Glial Fibrillary Acidic Protein: A Biomarker and Drug Target for Alzheimer’s Disease

Glial fibrillary acidic protein (GFAP) is an intermediate filament structural protein involved in cytoskeleton assembly and integrity, expressed in high abundance in activated glial cells. GFAP is neuroprotective, as knockout mice are hypersensitive to traumatic brain injury. GFAP in cerebrospinal fluid is a biomarker of Alzheimer’s disease (AD), dementia with Lewy bodies, and frontotemporal dementia (FTD). Here, we present novel evidence that GFAP is markedly overexpressed and differentially phosphorylated in AD hippocampus, especially in AD with the apolipoprotein E [ε4, ε4] genotype, relative to age-matched controls (AMCs). Kinases that phosphorylate GFAP are upregulated in AD relative to AMC. A knockdown of these kinases in SH-SY5Y-APPSw human neuroblastoma cells reduced amyloid accrual and lowered protein aggregation and associated behavioral traits in C. elegans models of polyglutamine aggregation (as observed in Huntington’s disease) and of Alzheimer’s-like amyloid formation. In silico screening of the ChemBridge structural library identified a small molecule, MSR1, with stable and specific binding to GFAP. Both MSR1 exposure and GF AP-specific RNAi knockdown reduce aggregation with remarkably high concordance of aggregate proteins depleted. These data imply that GFAP and its phosphorylation play key roles in neuropathic aggregate accrual and provide valuable new biomarkers, as well as novel therapeutic targets to alleviate, delay, or prevent AD.

GFAP mutations, favoring aggregation to form Rosenthal fibers, are both causal and diagnostic for chronic gliosis, as found in Alexander's disease [3], and several singlenucleotide polymorphisms in GFAP are strongly associated with this demyelination disorder [3]. GFAP expression is transcriptionally modulated by multiple growth factors and

Chemotaxis and Paralysis Assays in Aβ-Transgenic C. elegans Strains CL2355 and CL4176
Transgenic C. elegans strains, expressing Aβ 1-42 in neurons (CL2355) or in muscle (CL4176), were grown at 20 • C with ample E. coli (OP50) bacteria, and on day 1 of adult life, worms were lysed to release eggs and thus generate a synchronized cohort. Eggs were then placed on 100 mm NGM-agar Petri dishes seeded with RNAi-expressing bacteria (E. coli HT115) targeting a GFAP ortholog or empty-vector control bacteria. Worms at the L3-L4 transition were upshifted to 25.5 • C to induce human Aβ 1-42 transgene expression and assayed after 48 h. Worms were returned to 20 • C, and chemotaxis [31] and paralysis [32] assays were performed as described [29,[33][34][35].

Paralysis Assays in Human Tau-Expressing C. elegans Strain VH255
Transgenic C. elegans strain VH255, with pan-neuronal expression of human tau [36], was maintained at 25 • C on agar plates overlaid with a lawn of E. coli (OP50) bacteria. Worms were lysed in alkaline hypochlorite solution to obtain unlaid eggs used to initiate a synchronized culture. The eggs were then transferred to 100 mm NGM-agar plates, to which drug candidates MSR1 or MSR2 were added to final concentrations of 1 µM. For ifp-1 RNAi knockdown, the bacterial lawn consisted of strain HT115 expressing an exonic segment of double-stranded ifp-1 RNA (see next section). The worms were washed and added to fresh, drug-equilibrated plates every two days. Assays were performed on day-3 adults (i.e., 5.5 days after egg hatching) to assess the fraction of paralyzed worms.

