Neurofilament Light Chain as a Biomarker of Hereditary Transthyretin-Mediated Amyloidosis

The present case-controlled study includes analysis of plasma samples in a subset of patients enrolled in the APOLLO study who completed the study and consented to a biomarker discovery study. APOLLO eligibility criteria included Neuropathy Impairment Score of 5 to 130 (range 0–244, with higher scores indicating more impairment) and a polyneuropathy disability (PND) score of IIIb or lower.⁶ During the APOLLO study, patients received either patisiran 0.3 mg/kg or placebo intravenously every 3 weeks for 18 months.⁶ Further details of the APOLLO methodology have previously been described.⁶ Healthy control samples (n = 57) were collected separately and were age, sex, and race matched to the baseline demographics of the APOLLO patients included in this analysis (Dx Biosamples, LLC, San Diego, CA).

Plasma samples were taken from patients enrolled in the phase 3 APOLLO study (NCT01960348; ClinicalTrials.gov).⁶ APOLLO was conducted according to the guidelines of the International Conference on Harmonization, the World Health Organization Declaration of Helsinki, and the Health Insurance Portability and Accountability Act of 1996. Written informed consent was obtained from all patients. The APOLLO study protocol and all amendments were approved by the local institutional review boards and ethics committees. Healthy control samples were collected under institutional review board–approved protocols.

mNIS+7 is a composite measure of neuropathy that assesses motor, sensory, and autonomic neuropathy (range 0–304, with higher scores indicating more impairment) using a number of standardized assessments to quantify muscle weakness, muscle stretch reflexes, sensory loss, nerve conduction, and autonomic impairment.¹⁵ It represents the first Neuropathy Impairment Score specifically designed to assess patients with ATTRv amyloidosis in clinical trials. PND score is also used to assess severity of neuropathy in patients with ATTRv amyloidosis, with higher scores indicating greater impairments to ambulation. In the APOLLO study, all patients were assessed for mNIS+7 and PND scores at baseline and 9 and 18 months.⁶

Plasma samples of patients enrolled in APOLLO were collected according to the study protocol before dosing (n = 136 patisiran-treated, n = 53 placebo-treated patients).⁶ Levels of 1,161 unique proteins (1,196 total proteins) were measured with 13 human panels containing 92-plex immunoassays (Olink Proteomics, Watertown, MA) with intra-assay coefficients of variation ranging from 2% to 14% and interassay coefficients of variation ranging from 5% to 26%. Approximately 10% of the data points fell below the limit of detection of the assays with <1% of proteins not being detected in >50% of the 624 samples. Olink uses a proprietary proximity extension assay technology to combine a detection step involving oligonucleotide-labeled antibodies with a proximity-dependent DNA polymerization step and a real-time quantitative PCR amplification step to measure relative levels of multiple biomarkers simultaneously. Relative protein levels are reported as normalized protein expression (NPX) values, which are on a log₂ scale.

Subsequent quantitative measurements of neurofilament light chain (NfL) were made using an ultrasensitive single molecule array method with intra-assay coefficients of variation <12% and interassay coefficients of variation <10%, in duplicate, in a subset of patients from whom sufficient volume of plasma was available (n = 112 patisiran treated, n = 47 placebo treated) at baseline and 18 months. All NfL values were within the linear ranges of the assay except for 2 measurements, which were excluded.

The overall study size was determined by the number of patients who completed the APOLLO study and had plasma samples available at the 3 assessment time points. The number of healthy controls was more than double what was needed to detect a 20% difference in protein levels between healthy controls and baseline APOLLO participants, with 80% power and 5% significance to account for variable potential variance of proteins measured and to closely match the patient demographics.

To determine which protein levels changed the most over time as a result of placebo vs patisiran treatment, a linear mixed model regression analysis was performed for each protein to determine whether there was a significant differential time profile based on treatment. The Wilkinson notation representation of the model is given by the following: protein NPX ∼ treatment + time + treatment:time + age + sex + (1|patient), in which the response protein NPX represents protein levels, treatment denotes whether the patient was administered placebo or patisiran, time is a numeric quantity representing 0, 9, or 18 months, treatment:time is the interaction term, age and sex are additional covariates, and individual patients (patient) are assigned random intercepts. The model was regressed for each protein using the fit linear mixed-effects model (fitlme) function (MATLAB 9.3; MathWorks, Natick, MA), and the coefficients and p values associated with the treatment:time interaction coefficient were stored and plotted.

