Improved protocol for measurement of plasma β-amyloid in longitudinal evaluation of Alzheimer's Disease Neuroimaging Initiative study patients

Abstract

Background

The interassay variability and inconsistency of plasma β-amyloid (Aβ) measurements among centers are major factors precluding the interpretation of results and a substantial obstacle in the meta-analysis across studies of this biomarker. The goal of this investigation was to address these problems by improving the performance of the bioanalytical method.

Methods

We used the Luminex immunoassay platform with a multiplex microsphere-based reagent kit from Innogenetics. A robotic pipetting system was used to perform crucial steps of the procedure. The performance of this method was evaluated using two kit control samples and two quality control plasma samples from volunteer donors, and by retesting previously assayed patient samples in each run. This setup was applied to process 2454 patient plasma samples from the Alzheimer's Disease Neuroimaging Initiative study biofluid repository. We have additionally evaluated the correlations between our results and cerebrospinal fluid (CSF) biomarker data using mixed-effects modeling.

Results

The average precision values of the kit controls were 8.3% for Aβ_1-40 and 4.0% for Aβ_1-42, whereas the values for the plasma quality controls were 6.4% for Aβ_1-40 and 4.8% for Aβ_1-42. From the test–retest evaluation, the average precision was 7.2% for Aβ_1-40 and 4.5% for Aβ_1-42. The range of final plasma results for Alzheimer's Disease Neuroimaging Initiative patients was 13 to 372 pg/mL (median: 164 pg/mL) for Aβ_1-40 and 3.5 to 103 pg/mL (median: 39.3 pg/mL) for Aβ_1-42. We found that sample collection parameters (blood volume and time to freeze) have a small, but significant, influence on the result. No significant difference was found between plasma Aβ levels for patients with Alzheimer's disease and healthy control subjects. We have determined multiple significant correlations of plasma Aβ_1-42 levels with CSF biomarkers. The relatively strongest, although modest, correlation was found between plasma Aβ_1-42 levels and CSF p-tau₁₈₁/Aβ_1-42 ratio in patients with mild cognitive impairment. Plasma Aβ_1-40 correlations with CSF biomarkers were weaker and diminished completely when we used longitudinal data. No significant correlations were found for the plasma Aβ_1-42/Aβ_1-40 ratio.

Conclusions

The precision of our robotized method represents a substantial improvement over results reported in the literature. Multiple significant correlations between plasma and CSF biomarkers were found. Although these correlations are not strong enough to support the use of plasma Aβ measurement as a diagnostic screening test, plasma Aβ_1-42 levels are well suited for use as a pharmacodynamic marker.

1. Introduction

The formation and aggregation of β-amyloid (Aβ) into plaques in the brain is one of the major pathologic events associated with the development of Alzheimer's disease (AD). Hence, the research community has great interest in evaluating Aβ, particularly Aβ_1-42 (the least soluble member of the family of Aβ peptides), as a biomarker of the disease.

Measurements of Aβ_1-42 levels in cerebrospinal fluid (CSF), along with the tau peptide and its phosphorylated counterpart p-tau₁₈₁ (products of neuronal degeneration and the main components of tangles also found in autopsied brains of AD patients), already have a prominent role in the early detection of AD pathology. The clinical performance for the detection of AD using these three CSF biomarkers is relatively well characterized. Threshold levels were established for them using CSF samples collected premortem from AD patients who had autopsy-based diagnoses [1].

Nevertheless, owing to the invasiveness of CSF sampling, there is an increasing interest in the development and validation of biomarkers measurable in blood or plasma, and many researchers are investigating whether plasma Aβ could be one of them. Unfortunately, the reports are contradictory in many instances, where some researchers report higher and others report lower plasma Aβ levels in AD patients compared with control subjects. Moreover, the correlations reported for plasma Aβ levels are also conflicting; some authors report a higher plasma Aβ_1-42/Aβ_1-40 ratio in patients with less education [2], whereas others report a higher ratio in patients with the highest education level [3].

A substantial contributor to this lack of consistency, as many investigators agree, is the poor within- and between-assay precision and the low accuracy of the immunoassay test methods used in the analysis of Aβ. Additionally, there is a general difficulty of comparing the results from different studies, as the reports differ substantially from one another in the reported ranges and threshold values. These additional confounding factors hinder the interpretation of the data [4] and make the meta-analysis of Aβ results challenging.

The procedure presented in this study is a step toward further standardization and improved performance by implementing automation in critical steps of the procedure. The goal of this work was to improve the performance of the immunoassay method for measuring plasma Aβ_1-40 and Aβ_1-42 levels in terms of run-to-run reproducibility, and to apply this improved methodology to measure plasma Aβ levels collected at baseline and subsequent annual visits in the Alzheimer's Disease Neuroimaging Initiative (ADNI) study patients and compare them with CSF results.

