Genotype-Phenotype Association for 14 GFAP Variants in Alexander Disease

Abstract

Background and Objectives

Alexander disease is a rare monogenic disorder caused by dominant variants in GFAP (glial fibrillary acidic protein). Over 180 variants have been associated with the disease, with a wide spectrum of severity and clinical manifestations. Previous attempts at genotype-phenotype correlations have been hampered by the small numbers of cases that have been published for many of the variants. We sought to determine whether genotype-phenotype correlations could be discerned from available information.

Methods

We compiled a list of variants in GFAP for which 7 or more unrelated cases had been either published or identified through an ongoing natural history study and other sources (with a closing date of July 27, 2024). For each of these cases, we tabulated age at onset, age at death (or last contact), and sex. We used a Kruskal-Wallis test to evaluate statistical differences in age at onset in relation to variant. Differences in survival across variants were studied using Kaplan-Meier curves.

Results

Fourteen variants met our criteria for detailed analysis (10 with 7 or more unrelated cases and 4 additional variants involving 2 of the most commonly affected amino acids, R79 and R239) derived from a total of 231 cases. The variants seem to fall into 3 distinct groups—some with consistent early onsets (N77S, R79C and R79L, and most of the R239s), some with consistent late onsets (R70W and N386S), and some with more variable onsets (R416W). Pairwise comparison results found that R239H was associated with significantly earlier onsets than R239C. We found similar groupings for survival. Finally, we evaluated sex as a potential modifying factor for either age at onset or survival but found no significant association.

Discussion

Genotype-phenotype correlations do exist in Alexander disease, at least for a limited number of GFAP variants for which sufficient numbers of individual cases can be identified to allow valid statistical analysis.

Introduction

Alexander disease (AxD, ALXDRD) is a rare monogenic disorder of the CNS caused by dominant variants in GFAP, a member of the intermediate filament gene family that is a major component of the astrocyte cytoskeleton.¹ Only 1 population-based study of prevalence exists, from Japan, which arrived at an estimate of 1 in 2.7 million,² although others have suggested the value is likely closer to 1 in 1 million. Over 180 variants have been reported so far, and the clinical manifestations are remarkably broad, with disease onset ranging from in utero through the ninth decade of life.

Attempts at genotype-phenotype correlations are complicated by the rarity of the disease and the fact that nearly 2/3 of the reported variants are private, thus preventing any determination of consistency between cases.³ The published literature contains occasional mention of genotype-phenotype correlations for the very few variants for which a large number of reported cases exist, although none of these comments were based on a systematic analysis of available data. The first quantitative assessment was conducted by Prust et al.⁴ These authors studied 215 cases and calculated ages at onset associated with a number of variants, including those affecting 4 of the frequently involved amino acids (R79, R88, R239, and R416), but did not distinguish between different variants affecting each of these residues and grouped all other known variants together as a single entity. Recently, Grossi et al.⁵ reported an analysis of all known variants but reached few conclusions regarding the effects of any one variant. We believe that enough information is now available to permit evaluation of 2 specific aspects of the clinical phenotype—age at onset and survival—for many more variants than currently known. We greatly expand the available data set and show that certain variants offer clear predictions about age at onset and survival. This analysis will inform approaches to diagnosis, genetic counseling, and future research.

Methods

Cohort, Data Collection, and Data Curation

The cohort of cases used in this analysis was derived as follows. A bibliography of all publications related to Alexander disease, derived from weekly searches in PubMed, Web of Science, and Google Alerts, is maintained with daily updating by one of the authors at the University of Wisconsin-Madison.⁶ From this bibliography, we compiled a list of 364 publications, beginning with the initial genetic report in 2001. We found 419 reports of single nucleotide changes resulting in 161 different missense variants (maintained as a separate online resource⁷). These variants affected 94 of the 432 amino acids in GFAPα, the major isoform in the CNS. Several reports described a variant involving an amino acid found only in GFAPδ (R430H), the second most abundant isoform.^8-11 In addition, one single nucleotide change affecting GFAPδ resulted in a synonymous variant (R430R) that was still considered pathogenic because of a predicted change in splicing.⁸ Beyond single nucleotide changes that predicted alterations of single amino acids, 12 examples were reported of small insertions, deletions, or duplications and 8 examples of more complex changes such as early termination, major deletions due to exon skipping, and frameshifts that allowed continued translation but of an incorrect sequence. Of the putatively pathogenic variants that have been reported in the literature, 114, or 62.6% of the total, are considered private (i.e., only 1 reported individual or family in the data set). We then added unpublished individuals known to the authors through an ongoing natural history study at CHOP (NCT02714764) and other contacts.

