Structure-based discovery of fiber-binding compounds that reduce the cytotoxicity of amyloid beta

Open accessCopyright infoDownload PDFRelated contenteLife Lens

Structure-based discovery of fiber-binding compounds that reduce the cytotoxicity of amyloid beta

Affiliation details

DOI: http://dx.doi.org/10.7554/eLife.00857Published July 16, 2013 Cite as eLife 2013;2:e00857

Abstract

Amyloid protein aggregates are associated with dozens of devastating diseases including Alzheimer’s, Parkinson’s, ALS, and diabetes type 2. While structure-based discovery of compounds has been effective in combating numerous infectious and metabolic diseases, ignorance of amyloid structure has hindered similar approaches to amyloid disease. Here we show that knowledge of the atomic structure of one of the adhesive, steric-zipper segments of the amyloid-beta (Aβ) protein of Alzheimer’s disease, when coupled with computational methods, identifies eight diverse but mainly flat compounds and three compound derivatives that reduce Aβ cytotoxicity against mammalian cells by up to 90%. Although these compounds bind to Aβ fibers, they do not reduce fiber formation of Aβ. Structure-activity relationship studies of the fiber-binding compounds and their derivatives suggest that compound binding increases fiber stability and decreases fiber toxicity, perhaps by shifting the equilibrium of Aβ from oligomers to fibers.

DOI: http://dx.doi.org/10.7554/eLife.00857.001

eLife digest

Alzheimer’s disease is the most common form of dementia, estimated to affect roughly five million people in the United States, and its incidence is steadily increasing as the population ages. A pathological hallmark of Alzheimer’s disease is the presence in the brain of aggregates of two proteins: tangles of a protein called tau; and fibers and smaller units (oligomers) of a peptide called amyloid beta.

Many attempts have been made to screen libraries of natural and synthetic compounds to identify substances that might prevent the aggregation and toxicity of amyloid. Such studies revealed that polyphenols found in green tea and in the spice turmeric can inhibit the formation of amyloid fibrils. Moreover, a number of dyes reduce the toxic effects of amyloid on cells, although significant side effects prevent these from being used as drugs.

Structure-based drug design, in which the structure of a target protein is used to help identify compounds that will interact with it, has been used to generate therapeutic agents for a number of diseases. Here, Jiang et al. report the first application of this technique in the hunt for compounds that inhibit the cytotoxicity of amyloid beta. Using the known atomic structure of the protein in complex with a dye, Jiang et al. performed a computational screen of 18,000 compounds in search of those that are likely to bind effectively.

The compounds that showed the strongest predicted binding were then tested for their ability to interfere with the aggregation of amyloid beta and to protect cells grown in culture from its toxic effects. Compounds that reduced toxicity did not reduce the abundance of protein aggregates, but they appear to increase the stability of fibrils. This is consistent with other evidence suggesting that small, soluble forms (oligomers) of amyloid beta that break free from the fibrils may be the toxic agent in Alzheimer’s disease, rather than the fibrils themselves.

In addition to uncovering compounds with therapeutic potential in Alzheimer’s disease, this work presents a new approach for identifying proteins that bind to amyloid fibrils. Given that amyloid accumulation is a feature of many other diseases, including Parkinson’s disease, Huntington’s disease and type 2 diabetes, the approach could have broad therapeutic applications.

DOI: http://dx.doi.org/10.7554/eLife.00857.002

Main text

Introduction

Protein aggregates, both amyloid fibers and smaller amyloid oligomers, have been implicated in the pathology of Alzheimer’s and other neurodegeneration diseases (Chiti and Dobson, 2006; Eisenberg and Jucker, 2012). The increasing prevalence of Alzheimer’s disease in our aging societies, the associated tragedy for patients and their families, and the mounting economic burden for governments have all stimulated intense research into chemical interventions for this condition. Much work has been focused on screening compounds that prevent aggregation and the associated cytotoxicity of the amyloid β-peptide (Aβ) (reviews by Sacchettini and Kelly, 2002; Bartolini and Andrisano, 2010; Hard and Lendel, 2012).