2. Methods

2.1. ADNI study cohort

The ADNI study (http://adni.loni.ucla.edu) was launched in 2004 by the National Institute of Aging, the Foundation for the National Institutes of Health, and by a group of private–public partners as a precompetitive AD biomarker consortium that now is in its seventh year and will continue for approximately another 6 years. Enrollment target in the first phase of ADNI (ADNI-1) was 800 participants: 200 healthy control subjects, 400 subjects with amnesic mild cognitive impairment (MCI), and 200 subjects with mild AD (although this cohort has been enlarged in the second phase of ADNI [ADNI-2]). The ADNI protocol was approved by the human studies committee at the 58 institutions participating in the ADNI study in the United States and Canada. Written and verbal informed consent was obtained from participants at screening and enrollment. At sample collection, participants were ≥60 years of age and in good general health, with no other neurological, psychiatric, or major medical diagnoses that could contribute importantly to dementia. Further details regarding the ADNI study, including participant selection procedures and complete study protocol, have been presented elsewhere and can be found online at http://www.alzheimers.org/clinicaltrials/fullrec.asp?PrimaryKey=208. A schematic depiction of ADNI biomarker specimen collection, analytical procedures, and data flow is shown in Fig. 1.

Fig. 1.

Schematic diagram of the ADNI biomarker specimen collection, bioanalytical procedures, and data flow.

2.2. Samples

Plasma samples used in this study were obtained from the ADNI biofluid repository at the University of Pennsylvania. All ADNI study patients having a baseline and at least one scheduled follow-up plasma samples were included in the study (n = 722; 422 men [58.4%] and 300 women [41.6%], with 2454 samples). The mean age of all patients at baseline was 75.3±6.79 years. Patients were age-matched across genders (75.7 ± 6.73 years in male and 74.8 ± 6.87 years in female subjects, P =.12) and across diagnosis groups (75.9 ± 5.03 years in AD, 75.0 ± 7.3 years in MCI, and 75.3 ± 7.57 years in cognitively normal (NL) patients, P =.78).

The plasma samples were collected at the participating ADNI centers. After overnight fasting, plasma was collected in the morning by venipuncture into Vacutainer tubes (Becton Dickenson, Franklin Lakes, NJ) containing potassium K3 ethylene tetraacetate as an anticoagulant. After centrifugation, samples were placed in transfer tubes (13 mL polypropylene, Sarstedt Inc., Newton, NC, catalog number 60.541), frozen, and shipped on dry ice to the UPenn Biomarker Core Laboratory, where they were stored temporarily at −80°C. The average time from blood collection to freezing of plasma for shipment was 67 ± 41 minutes (95% confidence interval [CI]: 21–180 minutes). Within several weeks of receipt, the samples were thawed, aliquoted by 500 μL into aliquot tubes (1.5 mL polypropylene, Thermo Fisher Scientific, Waltham, MA, catalog number 05-408-129), and stored at −80°C pending biochemical analyses.

2.3. Analytical method

The levels of Aβ_1-40 and Aβ_1-42 peptides were quantified in plasma using “INNO-BIA plasma Aβ forms” (module A kit; Innogenetics NV, Ghent, Belgium), a multiplex microsphere-based xMAP technology research use-only reagent kit, on a Luminex 100 with IS version 2.3 software (Luminex, Austin, TX). The kit uses monoclonal antibodies (MAbs) covalently coupled to spectrally specific fluorescent beads to detect Aβ_1-42 (MAb 21F12) and Aβ_1-40 (MAb 2G3). Full details of the implementation of this kit on the Luminex platform are described elsewhere [5].

The same lot number of the kit reagents as well as of calibrators and controls was used throughout the study to avoid possible lot-to-lot variability. Reagents were prepared according to the manufacturer's instructions, except for the coated beads' dilution, which was decreased by a factor of two (to ×50, a twofold increase in the number of beads per well) to ensure a consistently sufficient bead count number (100). Samples and calibration standards were run in duplicate, and every set of samples from a single subject was analyzed within the same run.

2.4. Sample preparation procedure

Sample preparation was performed using an automated pipetting system epMotion 5075 (Eppendorf, Hamburg, Germany), except for the filter plate draining step, which was performed using a stand-alone vacuum manifold (Milli-pore, Brussels, Belgium). The schematic diagram of the setup and automation is displayed in Fig. 2. Further details, including epMotion bench layout and program, can be found in Fig. S1 and Tables S1 and S2.

Fig. 2.

Schematic diagram of the analytical procedure. The steps inside the box are performed by the epMotion 5075 robotic pipettor.