siRNA Knockdowns and Thioflavin-T Staining of Human Cells
SH-SY5Y-APP Sw cells from an exponentially growing culture were trypsinized, rinsed, plated at 10,000 cells per well in 96-well plates, and grown for 16 h at 37 • C in "DMEM + F12" (Life Technologies, Waltham, MA, USA) supplemented with 10% (v/v) fetal bovine serum (FBS). Cells at~40% confluence were transfected with short interfering RNA (siRNA) constructs targeting GFAP (SAS1_Hs01 00227618), AKT2 (SAS1_Hs01 00035058), ROCK1 (SAS1 Hs 00065571), BARK (SAS1 Hs 00039321), or PKA (SAS1 Hs 00217223), all obtained from Millipore-Sigma (St. Louis, USA). Transfections with the indicated siRNAs were performed using RNAiMax reagent (Life Technologies) according to the manufacturer's directions. Cells at 48 h post-transfection were fixed in 4% v/v formaldehyde and stained in a dark container with 0.1% w/v Thioflavin T. After 4 washes in PBS, cells were covered with Antifade + DAPI (EMD/Millipore-Sigma), and fluorescence was captured in green and blue channels, using a Keyence fluorescence microscope with motorized stage for automated well-by-well imaging, 9 fields per well. Thioflavin-T fluorescence intensity was divided by the number of DAPI + nuclei in each well, yielding ratios of amyloid per cell, summarized as mean ± SD.

MSR1 Treatment of SH-SY5Y-APP Sw Cells
Human neuroblastoma cells were grown as previously described [33,40]. These SH-SY5Y-APP Sw cells, expressing an aggregation-prone "Swedish" double mutant of amyloid precursor protein (APP Sw ), were maintained in DMEM plus 10% (v/v) FBS at 37 • C. Cells were suspended in trypsin/EDTA and rinsed in buffer prior to replating or harvesting. Immediately preceding assay, cells were grown for 48 h in the presence of 10 µM MSR1 dissolved in DMSO (0.02% final concentration) or 0.02% DMSO (solvent alone) for control cells. Cells were harvested, total protein was isolated, and insoluble aggregate proteins were purified as described below.

Modeling and MD Simulation of GFAP Structure
The three-dimensional structure of GFAP was modeled using fold recognition and ab initio structure prediction methods implemented by I-TASSER server-based algorithms. I-TASSER generates 5 different models, of which the lowest-energy conformer was chosen for further processing. For molecular dynamic (MD) simulations, we used the proteinpreparation wizard within the Schrödinger Desmond simulation suite to prepare modeled structures. Physiological conditions were approximated during simulations by creating an orthorhombic simulation box filled with simple point charge (SPC) water, neutralization of locally charged sites with appropriate counterions (Na + , Cl − ), and further addition of 0.15 M NaCl to achieve a physiological isotonic state. For equilibration, temperature and pressure were held at 300 • K and 1.1023 bar, respectively. The random sampling seed input was changed for each run, and each 200 to 500 ns simulation was repeated at least three times. Simulations with phosphorylations were incorporated using the Maestro "Mutate residue" plug-in, in which specified residues were converted to their phosphorylated form. Trajectories were visualized with VMD and analyzed using BIOVIA Discovery Studio (Dassault Systemes, Waltham, MA, USA).

Virtual Screening of a Target Protein against Molecular Structure Libraries
High-throughput virtual screening of the ChemBridge molecular structure library for docking to GFAP was initially conducted using the Glide module of Schrödinger Suite. ChemBridge structures were retrieved in 2D format and prepared using LigPrep wizard (Schrödinger Suite). A 3-stage strategy was employed to improve the efficiency of virtual drug screening: (i.) the entire library of~750,000 molecular structures was virtually docked to the GFAP protein in Glide's high-throughput mode; (ii.) the top 1% of structures from high-throughput screening were then docked again to GFAP in Glide's standard precision mode; and (iii.) for the top 1% of structures from standard precision docking, binding free energies were predicted under MM-GBSA conditions using the Schrödinger Suite Prime module. Protein-ligand complexes with highest avidity (lowest ∆G binding ) were simulated using the Schrödinger Desmond module to assess their stability over time.

Statistical Analyses
For replicated assays of protein aggregation, chemotaxis, and paralysis, differences between control and experimental groups were assessed for significance with the Fisher-Behrens heteroscedastic t-test (appropriate for samples of unequal or unknown variance), treating each experiment as a single point. Within experiments, differences in proportions (fractional paralysis or chemotaxis) were evaluated with chi-squared or Fisher exact tests, as appropriate, based on sample counts.
We also used PEAKS software (PTM module) to screen several other PTMs, but none were useful in distinguishing AD from AMC. In all groups, methylated arginine and lysine were observed at R88 and K95, whereas K107 was dimethylated, all >90%. Deamidation was not observed in >10% of spectral counts for any GFAP peptide, and pyroglutamine never exceeded 25% (data not shown).