Screens have often focused on natural products from plants and lichens. These include polyphenols, such as epigallocatechin gallate (EGCG) from green tea (Ehrnhoefer et al., 2008) and curcumin from the spice turmeric (Yang et al., 2005). These natural polyphenolic compounds show inhibition on the fibrillation of a variety of amyloid proteins, including Aβ40 as well as α-synuclein, IAPP and PrP (Porat et al., 2006; Dasilva et al., 2010; Ono et al., 2012). Several dyes have also been found to ameliorate amyloid toxicity. Orcein from lichens appears to diminish toxic oligomers and enhance fiber formation (Bieschke et al., 2011). Congo red, thioflavin T and their analogs, commonly used as staining reagents for amyloid detection, exhibit ameliorative effects on neurodegenerative disorders, such as Alzheimer’s, Parkinson’s, Huntington’s, and prion diseases (Frid et al., 2007; Alavez et al., 2011), however their application is limited by significant side effects (Klunk et al., 2004).

Additional screens have identified a variety of molecules, including proteins (Evans et al., 2006), antibodies (Kayed et al., 2003; Ladiwala et al., 2012), synthetic peptide mimetics (Findeis, 2002; Kokkoni et al., 2006; Takahashi and Mihara, 2008; Cheng et al., 2012) and small molecules (Wood et al., 1996; Williams et al., 2005; McLaurin et al., 2006; Necula et al., 2007; Bartolini and Andrisano, 2010; De Felice et al., 2001; Ladiwala et al., 2011; Hard and Lendel, 2012; Kroth et al., 2012), that inhibit Aβ fibrillogenesis and/or Aβ-associated cytotoxicity in vitro. While most efforts have targeted the deposition of Aβ fibers as the hallmark of Alzheimer’s, smaller amyloid oligomers are now receiving greater attention as the possible toxic entities in Alzheimer’s and other neurodegenerative diseases (Hartley et al., 1999; Cleary et al., 2005; Silveira et al., 2005). Furthermore, emerging evidence suggests that mature, end-stage amyloid fibers may serve as a reservoir, prone to releasing toxic oligomer (Xue et al., 2009; Cremades et al., 2012; Krishnan et al., 2012; Shahnawaz and Soto, 2012). Recent screens have identified compounds that reduce Aβ cytotoxicity, without interfering with Aβ fibrillation (Chen et al., 2010) or promoting the formation of stable Aβ aggregates (Bieschke et al., 2011).

Structural information about protein targets often aids drug development, so here we take a structure-based approach, combined with computational screening, to discover amyloid interacting compounds that reduce amyloid toxicity. This approach has been enabled by the determination of atomic structures of the adhesive segments of amyloid fibers, termed steric zippers (Nelson et al., 2005), and of solid state NMR-based structures of amyloid fibers (such as full-length Aβ fibers [Luhrs et al., 2005; Petkova et al., 2005] and the HET-s prion domain complexed with Congo Red [Schutz et al., 2011]). The steric zipper structures reveal a common motif for the spine of amyloid fibers, in which a pair of fibrillar β-sheets is held together by the side-chain interdigitation (Sawaya et al., 2007). We focus on Aβ, a peptide of 39–42 residues cleaved from the Amyloid precursor protein (APP) associated with Alzheimer’s, as a target for inhibitor discovery. The segment Aβ16–21 with the sequence KLVFFA is an amyloid-forming peptide, which packs in a steric zipper form, and has been identified as the spine of the full-length Aβ fiber (Luhrs et al., 2005; Petkova et al., 2006; Colletier et al., 2011). Co-crystal structures have been determined for small molecules in complex with the fibrillar β-sheets of Aβ16–21 (Landau et al., 2011). One of these structures—Aβ16–21 with the dye Orange G—reveals the specific pattern of hydrogen bonds and apolar interactions between orange G and the steric zipper: the negatively charged dye binds specifically to lysine side chains of adjacent sheets, and its planar aromatic portion packs against apolar residues (phenylalanine and valine) of adjacent sheets. By creating a tight, low energy interface across several β-strands within fiber core, this fiber-binding molecule appears to stabilize the fiber structure. With this atomic structure as a basis, we are able to screen for small molecular compounds that bind to amyloid fibers, stabilizing them and possibly reducing amyloid toxicity. Applying our structure-based screening procedure, we screen computationally for compounds that bind to Aβ fibers, termed BAFs (Binders of Amyloid Fibers) and then experimentally test their effects on Aβ aggregation and cytotoxicity.

Results

Structure-based screening procedure

We have devised a structure-based procedure for the identification of small molecules that bind to amyloid and affect amyloid toxicity (Figure 1). The procedure starts from a co-crystal structure of a ligand bound to an amyloidogenic segment of Aβ (Landau et al., 2011), the dye orange G bound to the fiber-like crystal structure of KLVFFA(Aβ16–21) segment. This structure reveals the chemical environment or ‘pharmacophore’ presented by the ligand binding site of this Aβ segment, that is, orange G binds to stacked β-sheets of Aβ. Knowledge of the amyloid pharmacophore (Figure 1A) permitted us to screen for compounds that could be expected to bind in this chemical environment, possibly stabilizing amyloid fibers.

