Identification of novel diagnostic panel for mild cognitive impairment and Alzheimer’s disease: findings based on urine proteomics and machine learning

This study was a cross-sectional study that enrolled participants from China-Japan Friendship Hospital from April 2022 to November 2022. A total of 162 participants, over 50 years old, including 57 AD patients, 43 MCI patients, and 62 CN subjects were included in the final analysis. Risk factors were collected, and APOE genotypes were classified into ε4 carriers and non-carriers. Sex, living status, education, smoking status, and family histories matched among the groups. Besides, the distribution of hypertension, diabetes, hyperlipidemia, heart diseases, and cerebrovascular diseases among the three groups did not reach statistical significance. Age, the most important risk factor of AD, was more senior in the AD group compared to the CN group. APOE ε4, the main genetic risk factor for sporadic AD, was more prevalent in the AD and MCI groups compared with the CN group. The overall information is summarized in Table 1.

All subjects underwent medical history collection, a battery of neuropsychological assessments and apolipoprotein E (APOE) genotype test. Most individuals underwent quantitative electroencephalography (qEEG) and magnetic resonance imaging (MRI). The study protocol was approved by the China-Japan Friendship Hospital ethics committee and institutions (Ethics ID: 2020–31-Y06-32). Consent forms were obtained from all participants.

AD is clinically diagnosed using the 2011 National Institute on Aging-Alzheimer’s Association (NIA-AA) criteria [5]. The contents are as follows: (1) meet the core clinical criteria including interference with the ability to complete daily activities and a decline from previous levels, (2) characterized by insidious onset and clear-cut history of decline of cognition, and (3) excluding dementia due to other etiologies.

MCI is defined with the 2011 NIA-AA diagnostic criteria [22], as the following shows: (1) concern about a cognition decline compared with the previous status, reported by the patient himself, the informant, or a skilled physician; (2) decline in at least one cognitive domain after age and education adjustment; (3) maintenance of independent function in daily life activities; and (4) not meeting the diagnostic criteria for dementia.

CN controls were those who performed normally on the standardized neuropsychological tests and with or without cognitive complaints or concerns during the structured interview.

Briefly, MMSE cutoff points for dementia/non-dementia were 16/17 for illiterate, 19/20 for individuals with 1–6 years of education, and 23/24 for individuals with 7 or more years of education [23]. The ADL cutoff was 26. The definition of cognitive decline in domains was a decrease of more than 1.5 standard deviations in at least one test. Besides, medical history and imaging evidence were taken into consideration. In summary, patients were diagnosed according to the clinical criteria based on comprehensive assessments.

The exclusion criteria are as follows: (1) cognitive decline caused by severe psychiatric disorders or mental retardation; (2) cognitive impairment caused by other neurological diseases, such as trauma, stroke, tumor, parkinsonism, encephalitis or epilepsy, or other types of dementia, such as frontotemporal dementia (FTD), Lewy body dementia (LBD), and vascular dementia (VaD); (3) cognitive impairment caused by diseases of other systems such as severe anemia and thyroid disorders; (4) a history of urinary system disorders, malignant tumor, or other severe diseases; and (5) inability to cooperate in completing neuropsychological tests or incomplete clinical data.

The neuropsychological test battery included measures of global cognition and cognitive performance in the domains of memory, executive function, attention, language, and visuospatial ability. Participants were administered the Mini-Mental State Examination (MMSE) and Montreal Cognitive Assessment (MoCA) for global cognition. The Activity of Daily Living Scale (ADL) was used for accessing the function ability during daily life. The Rey Auditory Verbal Learning Test-immediate recall (RAVLT-I) and Rey Auditory Verbal Learning Test-delayed recall (RAVLT-D) were administered to assess memory; Digit Span Test (DST)-Backward and Stroop Color and Word Test (SCWT) were used for accessing executive function; DST-Forward and Symbol Digit Modalities Test (SDMT) were used for accessing attention; Boston Naming Test (BNT) and Verbal Fluency Test (VFT) were administered to assess language. In addition, the Clock Drawing Test (CDT) and Rey Complex Figure Test (RCFT) were utilized to assess visuospatial ability. The above scales have been applied in clinical practice and published in previous articles from our team [24].