We limited our analysis to variants for which there are reports or records of at least 7 or more unrelated individuals, the goal being to have sufficient independent data points to assess relevant clinical features as defined below. We considered the potential value of using a lower threshold for inclusion but encountered problems with 2 variants at the level of 6 independent reports (one contained a patient who was homozygous for the variant rather than the typical heterozygosity and the second involved uncertainty about potential familial connections and whether the separate publications reflected truly independent cases). Instances of twins or familial inherited forms, and even the large families such as those reported by Stumpf et al.,¹² Messing et al.,¹³ and Helman et al.,⁸ are each considered as 1 occurrence. To avoid bias from identical twins, we counted only the first of each twin in any listing. In reports of parent-child pairs, we counted only the parent to avoid the potential bias of skewing to earlier recognition of age at onset in the child. In multigeneration examples, we chose the proband as described in the publication. Extracting information from these reports, we tabulated 3 factors—sex, age at onset, and age at last contact or age at death. We took note of duplicate publications of the same individuals because these should not be counted twice, yet the later publications could offer updated information on survival. We have updated the ages at last contact or ages at death (information collected from April to July, 2024) to the extent possible. The sources for all cases are provided in eTable 1 (without potentially identifiable information such as ages and sex per journal guidelines), using asterisks to denote published cases where information has been updated or corrected from the original. Three cases were omitted from our analysis because of ambiguity over age at onset or genetic diagnosis, although these are included in the eTable(s) and marked accordingly. Those variants for which 3–6 unrelated cases have been published or known are presented in eTable 2, although those are not included in our statistical analysis.

As of 2021, some individuals with AxD began participating in a clinical trial of a potential disease-modifying therapy using GFAP-targeted antisense oligonucleotides (NCT04849741). Hence, for these participants, we used the age at enrollment as the age at last contact without consideration of their subsequent course. Among the unpublished cases, most age calculations could be based on our knowledge of exact dates of birth, although for some, both dates of birth and death were only known by month and year.

Definitions

We define age at onset as the onset of clinical signs or symptoms considered relevant to a potential diagnosis of Alexander disease or MRI evidence of leukodystrophy. The latter criterion applies even to individuals otherwise considered asymptomatic (as occurred in 2 examples where imaging was performed for other reasons but revealed white matter lesions that prompted subsequent genetic testing—omitting these 2 individuals from the data set did not significantly alter the results or conclusions). Most publications provided a specific number for age at onset, but some were less specific. Because such terminology is not useful for statistical analysis, systematic criteria were developed by the authors. For instance, if age at onset was described as < 1, as, for example, cases # 5 and 6 in the first report of GFAP genetics in Alexander disease,¹⁴ or as infancy, these were rounded up to 1. The full list of such criteria are available upon request.

Plan for Statistical Analysis

Owing to anticipated violations of ANOVA assumptions, we use a Kruskal-Wallis test to evaluate statistical differences in age at onset in relation to variant. In the Kruskal-Wallis test, age at onset observations are combined across variants and replaced by ranks. A test for mean rank differences across variants is then performed using the Kruskal-Wallis H statistic, which is asymptotically χ² distributed under a null hypothesis of no variant differences. A significance test leads to post hoc pairwise comparisons between each pair of variants using Dunn tests.¹⁵ Dunn tests entail a z-test approximation based on the difference in mean ranks divided by a pooled variance estimate of the within-group ranks across the pair. Owing to the large number of pairwise comparisons, Bonferroni correction was applied to control familywise Type I error at 0.05 across all pairs.

Differences in survival across variants were studied using Kaplan-Meier curves. Cases for which a death had not occurred were treated as right censored at the age at last contact. To evaluate differences between survival curves in relation to variant, Mantel-Cox log-rank tests were applied.¹⁶ The log-rank test compares across groups the hazard of event (death) occurrence using a χ² statistic. A significant test result is followed by pairwise comparison log-rank tests for variant pairs; in this analysis, we focus on pairwise comparisons within the 3 subcategories of variants mentioned above, applying Bonferroni correction within the single disease-causing variant category.

Standard Protocol Approvals, Registrations, and Patient Consents

Data were collected under protocols approved by institutional review boards at the University of Wisconsin-Madison and Children's Hospital of Philadelphia. Consents were provided by the patient or legally authorized representative, where appropriate, and assent was obtained from minors when capable or adults with diminished capacity who were unable to provide consent.

Data Availability

Anonymized data not published within this article will be made available by request from any qualified investigator.