Molecular Dynamic Simulations Predict GFAP Unfolding and Identify a Druggable Pocket
Molecular dynamic simulations of the hP-GFAP structures observed in AD aggregates (rendered by phosphomimetic substitutions) predict a more malleable GFAP structure in ApoE(3,3) individuals but relatively greater structural rigidity in AD(4,4) compared to unphosphorylated GFAP, as seen in AMC aggregates ( Figure 2). Since GFAP is a largely disordered protein, its full-length structure has not been experimentally determined. We therefore predicted its three-dimensional structure using fold recognition and ab initio procedures in I-TASSER [44]. The resulting hypothetical structure comprises helices and loops (Figure 2a), forming a small pocket or cavity near the inner groove of the protein (yellow region).
In view of its disordered nature, the predicted protein structure is expected to be unstable and to unfold spontaneously, altering the orientation and drug accessibility of the pocket. To explore the protein unfolding trajectory, we simulated the predicted structure of fully solvated GFAP for 0.5 µs (500 ns). The volume of the druggable pocket (Figure 2b

Molecular Dynamic Simulations Predict GFAP Unfolding and Identify a Druggable Pocket
Molecular dynamic simulations of the hP-GFAP structures observed in AD aggregates (rendered by phosphomimetic substitutions) predict a more malleable GFAP structure in ApoE(3,3) individuals but relatively greater structural rigidity in AD(4,4) compared to unphosphorylated GFAP, as seen in AMC aggregates ( Figure 2). Since GFAP is a largely disordered protein, its full-length structure has not been experimentally determined. We therefore predicted its three-dimensional structure using fold recognition and ab initio procedures in I-TASSER [44]. The resulting hypothetical structure comprises helices and loops (Figure 2a), forming a small pocket or cavity near the inner groove of the protein (yellow region).
In view of its disordered nature, the predicted protein structure is expected to be unstable and to unfold spontaneously, altering the orientation and drug accessibility of the pocket. To explore the protein unfolding trajectory, we simulated the predicted structure of fully solvated GFAP for 0.5 μs (500 ns). The volume of the druggable pocket ( Figure  2b), and several measures of atomic positional variation (see below), provide useful descriptors of structural change. These analyses support the anticipated unfolding of the The observed differential phosphorylation of GFAP in AD aggregates ( Figure 1) may be expected to impact GFAP structural dynamics. To evaluate this possibility, we computationally "mutated" the observed phosphorylated sites to glutamic acid (phosphomimetic substitutions) and simulated the resulting structures for 0.5 µs. The results support the expectation that GFAP hyperphosphorylation in the AD brain is likely to alter its structural stability. Root mean square deviation (RMSD) of the GFAP atomic coordinates over this time interval is consistently lower for the unmodified molecule, "AMC(3,3) GFAP", than for "AD(3,3) GFAP", a phosphomimetic structure similar to that observed in AD(3,3) aggregates ( Figure 2d). "AD(4,4) GFAP" (a phosphomimic of GFAP observed in AD(4,4) aggregates) is initially a little more variable than AMC(3,3) GFAP but achieves a relatively stable structure, which is maintained from~235 ns onward (Figure 2d).
We also plotted the average root mean square fluctuation (RMSF) of individual residues over time, which indicates moderate to high positional fluctuation across the GFAP molecule, for both AMC(3,3) and AD(3,3) structures (Figure 2e), in contrast to the relative rigidity of the AD(4,4) phosphomimetic structure. Together, these data (Figure 2d,e) support the prediction that AD-associated differential phosphorylations can modify GFAP aggregates) is initially a little more variable than AMC(3,3) GFAP but achieves a relatively stable structure, which is maintained from ~235 ns onward (Figure 2d).
We also plotted the average root mean square fluctuation (RMSF) of individual residues over time, which indicates moderate to high positional fluctuation across the GFAP molecule, for both AMC(3,3) and AD(3,3) structures (Figure 2e), in contrast to the relative rigidity of the AD(4,4) phosphomimetic structure. Together, these data (Figure 2d,e) support the prediction that AD-associated differential phosphorylations can modify GFAP structure, somewhat unstably in ApoE(3,3), but creating a relatively invariant conformation in ApoE(4,4).