Construction of compound libraries for computational screening

For assembling the compounds in our screening library, we sought three characteristics: (a) commercially available compounds since we intended to follow the in silico screening with experimental validation; (b) compounds with known three-dimensional structures such that our screening would be as realistic as possible; (c) generally flat compounds able to bind to the β-sheets of the steric zipper, as does orange G. Some ∼11,000 compounds having the first two characteristics (CSD-ZINC set) were selected as the intersection of molecules found both in the Cambridge Structure Database (http://www.ccdc.cam.ac.uk) and the Zinc Database of purchasable compounds (http://zinc.docking.org/) (Irwin and Shoichet, 2005). This CSD-ZINC set spans a variety of structural shapes and molecular properties. A second set of ∼7000 compounds, the Flat Compound Set, was gathered from the ZINC database to include molecules expected to bind to the flat surface of a steric zipper. The members of this set contain multiple aromatic rings or one aromatic ring with additional planar groups.

Computational screening of compounds that bind to Aβ fibers

Computational screening was carried out with the RosettaLigand program (Davis and Baker, 2009), after adapting its docking approach to carry out high-throughput screening (Figure 2). The conformational flexibilities of ligand and protein side chains are in a ‘near-native’ perturbation fashion, meaning that the fine sampling of conformations was restrained to be close to the starting conformation. A balance was achieved between extensive sampling and the speed required for screening a large compound library by fine sampling of side chain and ligand torsion angles only around their starting conformations, as illustrated by sticks in Figure 2C.

In the screening steps of computational docking (Figure 2A), a library of ∼18,000 purchasable compounds (Sets 1 and 2) was scanned computationally for structural compatibility with the pharmacophore (ligand binding site) presented by a single sheet of the Aβ16–21 steric zipper. Structural compatibility was assessed by a combination of binding energy (Meiler and Baker, 2006) and steric complementarity (Lawrence and Colman, 1993). After computational docking, the distribution of calculated binding energies suggests that, statistically the flat compounds from Set 2 fit more snugly on the flat surfaces of Aβ16–21 fibers than those with diverse shapes in Set 1 (Figure 2B). The best scoring compounds were screened further by requiring that each is also structurally compatible with the solid-state NMR-derived model of the Aβ full-length fiber structure (Petkova et al., 2006) (Figure 1C and Figure 1—figure supplement 3).

Experimental characterization of BAFs

After in silico screening of a library of ∼18,000 purchasable compounds, twenty-five of the top-ranking compounds all with better scores for binding energy and steric complementarity than orange G (Figure 1D, Figure 2—figure supplement 1), were selected for experimental validation. First these 25 compounds were tested for their ability to protect mammalian cells from Aβ toxicity (Figure 1E, Tables 1 and 2), and five of them were found to reduce the toxic effects of Aβ. These five were tested for binding to both Aβ1–42 and Aβ16–21 fibers by NMR. Two were found to have tighter binding than orange G, and the others gave insufficient NMR signals for detection. To expand this set of the five compounds, a second cycle of inhibitor discovery was performed. From the computed positions of the five compounds, a refined pharmacophore was inferred (Figure 1F), and used in the next cycle of screening. Added to the compound set were nine additional compounds apparently related to the five lead compounds from the initial cycle, plus 17 chemical derivatives of compounds (Tables 1 and 3). The second cycle produced three additional compounds and three compound derivatives that also protected the mammalian cells from Aβ fibers. One of these compounds was confirmed by NMR to bind to Aβ fibers. The detailed description of those experimental results is as follows.

View this table:
Table 1.

List of all tested BAF compounds

DOI: http://dx.doi.org/10.7554/eLife.00857.010

View this table:
Table 2.

Detailed list of the active BAF compounds

DOI: http://dx.doi.org/10.7554/eLife.00857.011

View this table:
Table 3.