Results

Ten variants, affecting 8 different amino acids, met our criterion for inclusion of at least 7 unrelated individuals or families (Table 1). To the cohort with these variants that we derived from the published literature, which numbered 183 individuals, we added 48 individuals known through the natural history study that is underway at CHOP, thus yielding a total of 231 individuals (see CONSORT diagram, Figure 1). The total cohort included 117 male and 102 female individuals, as well as 12 for whom sex was not identified in the publications. We were also able to update the survival information for 21 individuals who had previously been described in the literature, either because they were participants in our own studies and we were in continued contact or because they participated in studies of our colleagues. We evaluated differences across all 10 variants, as well as for 3 subcategories of variants—the 2 R79 variants, the 2 R239 variants, and the remaining 6 variants representing amino acids with single disease-causing variants. Separately, we examined 4 variants (1 at R79 and 3 at R239) for which the number of independent cases did not meet our minimum for inclusion in statistical analysis but were nevertheless considered valuable for comparison.

Table 1.

Total Cohort

Nucleotide changea	Variant	Total	Male	Female	Sex n.s.	Updated from publications	CHOP
c.208 C>T	R70W	7	3	4
c.230 A>G	N77S	8	3	5		2	2
c.235 C>T	R79C	40	21	19		3	16
c.236 G>A	R79H	35	15	16	4	2	6
c.262 C>T	R88C	33	16	15	2	4	7
c.715 C>T	R239C	43	17	22	4	2	10
c.716 G>A	R239H	23	13	9	1	1	2
c.772 C>T	R258C	7	6	1
c.1157 A>G	N386S	11	6	4	1	1
c.1246 C>T	R416W	24	17	7		4	5
	Totals	231	117	102	12	21	48

Abbreviation: n.s. = not specified.

Those for whom survival information was updated from an original publication are indicated. Cases newly ascertained from the CHOP natural history study rather than a prior publication are tabulated separately.

^aReference sequence = NM_002055.5.

Figure 1. CONSORT Flow Diagram.

CONSORT flow diagram shows the selection of case reports used in the analysis of disease phenotypes.

Descriptive Findings

We first considered age at onset as a function of the GFAP variant. The results for the cases compiled through our combined literature search and natural history study are shown graphically in Figure 2A. Based on visual inspection alone, the variants seem to fall into 3 distinct groups—some with consistent early onsets (N77S, R79C, R88C, R239C, and R239H), some with consistent late onsets (R70W, R258C [with 1 exception], and N386S), and some with more variable onsets (R79H and R416W).

Figure 2. Ages at Onset.

Ages at onset for the 10 variants in GFAP for which at least 7 unrelated cases are known. Each symbol represents an individual case, with the means indicated with a horizontal red line. (A) All 10 variants illustrated together, (B) 2 variants at the R79 position, and (C) 2 variants at the R239 position. Male individuals (blue squares), female individuals (pink squares), sex undefined (gray triangles).

Statistical Analysis of Genotype-Phenotype Associations

We begin with the question of age at onset. eTable 3 provides the mean, SD, and 95% bootstrapped confidence intervals for the mean age at onset of each variant. The overall Kruskal-Wallis test resulted in H = 121.01, df = 9, p < 0.001, implying differences across variants in age at onset. A corresponding effect size estimate given by η² = (H − k + 1)/(n − k), where k is the number of variants (10) and n is overall sample size (231), yielded η² = 0.51, indicating a large effect.¹⁷ For those amino acids with more than 1 variant, specifically R79 and R239, the follow-up pairwise comparisons assessed whether age at onset was additionally affected by the specific change that occurred. In the case of the 2 R79 variants available for study, the difference between variants was not significant (z = 0.96, p = 0.337, adj. p = 1.000) (Figure 2B). In the case of the 2 R239 variants available for study (Figure 2C), the pairwise comparison result found that R239H was associated with significantly earlier onsets than R239C (z = −3.30, p < 0.001, adj. p = 0.043). Formal pairwise comparisons for these and all other pairs of the full set of variants are presented in eTable 4 (which gives the mean ranked differences and accompanying Dunn statistical tests).