Briefly, thawed plasma samples were centrifuged at room temperature (3000 rpm for 10 minutes) to remove solid particles that could cause problems during further steps in the analyses. Subsequently, a 1:3 dilution of samples was prepared by adding 160 μL of diluent to 80 μL of plasma. Then, a 96-well filter plate was washed with 225 μL of washing solution per well and drained using the vacuum manifold, followed by the addition of a mixture of the coated beads (100 μL/well) to all wells. The wells were drained, and “conjugate 1” (25 μL/well) was added, followed by 75 μL/well of blanks, standards, quality controls (“ConA” and “ConB”), diluted quality control plasma samples (collected from individual volunteer donors), and diluted patient plasma samples. After an overnight incubation at 2°C to 8°C in the dark (plate covered with aluminum foil, continuous shaking), the filter plate was drained and washed three times. Subsequently, phycoerythrin-labeled streptavidin was added (100 μL/well), and the plate was incubated on a shaker for another 1 hour in the dark at room temperature. After the subsequent washing step, reading solution (phosphate-buffered saline, 100 μL/well) was added and again incubated at room temperature for 3 minutes in the dark. The assay was read on a Luminex 100 IS instrument. For each set of micro-spheres, 100 beads were analyzed, and the median fluorescence intensity (MFI) was used for quantification.

2.5. Evaluation of method performance

To closely evaluate the performance of this procedure, two additional quality control samples (“QC1” and “QC2”) were run on each plate apart from the kit quality controls (ConA and ConB). These samples were plasma aliquots from volunteer donors. Additionally, on every plate except the first one, we included 2 to 3 never-before-thawed aliquots of samples that were run previously within this project, thus implementing a continuous test–retest evaluation.

2.6. Assay data processing

For efficient processing of large amounts of data from the multiple runs performed using the aforementioned procedure, we developed a set of scripts for the “R” statistical analysis language (R version 2.12.2 [6]). The scripts were collected in a library named “immunoassay,” currently available from R-Forge (version 0.3, http://immunoassay.r-forge.r-project.org) [7]. An important functionality of this library from our perspective was the ability to batch process the data from multiple Luminex runs and create a cumulative table of results.

In brief, the functions in the immunoassay library were used to load run data into R. Only the tables of bead counts and MFI were loaded, along with basic run information. Kit information (concentrations of analytes in standards and kit quality controls) was loaded from a separate table. Subsequently, the fitting of sigmoidal models to the calibrators was performed using five-parametric logistic model with either “1/y” or “sqrt(1/y)” weighting, and the fitting function was allowed to automatically select the best fit based on the criteria of residual sum of squares [8]. Final results for plasma samples, including QC1 and QC2, were multiplied by the factor of 3 because of the initial step of dilution of the plasma samples.

Subsequently, the final results were evaluated for acceptability based on three predefined criteria: minimum bead counts of 20, minimum MFI greater than or equal to the MFI of the lowest calibrator, and maximum percent coefficient of variation (%CV) between replicates of 20%. Results not meeting these criteria were removed from the data set.

2.7. Statistical analysis of results

For statistical analysis of results, the R software package version 2.12.2 was used [6]. The results from the “test–retest” experiment were analyzed with the Passing–Bablok [9] and Deming method [10] using the “MethComp” library version 1.3 [11]. By-patient means of Aβ results across all visits were used in the analysis of time-independent variables [12]. Analysis of time-dependent variables in the longitudinal data set was performed with mixed-effects models using the library “nlme” version 3.1–98 [13]. Determination of statistical significance of model terms was done by comparing the “maximum likelihood” estimated models with and without those terms using analysis of variance [14]. Summary results are presented as mean ± standard deviation, and statistical tests for significance were performed using nonparametric Kruskal–Wallis rank sum test, unless noted otherwise. Correlations were evaluated using Spearman test. A P value of .05 was assumed to be the threshold for statistical significance.

3. Results

3.1. Analytical performance

A total of 2454 plasma samples from the ADNI study patients were run in duplicate using 76 plates. The first 24 plates were run with calibrators St7 and St8 prepared by sequential 1:3 dilution of the St6 calibrator to obtain concentrations of, respectively, 5 and 1.6 pg/mL for Aβ_1-40 and of 1.6 and 0.53 pg/mL for Aβ_1-42. However, the results for St8 of Aβ_1-42 were not satisfactory: the average imprecision amounted to 31.8 %CV, and the average inaccuracy amounted to −18.5%. Therefore, according to the US Food and Drug Administration guidance [15], this calibrator was removed from consideration for those 24 plates, and St7 was considered the lowest calibrator. For further runs, we replaced St7 and St8 with sequential 1:2 dilutions of St6, resulting in the concentrations of, respectively, 7.5 and 3.75 pg/mL for Aβ_1-40 and 2.5 and 1.25 pg/mL for Aβ_1-42. Performance of those calibrators (back-calculated values) is summarized in Table 1. Summary of the sigmoidal fits to those calibrators is also displayed on log concentration versus MFI plots in Fig. 3, along with box-and-whisker plots of measured patient results. The fitting function selected “1/y” weighting in 38 runs and “sqrt(1/y)” weighting in 40 runs for Aβ_1-40, and “1/y” weighting in 49 runs and “sqrt(1/y)” weighting in 29 runs for Aβ₄₂.

Table 1.