List of the representative BAFs 11, 30, 31 and their derivatives

DOI: http://dx.doi.org/10.7554/eLife.00857.012

Inhibition of Aβ1–42 toxicity by BAFs

Having identified compounds that bind Aβ fibers, by a structure-based procedure, we tested their effects on the cytotoxicity of Aβ1–42 fiber against two mammalian cell lines: PC12 and HeLa (Figure 3). Five BAFs—1,4,8,11, and 12—in the initial cycle and three additional BAFs—26, 30, and 31—from the second cycle, with diversified chemical structures shown in Figure 4, significantly increased both PC12 and HeLa cell survival after 24 hr incubation with Aβ1–42 (0.5 µM) at concentration of 2.5 µM, while the BAFs alone had little or no effect on cell survival (Figure 3—figure supplement 1). Three BAFs—11, 26, and 31—showed clear dose-response profiles in their protection of both PC12 and HeLa cells (Figure 3B). Among them, the two best BAFs—26 and 31—were tested and did not affect the cytotoxicity of amyloid fibers other than Aβ (Figure 3—figure supplement 2). Although all of these BAFs provide protection against Aβ toxicity, none diminish the amount of Aβ fibers in electron micrographs (Figure 3C).

View this table:
Table 4.

Student’s t-test and p value analysis suggests that BAFs reduce the cytotoxicity of Aβ fibers significantly

DOI: http://dx.doi.org/10.7554/eLife.00857.016

Figure 4.
Download figureOpen in new tabDownload powerpointFigure 4. Diversified chemical structures of 8 active BAF compounds that reduce Aβ toxicity.

Orange G in an orange box is also displayed for comparison.

DOI: http://dx.doi.org/10.7554/eLife.00857.017

Validation of compound binding by NMR titration

Promising candidate binders from in silico screening and toxicity tests were validated by titration of Aβ fibers into solutions of each compound, as monitored by NMR signals of aromatic protons of the compound (Figure 5). The proton resonances of the freely rotating compounds disappear as the compound binds to the fibers. By increasing the amount of fibers, an apparent Kd for compound binding can be estimated. From in silico screening, all tested BAF compounds are calculated to bind more tightly to Aβ fibers than orange G. In NMR studies, the apparent Kd of orange G binding to Aβ16–21 fibers was found to be 43 ± 21 µM, whereas the apparent Kd of BAF1 binding to Aβ16–21 fibers is 12 ± 7 µM. BAFs were found to bind to both Aβ16–21 fibers and Aβ1–42 fibers. Figure 5F shows a notable correlation between the calculated binding energies and the reduction in NMR peak areas upon Aβ binding. That is, all BAFs with predicted binding energy better than orange G also reduce NMR peak areas more than orange G. On the other hand, BAF31ΔOH, a derivative of BAF31 by removal of a key hydroxyl group essential for binding, exhibits both a worse calculated binding energy and a diminished reduction of NMR peak upon titration of Aβ1–42 fibers.

View this table:
Table 5.

Predicted binding energy and experimental measurement of the binding of two BAFs and orange G against both Aβ16–21 (KLVFFA) and full-length Aβ fibers

DOI: http://dx.doi.org/10.7554/eLife.00857.022

View this table:
Table 6.

Comparison of the measured binding parameters of the representative BAFs with orange G by NMR titrations

DOI: http://dx.doi.org/10.7554/eLife.00857.023

Structure-activity relationship studies of the Aβ pharmacorphore

Based on the lead compounds found in the initial cycle of the procedure, we carried out a second cycle to expand our understanding of the Aβ pharmacorphore. BAF11 (Figure 6A), one of the lead compounds in the initial cycle, was used to perform structure-activity relationship studies. Twelve derivatives of BAF11 were scanned to pinpoint the essential apolar and polar interactions for the pharmacorphore refinement (Figure 6B, Figure 6—figure supplement 1). These derivatives are grouped in five classes, whose effects on Aβ toxicity have been tested (Figure 6C). Classes I and II assess the polar region of BAF11, which makes hydrogen bonds to charged Lys16 ladders of the Aβ fiber: the deletion of the hydroxyl group (Class I) significantly decreased the inhibition of toxicity; the swapping of the hydroxyl group with the aromatic tail (Class II) almost abolished inhibition of toxicity. Classes III, IV, and V focused on the aromatic moieties of BAF11: altering the sizes of aromatic groups (Class III) showed little change in inhibition of toxicity while adding charged or polar groups within aromatic region (Classes IV and V) resulted in a significant decrease of inhibition of toxicity. These differences among BAF11 derivatives in inhibition of toxicity (Figure 6C) further validated our structure-based approach and provided guidelines for the refinement of Aβ pharmacophore.