We next examined the relationship between variant and survival, defined by the event of death, with all living individuals censored at the age at last known contact or age at entry into the antisense clinical trial. The Mantel-Cox log-rank test comparing survival across all 10 variants yields χ² = 152.26, df = 9, p < 0.001, implying significant differences in survival. Owing to the large number of variants, to facilitate comparison, we focus on variant subsets. First, we show Kaplan-Meier curves for the 5 amino acids with single disease-causing variants (Figure 3A). The Mantel-Cox log-rank χ² = 37.15, df = 5, p < 0.001 implies significant differences within this variant subset. Similar to the pattern described above for ages at onset, we could discern 3 types of curves for survival—3 with relatively long survivals (R70W, R258C, and N386S), 2 with intermediate survivals (R88C and R416W), and 1 with a shorter survival period (N77S). Post hoc pairwise comparisons based on log-rank tests largely confirm the distinctions between these variant groupings described above after Bonferroni correction, with all pairs having adjusted p < 0.05, except the distinctions between R70W and N386S [adj. p = 0.380], R70W and R258C [adj. p = 1.00], R258C and N77S [adj. p = 0.382], and R88C and R416W [adj. p = 1.000] clearly failed to achieve significance, the differences between R258C and N77S [adj. p = 0.058] and between R70W and N77S [adj. p = 0.055] being slightly above the significance threshold. We next considered the 2 variants at R79, for which R79C seemingly showed a more rapid decline than R79H (Figure 3B), although the difference for this pairwise comparison was just above the significance threshold (Mantel-Cox log-rank χ² = 3.50, df = 1, p = 0.062), likely because of the relatively young ages at which most of the R79C individuals are censored. Finally, we examined the 2 R239 variants. At this position, we observe a significant difference between R239C and R239H variants (Mantel-Cox log-rank χ² = 25.12, df = 1, p < 0.001), with H showing a reduced survival compared with C (adj. p < 0.001) (note that this is the opposite of that observed, but not statistically confirmed, for R79) (Figure 3C).

Figure 3. Kaplan-Meier Survival Curves.

(A) Kaplan-Meier curves for individuals with the R70W, N77S, R88C, R258C, N386S, and R416W variants in GFAP. Omnibus test implies significant differences within this variant subset (Mantel-Cox log-rank χ² = 37.15, df = 5, p < 0.001). (B) Kaplan-Meier curves for individuals with the R79C and R79H variants. The R79C curve appears more rapid than R79H, although the difference is just above the statistical significance threshold (Mantel-Cox log-rank χ² = 3.50, df = 1, p = 0.62). (C) Kaplan-Meier curves for individuals with the R239C and R239H variants. The R239H curve is significantly more rapid than for R239C (Mantel-Cox log-rank χ² = 25.12, df = 1, p < 0.001).

We asked whether sex was a potential modifying factor for either age at onset or survival. Both tests were nonsignificant, with Kruskal-Wallis H χ² = 0.35, df = 1, p = 0.555 for age at onset and Mantel-Cox log-rank χ² = 0.87, df = 1, p = 0.350. Furthermore, none of the variants yielded significant differences for either measure when individually analyzed for sex differences.

Other Variants of Interest

Two of the most commonly affected amino acids, R79 and R239, are associated with variants in addition to those listed in Table 1 but reported in fewer than the minimum 7 independent occurrences as set forth in Methods (R79L [3], R239G [3], R239L [6], and R239P [1]). While the numbers of such cases are insufficient for formal statistical analysis, the results of each are largely consistent with that of the other variants at these positions, with the possible exception of R239G, which seems unusually mild compared with all other R239 variants (publications describing individual case reports for individuals with these variants are included in eTable 2).

Discussion

We report that genotype-phenotype correlations do exist in Alexander disease, at least for a limited number of GFAP variants for which sufficient numbers of individual case reports can be identified to allow valid statistical analysis. We find that certain variants reliably cause onset at a very early age (such as N77S and R239H), whereas others consistently cause onset at the opposite end of the age spectrum (such as R70W and N386S). We find similar differences in survival, which is to be expected based on previous reports linking survival to age at onset.^4,18 None of these correlations seem to be related to sex, as noted previously.⁴ Our findings are consistent with an earlier study by Yasuda et al.,¹⁹ who examined ages at onset for 4 of the variants included in our analysis with a smaller number of cases.