ROCK1 Is Increased in ApoE-Expressing Glioblastoma Cells
The Rho-associated protein kinase 1 (ROCK1) protein level is elevated in mild cognitive impairment and AD, and its reduction by hemizygous knockout blunted the high amyloid levels seen in a mouse model of AD [45]. Since we observed a significant benefit from reducing ROCK1 levels both in human cells in culture and in intact C. elegans aggregation models, we measured its levels in T98G glioblastoma cells that overexpress either the APOE3 or APOE4 allele. ROCK1 protein levels were at least 6-fold higher in human glial cells that overexpress the APOE4 allele than in the same cells expressing an APOE3 transgene (p < 0.0001; Figure 4a,b), potentially contributing to the additional GFAP phosphorylation sites observed in AD aggregates from ApoE(4,4) vs. ApoE(3,3) individuals (Figure 1b,c). We also compared the impact of kinase-targeted knockdowns in C. elegans models of protein aggregation, using RNAi constructs to silence the closest nematode orthologs of these human kinases. We used aggregation models that simulate Huntington's disease via polyglutamine array aggregation (strain AM141) and Alzheimer's disease using neuronal expression of Aβ 1-42 peptide to form amyloid (CL2355) or muscle expression of human tau to form toxic aggregates that lead to paralysis (VH255). In the Huntington model, total aggregate intensity per worm (Supplementary Figure S1a), i.e., the product of aggregate count per worm and mean YFP fluorescence per aggregate, was reduced 50-60% by knockdown of C. elegans genes orthologous to BARK/GRK or ROCK1 (each p < 0.00005), similar to the effect of RNAi against ifp-1 (an intermediate filament protein with extensive homology to human GFAP). In a C. elegans model of AD-like neuronal amyloidosis (CL2355), chemotaxis declined after neuronal induction of an Aβ transgene, causing fewer worms to migrate toward n-butanol (chemo-attractant). RNAi knockdowns of ROCK1, AKT2, or BARK/GRK orthologs (let-502, akt-2, and grk-2, respectively) rescued the chemotaxis defect at least as well as KD of ifp-1, largely homologous to human GFAP (Supplementary Figure S1b). These C. elegans results are similar to those observed in the human glioblastoma cell line, T98G, in which siRNA KD of ROCK1, AKT2, or PKA reduced total aggregate protein by 50-60% (Supplementary Figure S1c).

ROCK1 Is Increased in ApoE-Expressing Glioblastoma Cells
The Rho-associated protein kinase 1 (ROCK1) protein level is elevated in mild cognitive impairment and AD, and its reduction by hemizygous knockout blunted the high amyloid levels seen in a mouse model of AD [45]. Since we observed a significant benefit from reducing ROCK1 levels both in human cells in culture and in intact C. elegans aggregation models, we measured its levels in T98G glioblastoma cells that overexpress either the APOE3 or APOE4 allele. ROCK1 protein levels were at least 6-fold higher in human glial cells that overexpress the APOE4 allele than in the same cells expressing an APOE3 transgene (p < 0.0001; Figure 4a,b), potentially contributing to the additional GFAP phosphorylation sites observed in AD aggregates from ApoE(4,4) vs. ApoE(3,3) individuals (Figure 1b,c).

Computational Screening Identifies Novel Small Molecules Predicted to Bind Stably to GFAP
To identify novel GFAP-specific inhibitors, we screened structures from the Chem-Bridge library, comprising ~750,000 small molecules, using in silico docking simulations. For target-based docking, we chose the predicted druggable pocket in the transitional GFAP structure used previously (Figure 2c, at 200 ns). To improve the predictive throughput, we conducted computational screening in three stages, increasing docking stringencies at each stage (see Methods), within the Schrödinger Glide docking module [54]. We first performed virtual docking of the entire ChemBridge library in high-throughput mode, followed by redocking of the top 1% of lead molecules in standard precision mode. The top 1% of molecules emerging from stage 2 (74 structures) were analyzed for implicitsolvent-based free energy of interaction, using Schrödinger's MM-GBSA module [54]. The molecules predicted to have the lowest ΔGbinding to GFAP were pursued in vivo for their impact on protein aggregation. This three-stage procedure for simulated docking produced a set of molecules predicted to bind avidly and stably to the GFAP target pocket. The three best candidates (labeled MSR1, MSR2, and MSR3) were predicted to have ΔGbinding surpassing -46 kcal/mol and to fit well within the modeled GFAP pocket (Figure 2c). Counter-screening for binding to tubulin predicted negligible affinity of these drugs to α or β tubulins, or to oligomers of α and β tubulin, although most other top-ranked drugs were predicted to bind tubulin as avidly as they bound GFAP (data not shown).