Summary of calibration information

	Aβ1-40				Aβ1-42
Standard	Nominal (pg/mL)	Mean ± SD (pg/mL)	Precision (%)	Percent inaccuracy	Nominal (pg/mL)	Mean ± SD (pg/mL)	Precision (%)	Percent inaccuracy
St1	1256	1046 ± 224	21.4	16.75	530	521 ± 26.3	5.06	1.78
St2	438	556 ± 93.0	16.7	−27.01	180	187 ± 9.66	5.15	−4.08
St3	198	176 ± 18.5	10.5	11.04	73	70.8 ± 3.34	4.72	3.05
St4	58	59.2 ± 3.61	6.10	−2.14	30	29.6 ± 0.98	3.30	1.44
St5	32	32.4 ± 1.94	5.98	−1.25	12	12.9 ± 0.47	3.67	−7.24
St6	15	14.5 ± 0.79	5.43	3.13	5	4.65 ± 0.26	5.53	7.10
St7	7.5	7.66 ± 0.41	5.33	−2.14	2.5	2.65 ± 0.21	7.99	−5.87
St8	3.75	3.72 ± 0.15	3.95	0.69	1.25	1.31 ± 0.20	15.0	−4.45

Abbreviations: Aβ, β-amyloid; SD, standard deviation.

Summary for 54 runs (of 78) performed with St7 and St8 as subsequent 1:2 dilutions of St6.

Fig. 3.

Summary log concentration versus MFI plots of the calibration curves for the 76 runs and box-and-whisker plots of the results. Solid thick line on the top of the individual run lines represents median.

A summary of the kit controls and the human plasma quality control samples is shown in Table 2 and in the form of stability plots in Fig. 4. In the first 42 runs, we used sample “a” as the human plasma QC2 sample; subsequently, we replaced the QC2 with sample “b” from another individual because sample “a” was similar in concentration to QC1 and did not offer enough coverage of the measured concentration range.

Table 2.

Quality control sample summary

	Aβ1-40			Aβ1-42
Sample	n	Mean ± SD (pg/mL)	CV (%)	n	Mean ± SD (pg/mL)	CV (%)
ConA	78	52.5 ± 2.72	5.19	78	179 ± 8.28	4.62
ConB	78	506 ± 57.3	11.3	78	33.4 ± 1.14	3.42
QC1	75	63.8 ± 4.49	7.04	75	14.7 ± 0.60	4.09
QC2a	42	57.1 ± 3.15	5.52	42	15.9 ± 0.82	5.15
QC2b	39	38.7 ± 2.59	6.70	39	7.73 ± 0.39	5.08

Abbreviation: CV, coefficient of variation.

“ConA” and “ConB” are kit controls prepared in artificial matrix by the kit manufacturer. “QC1,” “QC2a,” and “QC2b” are human plasma samples obtained from volunteer donors.

Fig. 4.

Stability plots for kit control samples (ConA and ConB) and for two human plasma controls (QC1 and QC2b). Thick solid line represents the average concentration, and dark and light shaded areas around the mean represent ±1 standard deviation and ±2 standard deviations, respectively.

A total of 195 samples were retested in the continuous test–retest performance evaluation. The average magnitudes of differences were 8.4 pg/mL for Aβ_1-40 (5.2% of the average concentration of 160.7 pg/mL) and 1.1 pg/mL for Aβ_1-42 (2.8% of the average concentration of 39.0 pg/mL). The slope of the Passing–Bablok regression line for test versus retest results was 0.971 (95% CI: 0.9057–1.0335) for Ab_1-40 and 0.936 (95% CI: 0.8939–0.9789) for Aβ_1-42. The parameters used were as follows: R² of 0.746, root-mean-square error of 7.73, average bias of −4.14 pg/mL, and 7.24% average coefficient of variation of residuals for the Aβ_1-40 fit; R² of 0.908, root-mean-square error of 1.18, average bias of −2.67 pg/mL, and 4.51% average coefficient of variation of residuals for Aβ_1-42 fit. The results of this experiment are graphically summarized in Fig. 5.

Fig. 5.

Passing–Bablok regression and Bland–Altman plots for the test–retest experiment. On the Passing–Bablok regression plots, the thick solid line represents the regression line, the dashed line is the line of identity, and the shaded area is the 95% confidence interval for the fit. In the Bland–Altman plots, the thick solid line represents the average difference between test and retest results, whereas dashed lines are drawn at ±1.96 standard deviation (95% confidence interval). Results are expressed as picograms per milliliter.

3.2. Patient results

For the 2454 samples from the ADNI study patients, 33 results did not meet the acceptability criteria for Aβ_1-40 (2 owing to bead count criteria, 8 owing to %CV criteria, and 23 owing to MFI criteria), and 24 results did not meet the criteria for Aβ_1-42 (5 owing to %CV criteria and 19 owing to MFI criteria). The range of the final results was from 12.6 to 371.7 pg/mL (95% CI: 54.8–256.2 pg/mL, median: 163.6 pg/mL) for Aβ_1-40 and from 3.47 to 102.8 pg/mL (95% CI: 17.9–62.1 pg/mL, median: 39.3 pg/mL) for Aβ_1-42.

