidence in the assignment), with dots colored to reference the diagnosis group. Asterisks across the waterfall plot indicate C to MCI converters with a confirmed AD diagnosis at autopsy.

Figure 6

MCI participants who progressed to AD are accurately identified by their gene expression signature. (a) Comparing the MCI to AD converters to the MCI group that did not convert revealed 132 significantly differentially expressed transcripts. (b) Box plots show expression levels for representative DE sequences in the MCI, MCI to AD, and AD groups. (c) Overlap between significantly different transcripts in MCI to AD (relative to MCI) and AD (relative to MCI) is restricted to 15 genes as shown in a Venn diagram. (d) The ELNET model distinguishing AD from MCI identified most converters as AD. In the Venn diagram, ↑ arrows denote transcripts that are upregulated relative to the control group, while ↓ arrows mark downregulated transcripts. Significance is denoted in the box plots as * (p < 0.05), and n.s. (not significant). The waterfall plot across the top of the heatmap demonstrates ELNET model confidence in assigning groups (0 to 1, with 0.5 being no confidence in assignment), with dots colored to reference the diagnosis group. Asterisks across the waterfall plot indicate MCI to AD converters with a confirmed AD diagnosis at autopsy.