A midstream of random urine was collected and stored at − 80 °C. A biosafety level II lab was used to prepare samples. The pellet from the urine was obtained after being centrifuged at 176,000 g for 1 h and then was re-suspended using 40 μL of resuspension buffer containing 50 mmol L⁻¹ Tris–HCl, 250 mmol L⁻¹ sucrose, pH 8.5, and then reduced with 50 mmol L⁻¹ dithiotheitol (DTT) at 65 °C for 30 min. After adding 160 μL wash buffer (10 mmol L⁻¹ Tris–HCl, pH 7.4, 100 mmol L⁻¹ NaCl), a second ultracentrifugation at 176,000 g was performed for 30 min. The pellet was re-suspended with 30 μL 50 mM NH₄HCO₃, heated for 3 min at 95 °C, cooled to room temperature, and then digested by trypsin at a protease-to-protein ratio of 1:100 (w/w), incubating overnight at 37 °C.

The digested peptides were vacuum-dried in a SpeedVac. Then, samples were stored at − 80 °C until further use. Peptide samples were re-dissolved in 0.1% formic acid (FA)-H₂O. One-microgram peptide samples were loaded onto a trap column (100 μm × 2 cm, homemade; particle size, 3 μm; pore size, 120 Å; SunChrom, USA). Solvent A was 0.1% FA in H₂O, and solvent B was 0.08% FA and 20% H₂O in Acetonitrile (ACN). Peptides were separated by a homemade silica microcolumn (150 μm × 10 cm, particle size, 1.9 μm; pore size, 120 Å; SunChrom, USA) with a gradient of 5–35% solvent B at a flow rate of 800 nL/min for 30 min. Liquid chromatography coupled to tandem mass spectrometry (LC–MS/MS) was performed on a Q Exactive HF-X mass spectrometer (Thermo Fisher Scientific, USA). The instrument was run in the data-dependent acquisition (DDA) mode. The whole scan was processed in the Orbitrap from m/z 300–1400 at a resolution of 60,000 with an automatic gain control (AGC) target of 3e⁶ and a 20-ms maximum injection time. With a normalized collision energy of 27%, the top 40 most intense ions in each scan cycle were chosen for high-energy collision dissociation (HCD) fragmentation. For the MS/MS scan, the fragment ions were identified in the Orbitrap with a resolution of 7500, an AGC target of 5e⁴, a maximum injection time of 12 ms, and a dynamic exclusion of 15 s. Trypsin digests of 293 T cells were used to prepare quality control samples which were then routinely evaluated to determine the sensitivity and reproducibility of LC–MS/MS. The mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium (http://​proteomecentral.​proteomexchange.​org) via the iProX partner repository [25, 26] with the dataset identifier PXD044672.

The Firmiana platform was used to process the mass spectrometry data [27]. The MASCOT search engine (Matrix Science, version 2.3.01) was used to identify proteins in the NCBI human RefSeq protein database (published on 04/07/2023, 33,118 entries). Precursor ion mass tolerance was set to 20 ppm, while product ion mass tolerance was set at 0.05 Da. Trypsin digestion may miss at most one cleavage. Dynamic modifications included methionine oxidation and N-terminal acetylation. For the following analyses, only ≥ 1 unique and strict peptide, ≥ 2 strict peptides (ion score > 20), or ≥ 3 strict peptides with protein levels equal to 1% FDR were employed. Protein quantification was carried out using the intensity-based absolute quantification (iBAQ) algorithm [28]. We converted the iBAQ to the fraction of total (FOT) to normalize the differences in sample amounts [29], which was calculated by the iBAQ value of each protein divided by the total iBAQ of the sample, multiplied by 10⁵. All missing values were replaced with zeros. Proteins detected in more than 50% of the samples were included for further analysis. A total of 608 proteins were retained, and the imputation of missing values was based on the k-nearest neighbor (KNN) method using the “Wu Kong” platform (https://​www.​omicsolution.​org/​wkomics/​main/​).