In the second cycle, nine new compounds were derived from the refined pharmacophore (Figure 7). Three of them detoxified Aβ in cell survival assay. BAF31, the best inhibitor which protected mammalian cells from Aβ toxicity in the second cycle, increased cell survival from the 40% induced by Aβ alone to >90% (Figure 3). A derivative of BAF31, BAF31ΔOH, lacking the hydroxyl group believed to bind to the Lys residue of the Aβ fiber (shown by the magenta oval in Figure 8B), is calculated no longer to bind to the Aβ fiber. NMR and cell viability assessments indicated that BAF31ΔOH binds much less strongly to Aβ fibers than BAF31 itself and shows significantly reduced power to inhibit toxicity (Figure 8E). Similarly, the detoxifying profile of derivatives of another inhibitor, BAF30, validated the key interactions of BAF30 across the binding interface (Figure 9). Our conclusion is that the NMR binding and toxicity results for the BAF derivatives studied are consistent with our model for the pharmacophore of Aβ (Figure 10).

Figure 7.
Download figureOpen in new tabDownload powerpointFigure 7. New BAFs derived from the refined amyloid pharmacophore.

(A). Amyloid pharmacophore based on the structural overlay of active BAFs and derivatives. The overlay of the lead compounds from the initial round (BAF4, BAF8, and BAF11) elucidated the consensus of polar and nonpolar interactions at fiber binding interfaces, which sheds light on the amyloid pharmacophore. The amyloid pharmacophore was further refined by iterative approaches of computational docking and experimental testing. The derivatives of those lead compounds were tested to explore the essential role of those consensus interactions, and the differences of binding patterns and toxicity inhibition effects of the BAF derivatives can provide a guideline for the further refinement of amyloid pharmacophore. (B). New BAFs were ‘designed’ based on the refined pharmacophore. One successful example, BAF31 (green sticks) derived from the pharmacophore (grey sticks), showed the enhanced capability of inhibiting Aβ toxicity (Figure 8C). The success of developing enhanced binder from pre-defined pharmacophore highlights the important role of iterative docking/test approach in structure-based drug development.

DOI: http://dx.doi.org/10.7554/eLife.00857.026

Figure 9.
Download figureOpen in new tabDownload powerpointFigure 9. Analysis of the lead compound BAF30 and its derivatives.

Structural models of BAF30 (green sticks) docked on Aβ fiber structure (in a light yellow color) are shown in (A and B). The important polar (black hydrogen bonds) interaction between BAF30 and single β-sheet of Aβ fiber, as well as shape complimentary between the aromatic rings of BAF30 and the hydrophobic patches of Aβ fiber are highlighted respectively. Schematic representation of the polar and nonpolar interactions of BAF30 with Aβ fiber is shown in panel (C). The magenta circles highlight two important hydroxyl groups which are absent in BAF30 derivatives. (D). The chemical structure of each derivative is listed. The dark blue open circles indicate the deletion of the important hydroxyl group. The red color in chemical structures indicates the addition of atoms or groups to BAF30. (E). HeLa cell survival rates in the presence of Aβ (0.5 µM monomer equivalent) and BAF30 or the derivatives are compared. The hydrogen bonds between BAF30 and Lys16 residues of Aβ fiber are important for binding of Aβ fiber and inhibition of Aβ toxicity. With additional groups at the opposite side of hydrogen binding sites, the derivative BAF30αR showed little change in toxicity inhibition. However, two BAF30 derivatives (σOHAαOH and σOHAΔOHBαCOO), which alter or delete the two important hydroxyl groups (magenta circles in panel C) of BAF30 that form hydrogen bonds to Lys16, showed a significant decrease in the toxicity inhibition. Furthermore, when BAF30 was modified by shifting both hydroxyl groups (A and B) to their neighboring positions, the derivative BAF30σOHABαCH3 almost lost the inhibition of Aβ toxicity. The rescuing percentage (%) is defined in ‘Materials and methods’.

DOI: http://dx.doi.org/10.7554/eLife.00857.029

Figure 10.
Download figureOpen in new tabDownload powerpointFigure 10. General rule of the essential interactions between BAFs and Aβ fiber can be derived from structure-based screening of Aβ toxicity inhibitor.

The carbonyl group is used to represent the H-bond acceptor (or negative charge) of BAFs, and the naphthalene ring is used to represent the planar aromatic portion of BAFs. Based on the rounds of computing search and experimental test, the detailed description about essential interactions and geometrical parameters are in ‘Materials and methods’.