A Top-Ranked Drug Suppresses In Vitro and In Vivo Aggregation as Effectively as Knockdown of GFAP and Suppresses the Same Co-Aggregate Constituents
Three lead compounds with the lowest predicted ΔGbinding values (Figure 5a) were

Computational Screening Identifies Novel Small Molecules Predicted to Bind Stably to GFAP
To identify novel GFAP-specific inhibitors, we screened structures from the Chem-Bridge library, comprising~750,000 small molecules, using in silico docking simulations. For target-based docking, we chose the predicted druggable pocket in the transitional GFAP structure used previously (Figure 2c, at 200 ns). To improve the predictive throughput, we conducted computational screening in three stages, increasing docking stringencies at each stage (see Methods), within the Schrödinger Glide docking module [54]. We first performed virtual docking of the entire ChemBridge library in high-throughput mode, followed by redocking of the top 1% of lead molecules in standard precision mode. The top 1% of molecules emerging from stage 2 (74 structures) were analyzed for implicitsolvent-based free energy of interaction, using Schrödinger's MM-GBSA module [54]. The molecules predicted to have the lowest ∆G binding to GFAP were pursued in vivo for their impact on protein aggregation. This three-stage procedure for simulated docking produced a set of molecules predicted to bind avidly and stably to the GFAP target pocket. The three best candidates (labeled MSR1, MSR2, and MSR3) were predicted to have ∆G binding surpassing -46 kcal/mol and to fit well within the modeled GFAP pocket (Figure 2c). Counter-screening for binding to tubulin predicted negligible affinity of these drugs to α or β tubulins, or to oligomers of α and β tubulin, although most other top-ranked drugs were predicted to bind tubulin as avidly as they bound GFAP (data not shown).