When evaluating these data, we noticed small, but significant, effects of certain sample collection parameters on the results: collected blood volume (Aβ_1-40: Spearman ρ = 0.098, coefficient = 2.33, P <.0001; Aβ_1-42: ρ = 0.063, coefficient=0.311, P=.0033) and time to freeze (Aβ_1-40: ρ=−0.068, coefficient = −0.13, P = .0015; Aβ_1-42: ρ = −0.056, coefficient = −0.034, P = .0006). Therefore, all further analyses were adjusted for these variables.

The Aβ_1-40 and Aβ_1-42 results were found to be moderately correlated (ρ =0.599, P <.0001), and the equation estimated using Deming method to predict Aβ_1-42 levels from those of Aβ_1-40 was as follows: Aβ_1-42 = 0.152×Aβ_1-40 + 4.9 pg/mL.

No statistically significant differences were found in plasma levels of Aβ_1-40 and Aβ_1-42 and the plasma Aβ_1-42/Aβ_1-40 ratio between the three ADNI diagnosis groups at any of the time points. The average concentrations of Aβ peptides in the diagnosis groups stratified by visit are displayed in Table 3, along with P values for statistical significance. However, both the Aβ_1-40 and the Aβ_1-42 results (but not the Aβ_1-42/Aβ_1-40 ratio) were found to be significantly correlated with time, as shown in Table 3. Spearman correlation coefficients were found to be 0.182 (P <.0001) for time and Aβ_1-40 levels, 0.226 (P <.0001) for time and Aβ₄₂ levels, and −0.039 (P = .053) for time and the Aβ_1-42/Aβ_1-40 ratio.

Table 3.

Average biomarker values for each diagnosis group stratified by visit

			Aβ1-40		Aβ1-42		Aβ1-42/Aβ1-40
Visit	Group	n	Mean ± SD (pg/mL)	P value	Mean ± SD (pg/mL)	P value	Mean ± SD (−)	P value
BL	NL	206	151 ± 48.6	.998	37.7 ± 12.1	.290	0.266 ± 0.080	.165
	MCI	350	151 ± 53.9		36.3 ± 11.9		0.261 ± 0.102
	AD	162	152 ± 46.7		36.4 ± 10.4		0.249 ± 0.059
M12	NL	201	164 ± 45.2	.830	39.5 ± 11.2	.863	0.249 ± 0.058	.373
	MCI	348	166 ± 46.0		39.7 ± 11.1		0.252 ± 0.090
	AD	155	162 ± 42.4		39.3 ± 9.89		0.250 ± 0.051
M24	NL	183	164 ± 43.8	.117	40.8 ± 10.4	.774	0.258 ± 0.067	.193
	MCI	286	173 ± 43.6		41.5 ± 10.5		0.252 ± 0.092
	AD	120	167 ± 35.5		40.1 ± 8.98		0.246 ± 0.046
M36	NL	167	173 ± 39.3	.644	41.8 ± 9.72	.602	0.252 ± 0.062	.204
	MCI	213	177 ± 44.8		41.8 ± 11.1		0.247 ± 0.080
	AD	9	190 ± 52.6		39.8 ± 8.76		0.219 ± 0.051
M48	NL	3	207 ± 43.1	.448	42.0 ± 10.6	.507	0.204 ± 0.035	.555
	MCI	2	185 ± 30.1		44.1 ± 2.75		0.241 ± 0.024
	AD	2	177 ± 23.3		41.4 ± 0.25		0.235 ± 0.032

Abbreviations: NL, cognitively normal; MCI, mild cognitive impairment; AD, Alzheimer's disease.

3.3. Correlations with CSF data

We have extended and modified the evaluation of plasma Aβ correlations with CSF biomarkers performed by Toledo et al by incorporating additional CSF biomarkers (p-tau₁₈₁ and p-tau₁₈₁/Aβ_1-42 ratio), by using longitudinal CSF data in addition to the baseline data, and by using additional adjustment factors.

When using the “baseline” visit data with 359 results (each for different patient), the results were adjusted for patient's gender, age, and the APOE ε allele count, which were previously found to be significantly associated with plasma Aβ levels [16], and additionally adjusted for collected blood volume and time to freeze, which retained statistical significance in this setting. We have found multiple significant correlations: plasma Aβ_1–42 was correlated to CSF p-tau₁₈₁ (ρ = −0.180, P = .0006), CSF p-tau₁₈₁/Aβ_1–42 ratio (ρ = −0.171, P = .0011), CSF tau/Aβ_1–42 ratio (ρ = −0.137, P = .009), CSF tau (ρ = −0.116, P = .028), and CSF Aβ_1–42 (ρ = 0.110, P = .036). Plasma Aβ_1–40 was significantly correlated to the CSF p-tau₁₈₁ (ρ = −0.149, P = .0046) and CSF p-tau₁₈₁/Aβ_1–42 ratio (ρ = −0.130, P = .013). The Aβ_1–42/Aβ_1–40 ratio was significantly correlated to CSF tau/Aβ_1–42 ratio (ρ = −0.127, P = .016), CSF tau (ρ = −0.121, P = .021), and CSF p-tau₁₈₁/Aβ_1–42 ratio (ρ = −0.112, P = .034).