Principal component analysis (PCA) is a statistical procedure used to convert a set of observations that may be correlated into a set of linearly uncorrelated values called principal components. This transformation results in the first principal component (PC1) accounting for the largest possible variance in the data, with each subsequent principal component accounting for less variance. Here, the analysis is used to display the levels of 66 proteins in various cohorts of individuals (healthy control, baseline APOLLO, APOLLO/placebo, and APOLLO/patisiran) on a simple 2-dimensional plot by plotting the 2 principal components that explain most of the variability in the data. Missing measurements were replaced with the limit of detection for PCA analysis.

All boxplots are based on standard metrics, namely the minimum value, maximum value, lower and upper quartiles, and median value. Outliers are denoted as those values that are >1.5 times the interquartile range above the third quartile or below the first quartile. Plots were made with the Python Matplotlib graphics environment.

Because NfL has been shown to vary with age, we also report adjustment for this covariate when applicable. All correlation analysis was performed with Pearson correlation and partial correlations for covariate adjustment. All comparisons of means between groups were performed with an unpaired 2-sample t test in R (R Foundation for Statistical Computing, Vienna, Austria). To adjust for age, intergroup comparisons were tested using linear regressions with age as a covariate. Comparisons between the same individuals over time were not age adjusted because the longest time frame looked at was 18 months.

PCA analysis was performed with the Python sklearn package, and the 99.9% confidence intervals (CIs) were drawn using the variance-covariance matrix of each treatment group's multivariate distribution using the Ellipse patch as part of Matplotlib. Receiver operating characteristic curves were drawn by plotting the true-positive rate against the false-positive rate for varying cutoff values. The area under the curve was calculated. The Youden index was used to calculate the NfL level that best discriminates the healthy control cohort from APOLLO patients at baseline.

The datasets generated and analyzed during the current study are not publicly available.

Plasma levels of 1,161 unique proteins were analyzed with a proximity extension assay. A linear mixed model was used to determine the impact of patisiran treatment on the time profile of the plasma level of each protein by analyzing levels at baseline and 9 and 18 months. A total of 66 proteins were found to show a significant change in levels in patients treated with placebo vs patisiran over time (p < 4.18 × 10⁻⁵, with Bonferroni correction), of which NfL was the most significant (p = 3.95 × 10⁻²¹, figure 2). This analysis revealed that on treatment with patisiran, levels of some proteins increased (e.g., neutral ceramidase), whereas levels of other proteins decreased (e.g., NfL, R-spondin 3, coiled-coil domain containing 80). Many of the proteins affected by patisiran treatment have not been described as biomarkers (table e-1, https://doi.org/10.5061/dryad.1vhhmgqqk); however, 1 previously described biomarker of cardiac health, N-terminal prohormone of brain-type natriuretic peptide (NT-proBNP), showed a significant decrease (p = 7.02 × 10⁻¹⁴) on patisiran treatment. The NT-proBNP levels predicted by the proximity extension assay data and linear mixed model were consistent with those measured with an orthogonal method in the APOLLO study.⁶

PCA was used to elucidate whether there was a systemic plasma proteomic shift in patients with ATTRv amyloidosis with polyneuropathy and whether there was a subsequent reversal after treatment with patisiran. The 66 most significant proteins found in the mixed model analysis (table e-1, https://doi.org/10.5061/dryad.1vhhmgqqk) were compared in samples from healthy controls and baseline samples from patients with ATTRv amyloidosis with polyneuropathy. Projecting each individual's data onto the first 2 principal components that account for the most variance revealed a distinct separation between healthy controls and patients with ATTRv amyloidosis with polyneuropathy at the APOLLO baseline (figure 3A), with most of the separation being driven by the PC1. The proteomes of patients treated with placebo at 18 months were projected onto the same principal components resulting in a leftward movement along PC1 of the 99.9% confidence ellipse, suggesting that the proteome of patients treated with placebo was further separating from healthy controls (figure 3B). At 18 months, it was found that plasma proteomes of patients treated with patisiran were more similar to those of healthy controls compared with patients treated with placebo (figure 3C). Because there is significant heterogeneity in the proteome of patients at baseline, individual patient trajectories were also captured in 66-dimensional space. For each patient, the course of their disease progression or reversal was defined with 2 metrics, the first representing the rate (φ) and a second representing whether the proteome moved away from or toward the healthy controls (θ; see figure 3D legend). Plotting these 2 parameters clearly separated patients treated with patisiran and patients treated with placebo (figure 3E), emphasizing the distinct differences in the proteomes of the 2 groups at 18 months.