A Top-Ranked Drug Suppresses In Vitro and In Vivo Aggregation as Effectively as Knockdown of GFAP and Suppresses the Same Co-Aggregate Constituents
Three lead compounds with the lowest predicted ∆G binding values (Figure 5a) were pursued for experimental validation in a human cell culture model of neurodegenerative amyloidosis and in C. elegans whole-animal models of AD-like aggregation. MSR3 was found in initial testing to be cytotoxic to neuroblastoma cells and to delay C. elegans development at all doses tested (data not shown) and was therefore not pursued. MSR1 and MSR2 were initially tested for in vivo efficacy in SH-SY5Y-APP Sw neuroblastoma cells; aggregation in this model was especially well suppressed by MSR1 (Figure 5b,c). In all assays, MSR1 was superior or similar in efficacy to MSR2. Figure 5b shows SH-SY5Y-APP Sw neuroblastoma cells stained for amyloid with thioflavin T after exposure to vehicle (control) or MSR1; in multiple experiments, thioflavin fluorescence declined approximately 2-fold in MSR1-treated cells. Total sarkosyl-insoluble aggregate proteins were isolated and separated on acrylamide-SDS gels, and protein was then stained with SYPRO Ruby. Gel lanes (Figure 5d, left to right) show vehicle-treated control cells, cells treated for 48 h with short interfering RNA (siRNA) targeting GFAP, or cells treated for 48 h with MSR1 or MRS2, drugs predicted to bind stably and selectively to GFAP. GFAP siRNA suppressed aggregate protein by 65-80%, while MSR1 provided 60-75% suppression, but MSR2 did not significantly reduce the amount of aggregate protein.
Proteomic identification of proteins in sarkosyl-insoluble aggregates revealed that proteins completely excluded from such aggregates by MSR1 treatment coincided remarkably well (87% concordance) with those eliminated by siRNA knockdown of GFAP (Figure 5e), whereas four proteins were found to be substantially upregulated by both treatments. The excluded group (which includes synapsin-1, plectin, dynactin-1, MAP1A, MAP2, and αtubulin) has a correlation coefficient of 0.77 between the effects of MRS1 and GFAP siRNA (p < 3 × 10 −280 ; Figure 5f). This observation of highly concordant and correlated effects of MSR1 and GFAP RNAi on aggregate composition provides compelling evidence that GFAP is the chief functional target by which MSR1 lowers aggregation.
Drugs MSR1 and MSR2 were also tested in several C. elegans models of human neurodegenerative diseases. In a tauopathy model strain (VH255) expressing normal human tau in C. elegans muscle, tau aggregation caused paralysis that was alleviated to a similar extent by treatment with 1 µM MSR1 or siRNA against ifp-1, the closest nematode homolog of GFAP (Supplementary Figure S1d). MSR1 also effected significant rescue of chemotaxis in a C. elegans model of neuronal amyloidosis (CL2355), in which migration toward a chemo-attractant was impaired by pan-neuronal expression of human Aβ 1-42 . The addition of 0.1 µM MSR1 restored chemotaxis to~90% (Supplementary Figure S1e), the same level as in wild-type or uninduced worms of the same age (not shown). Fluorescent muscle aggregates accumulate with age in a strain expressing Q40::YFP in muscle, mimicking the polyglutamine array expansion threshold for huntingtin protein, i.e., the array length sufficient to elicit symptoms of Huntington's disease in humans and paralysis in nematodes. Supplementary Figure S1f shows aggregate intensity per worm at 5 days of age post-hatch, reduced~50% by 10 µM MSR1, vs. 35% at 0.1 µM (each treatment differing from vehicle-only controls at p < 0.005). GFAP, one of the numerous proteins enriched in AD aggregates relative to agematched controls [29], now joins a small set of neuropathology-associated proteins that display disease-specific hyperphosphorylation. These include tau (AD, PD, ALS), Aβ1-42 (AD), TDP43 (ALS), and α-synuclein (PD) [29,[58][59][60][61][62][63]. We identified three GFAP serines that are highly phosphorylated in aggregates isolated from hippocampi of AD (3,3) individuals and additional serines and a threonine near the N-and C-termini for which phosphorylation was observed only in AD(4,4) aggregates (Figure 1c). We examined only