Using the mixed-effects model of the longitudinal data, we have evaluated these correlations in a multivariate way. There were a total of 140 patients having multiple results for both plasma and CSF studies at different points in time (404 total data points). Plasma Aβ_1–42 was found to be significantly correlated to the CSF p-tau₁₈₁/CSF Aβ_1–42 ratio (P = .0011) and the CSF p-tau₁₈₁ (P = .0234), and borderline significantly correlated to the CSF Aβ_1–42 (P = .0432). Interestingly, Aβ_1–40 levels and Aβ_1–42/Aβ_1–40 ratiowere not found to be significantly correlated to any of the CSF biomarkers using this approach.

Applying the autopsy-derived threshold values, we additionally used the CSF biomarker data from baseline visit to divide the plasma results into two groups [1]. We have used the following thresholds for CSF biomarkers: 192 pg/mL for Aβ_1–42, 93 pg/mL for tau, 23 pg/mL p-tau₁₈₁, 0.39 for tau/Aβ_1–42 ratio, and 0.1 for p-tau₁₈₁/Aβ_1–42 ratio [1]. For the 359 available data points, there were no significant differences in plasma Aβ_1–40 levels for any of the CSF biomarkers. For plasma Aβ_1–42 levels, the most significant differences were found between groups stratified by CSF Aβ_1–42 levels (38.6 ± 12.06 pg/mL in the “above” group and 35.1 ± 10.46 pg/mL in the “below” group, P = .0053) and CSF p-tau₁₈₁/Aβ_1–42 ratio (35.2 ± 10.45 pg/mL in the “above” group and 39.1 ± 12.34 pg/mL in the “below” group, P = .0029). Also significant differences were found between groups divided by CSF tau levels (34.4 ± 10.47 pg/mL in the “above” group and 37.5 ± 11.40 pg/mL in the “below” group, P = .0168), CSF tau/Aβ_1–42 ratio (35.3 ± 10.36 pg/mL in the “above” group and 37.9 ± 12.15 pg/mL in the “below” group, P = .0202), and CSF p-tau₁₈₁ levels (35.3 ± 10.49 pg/mL in the “above” group and 38.0 ± 12.04 pg/mL in the “below” group, P = .0210). Significant differences were also found for the plasma Aβ_1–42/Aβ_1–40 ratio when the groups were stratified by CSF tau/Aβ_1–42 ratio (0.251 ± 0.082 in the “above” group vs 0.271 ± 0.082 in the “below” group, P = .003), CSF Aβ_1–42 levels (0.271 ± 0.083 in the “above” group versus 0.253 ± 0.082 in the “below” group, P = .008), and CSF p-tau₁₈₁ levels (0.253 ± 0.083 in the “above” group versus 0.268 ± 0.080 in the “below” group, P = .012).

Drilling down these correlations, we have evaluated the by-CSF-threshold division separately in each diagnosis group. Again, there were no statistically significant differences in plasma Aβ_1–40 levels in any of the groups when stratified by any of the CSF biomarker thresholds. Interestingly, however, significant differences for Aβ_1–42 levels according to CSF biomarker thresholds were found exclusively in the MCI group, and not in the AD or NL groups. The plasma Aβ_1–42 levels stratified by CSF biomarker thresholds in the MCI group (n = 164) are displayed in Table 4, along with corresponding P values from comparisons of the “above-” and “below-threshold” groups. Additionally, Fig. 6 displays the plasma Aβ_1–42 levels in the MCI group, stratified by the thresholds of CSF Aβ_1–42 levels and the CSF p-tau₁₈₁/Aβ_1–42 ratio.

Table 4.

Plasma Aβ_1-42 values stratified by CSF biomarker thresholds in the MCI diagnosis group

	Above			Below
CSF biomarker (cutoff)	n	Mean ± SD (pg/mL)	95% confidence interval (pg/mL)	n	Mean ± SD (pg/mL)	95% confidence interval (pg/mL)	P value
Aβ1-42 (192)	45	40.1 ± 11.9	20.8–59.4	124	34.5 ± 10.3	11.3–55.5	.0064
Tau (93)	71	32.9 ± 10.8	10.1–56.6	98	38.3 ± 10.7	20.3–58.9	.0026
p-tau181 (23)	118	34.2 ± 10.5	10.8–55.8	51	40.3 ± 11.1	21.3–59.1	.0008
Tau/Aβ1-42 (0.39)	115	34.4 ± 10.4	10.8–56.1	54	39.4 ± 11.6	20.2–59.0	.0069
p-tau181/Aβ1-42 (0.1)	130	34.3 ± 10.2	12.1–55.4	39	41.7 ± 11.7	21.4–59.7	.0004

Abbreviation: CSF, cerebrospinal fluid.