It is interesting to note the marked variability in outcomes observed for the R79H and R416W variants. That such variability can exist has been known for some time (note the first 2 R79H patients ever reported, in reference 14), and variability has also been noted within intergenerational families carrying other variants.^8,12,13 A number of hypotheses have been suggested to account for this variability (discussed in reference ³ including environmental factors such as head trauma, alcohol exposure, or infection, or genetic modifiers within GFAP or other genes), but to date none have been proven. The 3 pairs of identical twins included in the current analysis were each remarkably consistent with each other, and we are aware of at least 4 additional sets of identical twins with other variants and similar concordance of phenotypes (one of these coming from the CHOP cohort).^18,20,21

Whether our findings can be interpreted in terms of mechanistic changes in the assembly and structural properties of the GFAP protein is not clear, because the literature on this subject is sparse. Few studies have even attempted direct comparisons between the particular variants listed in our Table 1 using the approaches typically used for analysis of intermediate filament proteins, such as in vitro assembly,²² transfection into cultured cells,^23-25 caspase cleavage,²⁶ or western blotting for GFAP and other proteins relevant to the pathogenic cascade.^27,28 In mouse knock-in models of the R79H and R239H variants (numbered 76 and 236 according to the mouse sequence), Hagemann et al.²⁹ found that this amino acid change at the R239 position resulted in a higher level of GFAP accumulation and more Rosenthal fibers than at R79, which Jany et al.³⁰ later attributed in part to transactivation of the GFAP promoter. Battaglia et al.²⁷ identified phosphorylation of Ser13 in GFAP as a potentially key event that distinguished early-onset from later-onset cases, but there were no obvious differences between N77S, R79C, R239C, and R239H in western blot analysis of the human brain. Yang et al.²² studied 14 variants of recombinant GFAP (i.e., produced in bacteria, and thus lacking post-translational modifications of normal GFAP) using in vitro assembly, including R79C and R79H, and again failed to find a significant difference in the behavior of these 2 variants. Most recently, using in vitro assembly and focusing specifically on R-to-C transitions, Lin et al.³¹ found that R239C was more prone to cross-linking and more susceptible to oxidative stress than R79C and R88C. In the same study, R239C was also more prone to spontaneous filament polymerization than R79C and R88C, all of which potentially point to explanations for why these variants might differ in the severity of disease as manifested by age at onset or survival.

Our study has several limitations. First, until ∼2005, a high proportion of patients were published, either as single case reports or as parts of series, but since that time many with the more common variants are no longer reported and hence become invisible for the purposes of our study. Second, there is no consensus definition for age at onset—sometimes the onset is precise, such as the date of first seizure, but other times it depends on failure to meet developmental milestones for which there is known heterogeneity in the healthy population. In some adult-onset cases, it relies on recollections of events that occurred decades earlier. In nearly all cases, we have relied on the publications' definitions for age at onset, with exceptions when clear statements were made that allow determination based on accepted criteria (such as milestones for some motor functions). Third, we have incomplete capture of survival data because many patients were not followed up after the initial publication. For many of the participants in our own studies, we were able to update this information because we have maintained contact, but for others such updating was impossible. Hence, our analyses related to survival should be recognized as referring to the most conservative end of the spectrum, and actual survival is certainly longer than portrayed in our figures. We also emphasize that our analysis covers patients born over a span of several decades, and it is likely that survival in part might have changed over time because of improvements in clinical care, especially for those with severe impairments.

We propose several approaches to rectify these limitations. Collaboration among pediatric and adult neurologists and other specialists who care for affected individuals will facilitate the publication of larger case series that will more rapidly add to the published literature for enhanced genotype-phenotype correlations. More accurate estimates of survival could be achieved through the development of comprehensive patient registries that permit periodic or even automatic updating of status. Development or implementation of quantitative measures of the disease phenotype, such as for seizures (type, frequency, and response to anticonvulsants), nausea, and motor and cognitive functions, are also needed to provide more refinement in the comparisons between different variants. Modifications of the clinical classification systems are underway, such as those recently proposed by Vaia and colleagues,³² and these may expand the ways in which genotype can be mapped onto phenotype. Nevertheless, the analysis of the 14 variants reported here should prove useful as a guide both for clinical practice and the design of future research.

Author Contributions

A. Messing: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data; study concept or design; analysis or interpretation of data. A.T. Waldman: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data; analysis or interpretation of data. D.M. Bolt: drafting/revision of the manuscript for content, including medical writing for content; study concept or design; analysis or interpretation of data.

Study Funding

This work was supported by grants from the NIH (P50HD105353, U54NS115052) and Elise's Corner.

Disclosure

A.T. Waldman has received research support from Ionis Pharmaceuticals (Investigator-initiated research), Ionis Pharmaceuticals (Clinical trial support - Alexander disease and Pelizaeus Merzbacher Disease), Roche/Genentech (Clinical trial support), Novartis (Clinical trial support), PassageBio (Clinical trial support), Sarepta (Clinical trial support), Pfizer (Clinical trial support); personal compensation for serving on a data safety monitoring board (SwanBio); and publishing royalties (UpToDate, MedLink Neurology). The other authors report no relevant disclosures. Go to Neurology.org/NG for full disclosures.