SPSS 23.0 was used for statistical analysis. The Shapiro–Wilk test was used to examine the normality of quantitative data. The mean (x ± s) was used for the description of normal data while non-normal data used median (P25, P75). Analysis of variance (ANOVA) was used for normal data mean comparison while the Kruskal–Wallis H test was utilized for non-normal data distribution comparison. For post hoc comparisons, p-values were Bonferroni-corrected. Besides, Pearson’s chi-square test or Fisher’s exact probability was used for the comparison of the proportions of categorical variables. Statistical significance was defined as a two-tailed p-value < 0.05. To construct a protein–protein interaction (PPI) network, we used the stringApp in cytoscape, and BiNGo in cytoscape was used for Gene Ontology (GO) enrichment with Benjamini–Hochberg corrected p-value < 0.05. In parallel, R (4.1.0) was used for bioinformatics analysis. Differential urinary proteins were filtered utilizing limma package [30] with a threshold of p < 0.05 and the absolute value of log2 fold change (log2FC) > 0.58 after log₂ transformation and normalization. Heatmap was presented using pHeatmap [31], and the volcano plot was presented using EnhancedVolcano [32]. The expression levels of selected proteins were shown in the boxplot by ggpubr [33] package. Gene set enrichment analysis (GSEA) was used to investigate various GO terms and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways that might be related with AD or MCI when compared to CN in all proteins. clusterProfiler package [34, 35] was utilized for enrichment analysis while enrichplot package [36] was utilized for visualization. Moreover, the corrplot [37] package was used for the visualization of the correlation relationship.

In order to distinguish AD from CN and MCI from CN, machine learning was utilized to determine the best multivariate signatures, which included both proteins and demographic information (age and APOE 4 status) as input parameters. The classifier consisted of feature selection and classifiers [38]. Briefly, the dataset was separated into a training set (0.7) and a test set (0.3). The least absolute shrinkage and selection operator (LASSO) was utilized to select the “n” top input variables that best differentiated AD or MCI diagnostic groups with minimum mean square error (MSE). On top of these “n” characteristics, support vector machine (SVM) classifiers were built to forecast the result under tenfold cross-validation. Linear, polynomial, radial, and sigmoid kernel functions were compared. Accuracy and area under the curve (AUC) (receiver operating characteristic (ROC) curve) were used for the diagnostic value evaluation when testing the model in the test set.

Table 2 presented the cognitive assessment results, percentage of abnormal qEEG, and medial temporal lobe atrophy (MTA) scales of each group. As for neuropsychological assessments, the results showed that there were significant differences among the three groups using the Kruskal–Wallis H test (p < 0.001). For post hoc comparisons, there were differences between the AD and CN groups as well as the MCI and CN groups in global cognition as indicated by MMSE and MoCA, memory domain as indicated by RAVLT-I and RAVLT-D, executive function as indicated by DST-Backward and SCWT, attention domain when indicated by SDMT, language as indicated by VFT and BNT, and visuospatial processing as indicated by CDT and RCFT. There were only differences between the AD and CN groups in ADL and DST-Forward. The individual basic information and results of neuropsychological tests for each participant were uploaded as Additional file 6: Table S1. Besides, the percentage of abnormal qEEG was higher in the AD group than in the CN group (p < 0.05). In parallel, there were differences between AD and CN in MTA scales (p < 0.001) in which the left-sided hippocampus atrophy of patients was more severe.

The proteomics analysis performed was a LFQ quantitative analysis in DDA mode. In total, 3366 proteins were identified. Only the protein that could be detected in the majority (more than 50%) of the samples was included, and at last, a total of 608 proteins were included for further analysis (Additional file 7: Table S2). After imputing missing values using the KNN method, a complete expression matrix was constructed. GSEA results of all proteins were shown in Additional file 1: Fig. S1. In AD samples, a number of biological pathways and processes related to the immune system were enriched, whereas in MCI samples, a number of biological pathways and processes related to metabolism were enriched.