Fig. 6.

Box-and-whisker plots with superimposed aligned dot plots of plasma Aβ_1-42 levels stratified by cerebrospinal fluid thresholds.

For the plasma Aβ_1–42/Aβ_1–40 ratio, the by-CSF-threshold evaluation separately for diagnostic groups revealed several significant differences between the AD and MCI diagnosis groups. In the AD group, the Aβ_1–42/Aβ_1–40 ratio was different when stratified by the CSF tau/Aβ_1–42 ratio (0.249 ± 0.068 in the “above-threshold” group versus 0.291 ± 0.046 in the “below-threshold” group, P = .013) and CSF Aβ_1–42 levels (0.286 ± 0.039 in the “above-threshold” group vs 0.251 ± 0.068 in the “below-threshold” group, P = .024). In the MCI group, significant difference in the Aβ_1–42/Aβ_1–40 ratio was found when stratified by CSF p-tau₁₈₁ levels (0.247 ± 0.088 in the “above-threshold” group versus 0.262 ± 0.078 in the “below-threshold” group, P = .030).

4. Discussion

In this study, we developed an improved protocol for analysis of Aβ_1–40 and Aβ_1–42 levels in human plasma and applied it to evaluate a cohort of 722 longitudinally followed patients from the ADNI study. We have automated the major steps of the procedure, as schematically depicted in Fig. 2, which enabled us to achieve substantial improvements in the precision and repeatability of the method.

As shown in Table 1, the precision of the calibrators for Aβ_1–42 levels was highest for the nominal concentrations in the range from 5 to 530 pg/mL and lower for the 1.25- to 2.5-pg/mL calibrators, whereas for Aβ1–40, it was best at the low limit of concentrations and more than 20% at the high end. However, the majority of measured results (95% CI: 18.3–85.4 pg/mL for Aβ_1–40 and 5.97–20.7 pg/mL for Aβ_1–42) were between the concentrations of St6 and St3, in the area of rising slope on the log concentration plot (Fig. 3). Therefore, in both cases, the measured results were within a well-defined part of the calibration curve, although the calibration ranges were exceedingly broad at the upper end.

The described procedure enabled us to achieve excellent kit control sample precision for Aβ_1–42 levels (3.4%–4.6%); the precision for Aβ_1–40 levels was also well within the acceptable range (5.2%–11.3%) [15], especially considering the fact that the “high” kit control sample ConB for Aβ_1–40 had an average concentration between calibrators “St1” and “St2”—the least precise calibrators for this biomarker. Other authors using the same assay have reported the average between-run precision of 9.3% for Aβ_1–42 and 9.9% for Aβ_1–40 results [2]. Between-run precision for another type of assay (enzyme-linked immunosorbent assay) was reported to be 17.3% on average [17]. Thus, the procedure described here provides a considerable improvement in measurement precision.

However, kit control samples are spiked samples prepared using an artificial aqueous matrix, and their performance does not always reflect the actual performance of the assay on patient samples. A better performance measure for the method was, therefore, reflected by the QC1 and QC2 samples that were plasma samples obtained from volunteers. Both of these samples had excellent precision ranging from 4.1% to 7%—on par with many mass spectrometry methods for much simpler analytes. With the exception of one run, as can be seen in Fig. 4, the results were stable across many runs, spanning approximately 7 weeks of time.

The results of the continuing test-retest evaluation also indicated good performance of the method, with an overall precision of 7.2 %CV for Aβ_1–40 and 4.5 %CV for Aβ_1–42—well within the acceptable range of 15 %CV [15]. There was considerably more variability in Aβ_1–40 results, as shown in Fig. 5, compared with Aβ_1–42, and Aβ_1–40 results additionally showed more bias (on average, −4.1% compared with −22.7% for Aβ_1–42). However, the bias, although statistically significant in cases of both peptides, was rather small in scale and lower than the general between-aliquot variability; therefore, we did not consider it a problem.

Many authors performing meta-analysis of existing plasma Aβ results have attributed the conflicting reports about the clinical value of plasma Aβ to general lack of consistency across analytical methods. Therefore, the method presented here is a substantial step toward eliminating this confounding factor. This improved protocol has provided results for 2454 samples from patients participating in the ADNI study. The in-depth analysis of this data set in the clinical context can be found in the article by Toledo et al [16]. Here we present only a brief summary of the data.

As expected, no significant differences in the Aβ_1–40 levels, the Aβ_1–42 levels, and the Aβ_1–42/Aβ_1–40 ratio were found between the clinical diagnostic groups, at any of the time points, as shown in Table 3. We have also studied the rates of the change of Aβ peptides and their ratio across diagnostic groups and found no significant differences (data not shown). Toledo et al have additionally evaluated the progression of the disease and found that plasma Aβ measures have limited value as prognostic factors [16]. However, to rigorously assess the potential utility of plasma Aβ measurements for prediction of progression, it is important to emphasize the need to apply this improved analytical methodology to earlier stages of MCI, which this study did not include.