NfL was the most significantly changed protein in the analysis when comparing patients treated with placebo and patisiran with ATTRv amyloidosis with polyneuropathy and was therefore investigated in more detail. Patients diagnosed with ATTRv amyloidosis with polyneuropathy had >4-fold higher levels of NfL in their plasma at baseline relative to healthy controls, and this result held up independently of age adjustment (figure 4A; log₂ scale; table e-2, https://doi.org/10.5061/dryad.1vhhmgqqk). As expected, plasma NfL levels at baseline did not differ between the patisiran and placebo groups. The patisiran-treated group showed a significant decline in plasma NfL levels at 9 months that was sustained at 18 months, whereas plasma NfL levels in the placebo group increased at 9 months relative to baseline, and this level was sustained at 18 months (figure 4B). At 18 months, patients treated with placebo had 2-fold higher NfL plasma levels than patients treated with patisiran. Treatment with patisiran significantly lowered NfL levels in patients with ATTRv amyloidosis toward levels observed in healthy controls.

To assess variability between individuals, plasma NfL levels were plotted over time for each patient receiving placebo (figure 4C) or patisiran (figure 4D) treatment. Consistent with the general trend in the previous analysis, patients receiving placebo showed increasing plasma NfL levels over time, whereas patients receiving patisiran demonstrated decreasing plasma NfL levels. Changes in plasma NfL levels were correlated with changes in mNIS+7 score at 18 months (R = 0.4 [95% CI 0.27–0.51]; figure 4E), indicating that decreasing plasma NfL levels are associated with an improvement in polyneuropathy. Notably, no correlation was observed between plasma NfL levels and mNIS+7 (R = 0.07, age adjusted R = 0.1) or PND score at baseline (table e-3, https://doi.org/10.5061/dryad.1vhhmgqqk), regardless of age adjustment, although changes in PND score over time also trended with changes in NfL levels over time (figure 5). In addition, patients belonging to the prespecified cardiac subpopulation of the APOLLO study (defined as baseline left ventricular wall thickness ≥13 mm in the absence of a history of aortic valve disease or hypertension)⁶ saw significant elevation of plasma NfL levels (figure 4F and table e-4, https://doi.org/10.5061/dryad.1vhhmgqqk) relative to healthy controls, as did patients who did not belong to this prespecified cardiac subpopulation. This suggests that NfL can serve as a biomarker of polyneuropathy in patients regardless of whether they meet specified criteria for cardiac involvement.

The initial NfL measurements were validated with a widely used ultrasensitive single molecule quantitative assay to provide absolute plasma concentrations of NfL in healthy controls and ATTRv amyloidosis patient samples from APOLLO at baseline and 18 months (figure 6 and table e-5, https://doi.org/10.5061/dryad.1vhhmgqqk). There was a strong correlation between the 2 assays used to measure plasma NfL levels (R = 0.96 [95% CI 0.96–0.97]; figure 6A). With the use of the quantitative assay, the mean NfL level in healthy controls was 16.3 pg/mL, consistent with published reports (figure 6B).¹⁶ Patients with ATTRv amyloidosis with polyneuropathy had plasma NfL levels that were ≈4-fold higher at baseline relative to healthy controls (69.4 pg/mL, effect size 53.1 [95% CI 45.9–60.5]), which, after 18 months, decreased after patisiran treatment (48.8 pg/mL, effect size −23.2 [95% CI –33.4 to −13.1]) and increased in the placebo group (99.5 pg/mL, effect size 36.3 [95% CI 16.5–56.1]). One outlier with NfL plasma levels of 747 pg/mL in the patisiran-treated group was excluded from the calculations (figure 6, B and E) because this individual experienced a study-unrelated cerebral infarct at month 17 that may have caused the observed elevations of NfL. Consistent with the initial plasma NfL measurements (figure 4), measurements using the quantitative assay demonstrated a strong correlation between change in mNIS+7 score and change in plasma NfL levels at 18 months for each patient (R = 0.43 [95% CI 0.29–0.55]; figure 6C). Receiver operating characteristic analysis determined that plasma NfL levels could discriminate between healthy controls and patients with ATTRv amyloidosis with polyneuropathy (area under the curve 0.956; figure 6D), and the distribution of each was plotted (figure 6E). With the current data, an NfL level of 37 pg/mL may distinguish between healthy controls and patients with diagnosed ATTRv amyloidosis with polyneuropathy, with a false-positive rate of 3.6% and true-positive rate of 84.9%.