Discussion
Aging is the most influential non-genetic risk factor for a variety of dementias and also for many other adult onset diseases that impose significant burdens on aging adults and on our healthcare system [55]. Neurodegenerative diseases, such as AD, Parkinson's, and amyotrophic lateral sclerosis (ALS), as well as conditions as diverse as hypertension [35], sarcopenia [37], and even several adult cancers [56], all show accrual of distinctive aggregate foci featuring disease-specific "diagnostic" proteins. We identified many proteins within immunopurified aggregate subsets by proteomics [29] and defined protein-protein interfaces by intra-aggregate cross-linking and recovery of linked peptides [57].
GFAP, one of the numerous proteins enriched in AD aggregates relative to age-matched controls [29], now joins a small set of neuropathology-associated proteins that display disease-specific hyperphosphorylation. These include tau (AD, PD, ALS), Aβ 1-42 (AD), TDP43 (ALS), and α-synuclein (PD) [29,[58][59][60][61][62][63]. We identified three GFAP serines that are highly phosphorylated in aggregates isolated from hippocampi of AD(3,3) individuals and additional serines and a threonine near the N-and C-termini for which phosphorylation was observed only in AD(4,4) aggregates (Figure 1c). We examined only APOE3 or APOE4 homozygous tissue, but it is reasonable to expect an intermediate outcome for APOE3/E4 heterozygotes. We also observed three methionine oxidation sites unique to AD-aggregate GFAP (Figure 1c), which may reflect either a more oxidative cell environment or greater misfolding of GFAP molecules in AD hippocampus.
The sites we observed for GFAP phosphorylation in AD are consistent with the known targets of several kinases previously implicated in AD pathogenesis. These include AKT2 (one of two mammalian AKT paralogs), Rho-associated Kinase 1 (ROCK1), Gprotein-coupled Receptor Kinase 2 (BARK/GRK2), and Protein Kinase A (PKA). Altered insulin signaling has been implicated in AD, and AKT is a key downstream effector of the kinase cascade that conveys insulin/insulin-like signaling [64][65][66]. GRK2, also known as BARK, was also shown to play a role in the development of cardiovascular disease. PKA, a cAMP-dependent kinase, is involved in multiple signaling pathways, contributes to tau hyperphosphorylation, and has been implicated in progression of several neurodegenerative disorders, including AD, PD, and HD [49,[67][68][69][70]. PKA was also shown to play roles in diabetes [70] and anxiety-related behavior [71].
Knockdowns of several kinases that could participate in the observed GFAP hyperphosphorylations decreased protein aggregation and associated physiological declines in C. elegans and in human neuroblastoma cells expressing amyloid precursor protein. We note that several of these same kinases have putative target sites in other AD-associated proteins, such as microtubule-associated protein tau, which are expected to further amplify their impact. Multiple AD-promoting targets of these kinases are suggested by results shown in Figure 3c,e, in which some kinase knockdowns provide more effective rescue from AD-like traits than knockdown of GFAP itself. We note, however, that the neuronal efficacy of siRNA knockdowns was not monitored in these experiments and is typically lower in neurons than in other target cells. While the noted comparisons may thus be misleading, a reduced siRNA efficacy in neurons would lead to underestimation of the target's impact.
Based on our data, GFAP provides a novel target to relieve aggregate burden in AD and possibly other aggregation-associated diseases. We therefore screened for small molecules that specifically target GFAP, beginning with the three-dimensional structure of GFAP predicted via powerful fold recognition and ab initio procedures [44]. The initial lowest-∆G structure included a small druggable pocket (Figure 2a). Molecular dynamic simulations of the GFAP structure over time revealed a transition to a metastable state in which the binding cavity was expanded~3-fold over its initial volume. The molecular structure of GFAP at 200 ns (Figure 2c) was selected as the target for drug screening. Several descriptors (RMSD, RMSF, and pocket volume) were used to monitor GFAP structural change during the simulation.
Proteomics of brain aggregates indicated strikingly differential phosphorylation of GFAP from AD tissue [29]. Although phosphomimetic substitutions are not the perfect mimics of actual protein phosphorylation, the computer simulations of the predicted AD (3,3) and AD(4,4) structures, incorporating pseudo-phosphorylated sites, as observed, were fully consistent with the hypothesis that phosphorylation status can alter GFAP structural dynamics. Simulation data do not permit us to infer the extent to which phosphorylation of any individual kinase target is responsible for destabilizing the GFAP structure or might alternatively favor accessibility of subsequently modified kinase sites.
We note that our strategy for computational screening of large drug structure libraries has several novel features. These include a three-tier docking screen with progressively increasing stringency and the elimination of drugs predicted to display off-target affinity toward tubulin chains, a common property we observed for most GFAP-binding drugs. We tested the top three candidates emerging from three successive screens for GFAP binding, plus a counter-screen to eliminate tubulin-binding drugs, for efficacy in a variety of aggregation model systems. Of these, MSR1 displayed an efficacy close to that of GFAP knockdown (using RNAi to GFAP or its closest nematode homolog) in each assay, and proteomic analysis of aggregates revealed a near-identical set of proteins depleted or eliminated from aggregates.

Conclusions
The present study is the first to demonstrate the impact of GFAP and its predicted upstream kinases on protein aggregation in human-cell and C. elegans models of neuropathic aggregation and the first successful in silico screen for a GFAP-specific drug, coupled to in vivo demonstrations of its anti-aggregative efficacy. Because the protective effects of MSR1 in nematode models of neurodegenerative pathology are paralleled by similarly positive results in cultured human neuroblastoma cells, GFAP and AD-associated kinases hold promise as novel targets for drug interventions to ameliorate AD-like neuropathies.
Supplementary Materials: The following are available online at https://www.mdpi.com/article/ 10.3390/pharmaceutics14071354/s1, Figure S1: MSR1 treatment attenuates aggregates and its associated phenotypic traits in C. elegans models of AD, Table S1: Kinases predicted to be capable of phosphorylating GFAP sites.  Institutional Review Board Statement: Ethical review and approval were waived for this study, due to use exclusively of postmortem tissue samples devoid of any personal identifiers.

Informed Consent Statement:
There were no human subjects involved in this study. Data Availability Statement: Detailed data will be provided upon request, unless prevented by intellectual property considerations.