However, there is a general steady increase of the Aβ_1–40 and Ab_1–42 levels with time [16], a fact reported previously by others although with less detail and statistical power [17,18]. The mixed-effects modeling of our data revealed that the age has both a “between-patient” and the “within-patient” effect. At the individual patient level, although peptide concentrations decreased in some patients, with some having too much variability to determine the trend, the concentrations increased in a majority of patients with time. In general, older patients in our group had significantly higher Aβ levels than younger patients, and the effect was strong enough to be detected in our generally well age-matched population. Substantial effect can also be attributed to patient's gender and APOE ε4 status, as described elsewhere [16]. Other allele types (2 and 3) did not seem to influence the plasma Aβ levels. In contrast, patients with two APOE ε4 alleles seem to have much increased variability of Aβ1–42/Aβ1–40 ratio, which is more stable in other patients.

We were able to determine significant correlations between the plasma Aβ_1-42 levels and the CSF biomarkers: Aβ_1-42, tau, and p-tau₁₈₁ levels as well as their ratios. Interestingly, these correlations were found exclusively in the MCI group of patients, and not in the AD or NL groups. There currently are very few reports in the literature on this topic, and the reported correlations between plasma and CSF biomarkers are rather weak [5,19]. However, such observations may highly depend on the choice of population, as also indicated by the conflicting reports relating plasma Aβ results to the diagnosis group.

Another interesting fact is that the correlation of baseline plasma Aβ_1-42 with CSF p-tau₁₈₁ results is much stronger than that with CSF Aβ_1-42 results. Using common sense, the correlation of the same species in two compartments of the body should be stronger than that of different species. This indicates that the dynamic balance between plasma and CSF biomarkers is very complex, and much more data are needed to characterize this, preferably covering a substantially broader period and spanning across the AD continuum.

Using the CSF biomarker thresholds [1], we found that the Aβ_1-42 levels in the MCI group were significantly different in the “above” and “below” subgroups when stratified by those thresholds. Nevertheless, as shown in Fig. 6, even for the strongest stratification by CSF p-tau₁₈₁/Aβ_1-42 ratio, the difference in plasma Aβ_1-42 levels was small between the “below”- and “above”-threshold groups. Using receiver operating characteristic analysis of the plasma Aβ_1-42 levels, we were able to predict to which CSF p-tau₁₈₁/Aβ_1-42 group a patient belongs only with 63% specificity and 55% selectivity at best. Therefore, plasma Aβ measurements are not yet suitable as a diagnostic test. Moreover, as in the case of univariate correlations, the association between CSF and plasma Aβ_1-42 levels was weaker than that of other CSF biomarkers when evaluated by CSF threshold stratification.

The plasma Aβ_1-42/Aβ_1-40 ratio, in contrast to many reports in the literature, did not turn out to be a good candidate for a biomarker in our study [20,21]. According to our data, the association of this ratio with various patient characteristics and CSF data was relatively weak.

Perhaps a diagnostic modifier could help extract more value from plasma Aβ measurements. In a recent animal model study, Takeda et al [4] reported that plasma Aβ levels respond differently to an oral glucose loading dose in mice with and without AD. Other authors have suggested other modifiers that could potentially be used to dynamically alter plasma Aβ levels. If these methods prove successful, plasma Aβ_1-42 level could potentially emerge as a valuable screening biomarker to be used before considering CSF testing, as it is a simpler and less invasive procedure.

Nevertheless, despite the current moderate diagnostic utility of this protocol, it also has practical utility in the determination of drug effect of anti-Aβ therapies. One report showed the effect of an investigational drug on plasma Aβ levels in a range from −75% to +150% throughout 24 hours [22]. Our protocol is perfectly suited for this type of study because of its high level of automation and high throughput, and it will allow researchers to determine Aβ concentrations with the greatest possible precision.

5. Conclusions

Using our robotic pipetting platform, we were able to substantially improve the analytical performance of the plasma Aβ_1-40 and Aβ_1-42 assay, eliminating a substantial amount of interassay variability. The excellent performance of our method was demonstrated using plasma quality control samples obtained from volunteers and run repeatedly across many plates throughout the study, as well as through retesting of the study patients' samples in every run. This improved protocol significantly reduces the effect of an important confounding factor that contributes to the lack of consistency between reports on plasma Aβ levels.

Using the described method, we processed 2454 samples from the ADNI study patients. We have determined significant correlations of the obtained results with sample collection parameters: time of storage at room temperature (time to freeze) and the volume of collected blood. We have confirmed significant correlations of plasma Aβ levels with age and gender and also with the APOE ε4 allele. Additionally, we have determined significant correlations between plasma Aβ levels and CSF biomarker levels and thresholds. However, these correlations are not strong enough to support research use as a diagnostic screening test at this point. Nevertheless, this biomarker test is well suited for use as a pharmacodynamic marker for therapeutic agents that alter the biosynthesis of Aβ or its clearance.