The protein expression levels of the samples were log2 transformed and normalized. Differential urinary proteins were filtered with a threshold of p < 0.05 and the absolute value of log₂ fold change (log₂FC) > 0.58. Compared to the CN group, significantly differential proteins were filtered in the AD group and MCI group by setting the threshold above. A table with the log₂FC, p-values, and Benjamini–Hochberg corrected p-values of the 608 proteins included in the analysis was uploaded as Additional file 8: Table S3. The expression of the differential proteins in the AD group was displayed as a heatmap and a volcano plot (Fig. 1A, B) while the expression of the differential proteins in the MCI group was shown in Fig. 1C, D. There were 33 significantly differential proteins between the AD and CN groups among the 608 proteins included in the analysis, including 21 upregulated ones and 12 downregulated ones. In parallel, there were 15 significantly differential proteins between the MCI and CN groups among the 608 proteins included in the analysis, including 7 upregulated ones and 8 downregulated ones. These differential proteins were respectively inputted in LASSO for diagnostic panel selection. GSTA1 was downregulated in both AD and MCI while EHD4 and C9 were both upregulated in AD and MCI urine samples. The differential proteins between the AD and MCI groups were shown in Additional file 2: Fig. S2. A Venn diagram showing the intersection between the groups was shown in Additional file 3: Fig. S3.

With the help of stringApp in cytoscape, differential proteins were inputted, and the PPI network was constructed (Fig. 2). While proteins with an unknown 3D structure were represented by empty nodes, those with a known or predicted 3D structure were represented by filled nodes. The red nodes indicated upregulated proteins, and the blue nodes indicated downregulated proteins. The size reflected relative fold change when compared to CN. Besides, 33 biological processes in the AD-CN group and 67 biological processes in the MCI-CN group mainly related to the immune system and metabolism were enriched (Benjamini–Hochberg corrected p-value < 0.05). The enrichment networks are shown in Additional file 4: Fig. S4, and relative details are shown in Additional file 9: Table S4.

Based on previous analysis, we extracted all differential proteins (33 in the AD-CN group and 15 in the MCI-CN group) plus age and APOE ε4 status to construct the LASSO model. For the AD-CN model, 13 proteins, age, and APOE ε4 status were identified when MSE reached minimum with the value of lambda (min) equaling 0.03225 (Fig. 3A). DDC, CTSC, EHD4, GSTA3, SLC44A4, GNS, GSTA1, ANXA4, PLD3, CTSH, HP, RPS3, CPVL, age, and APOE ε4 status were included in AD diagnostic panel. The boxplots showed the expression value of these proteins (Fig. 3B). Similarly, for the MCI-CN model, 10 proteins, age, and APOE ε4 status were identified when MSE reached minimum with the value of lambda (min) equaling 0.0191 (Fig. 3C). TUBB, SUCLG2, PROCR, TCP1, ACE, FLOT2, EHD4, PROZ, C9, SERPINA3, age, and APOE ε4 status were included in the MCI diagnostic panel. The boxplots showed the expression value of these proteins (Fig. 3D). EHD4 was considered valuable for both AD and MCI diagnosis.

Based on LASSO results, we built SVM classifiers with tenfold cross-validation to investigate the ideal multivariate signatures that distinguished AD or MCI from CN. After training in training sets, we compared the relative indicators using different kernel functions in SVM. Radial achieved the highest predictive value with an accuracy of 0.9881, an F1 measure of 0.9876, and an AUC of 0.9739 in the AD-CN group and an accuracy of 0.973, an F1 measure of 0.9688, and an AUC of 0.9985 in the MCI-CN group in the training set. The model achieved a high predictive value with an accuracy of 0.7714, an F1 measure of 0.6923, and an AUC of 0.8824 in the AD-CN group and an accuracy of 0.8387, an F1 measure of 0.7386, and an AUC of 0.8143 in the MCI-CN group in the test set. Figure 4 shows the ROC curve in the training sets and test sets either in the AD-CN group (Fig. 4A, B) or in the MCI-CN group (Fig. 4C, D).

Diagnostic proteins were found to be correlated with cognitive tests, although most weakly (Fig. 5). Significant labels were shown on the dots. Among 22 diagnostic proteins, DDC, CTSC, EHD4, GNS, GSTA1, RPS3, PROCR, and SERPINA3 were significantly correlated with more than half of cognitive tests while GSTA3, SLC44A4, ANXA4, PLD3, CTSH, CPVL, SUCLG2, TCP1, ACE, PROZ, and C9 were significantly correlated with less than half cognitive tests. Nevertheless, none of the correlations between HP, TUBB, or FLOT2 and cognitive domains reach significance. The relative ρ and p were shown in Additional file 10: Table S5, and scatter dot plots were shown in Additional file 5: Fig. S5.