DOI: http://dx.doi.org/10.7554/eLife.00857.030

Discussion

Structure-based discovery of compounds that bind amyloid fibers

Amyloid fibers differ fundamentally in structure from the enzymes and signaling proteins that are the traditional targets in structure based design of binding compounds, and thus their pharmacophores might be expected to differ fundamentally as would the types of compounds that bind. In general, the binding sites of the traditional targets are often concave pockets; in contrast, the surfaces of amyloid fibers are flat and repetitive along the fiber axis, without well-defined surface cavities. The widely used ligand-docking software, such as DOCK (Ewing et al., 2001), or AutoDock (Morris et al., 2009), is intended to fit well-defined protein pockets rather than shallow grooves at flat fiber surfaces.

Consequently we have adapted the RosettaLigand program (Davis and Baker, 2009) for docking a library of commercially available compounds onto the flat surface of amyloid fibers. Similarly to other software packages, RosettaLigand scores each candidate compound for its energetic fit to its binding site. The initial site is chosen near that occupied by a bound compound, as determined in a crystal structure. The conformational flexibilities of ligand and protein side chains are modeled in a ‘near-native’ perturbation fashion (‘Materials and methods’), meaning that the fine sampling of conformations was restrained to be close to the starting conformation. To find the position along the flat fibrillar surface of greatest binding energy for each candidate compound, our screening approach leverages the rotamer repacking algorithm (Leaver-Fay et al., 2011) and Rosetta energy function (Kuhlman and Baker, 2000) to account for flexibility of protein side chains and ligand, which is critical in modeling of such shallow grooves on the fiber surface.

Our procedure identified 34 BAF compounds predicted to bind to Aβ fibers, among which eight BAFs diminish the toxicity of the fibers in mammalian cells. We suggest that the same procedure can be used to discover other compounds that reduce the toxicity of Aβ fibers, starting from other co-crystal structures of Aβ segments with other bound ligands. Similarly, the same procedure can be applied to the discovery of compounds that bind to other amyloid proteins, for use as either toxicity inhibitors or imaging agents for amyloid diagnosis.

Mechanism of inhibition of Aβ toxicity

Our observation is that our tightest binding BAFs all diminish the toxicity of Aβ fibers, and yet do not substantially diminish the amount of fibers. Further study will be required to understand the molecular mechanism underlying the inhibition of Aβ toxicity, but here we offer the following hypothesis.

Emerging evidence suggests that amyloid oligomers, rather than amyloid fibers, are toxic entities (Hartley et al., 1999; Cleary et al., 2005; Silveira et al., 2005), and that perhaps toxic oligomers can be released from amyloid fibers (Xue et al., 2009; Cremades et al., 2012; Krishnan et al., 2012; Shahnawaz and Soto, 2012). By binding to fibers, BAFs stabilize them, thereby shifting the equilibrium of Aβ molecules from smaller, toxic entities towards the fibrillar state. The BAF compounds in their computationally docked sites on Aβ fibers contact several (as few as three and as many as six) adjacent β-strands of the fiber. By creating a low energy binding interface across several fiber strands, the BAFs apparently stabilize the Aβ fibers from breaking into smaller entities.

From previous studies, we expect BAFs to bind to amyloid fibers rather than oligomers. In recent work (Laganowsky et al., 2012; Liu et al., 2012), we proposed that amyloid forming proteins can enter either of two distinct aggregation pathways, which are separated by an energy barrier. One pathway leads to in-register fibers in which every β-strand lies directly above or below an identical strand in the fiber. The other pathway leads to out-of-register oligomers in which antiparallel β-strands are sheared relative to one another and roll into a β-barrel. We found that three out-of-register amyloid-like structures exhibit cytotoxicity (Laganowsky et al., 2012; Liu et al., 2012), which tend to be transient, equilibrating eventually into in-register fibers. In our approach, we search for BAFs based on in-register β-sheets rather than out-of-register β-strands found in toxic oligomeric structures, to which our BAFs are not expected to bind (Figure 11). We speculate that BAFs stabilize the in-register fibers revealed by our steric zippers, relative to out-of-register toxic oligomers, thereby shifting the equilibrium from toxic oligomers towards fibers (Figure 12). Supporting this is our result that diminished toxicity accompanies compound binding.

View this table:
Table 7.

BAFs reduce Aβ cyto-toxicity by targeting fibers rather than oligomers.

DOI: http://dx.doi.org/10.7554/eLife.00857.033

Figure 12.
Download figureOpen in new tabDownload powerpointFigure 12. Proposed mechanism of how compound binding increases fiber stability and decreases fiber toxicity.

BAFs (green) bind to the side of amyloid fibers, stabilizing the fiber, and shifting the equilibrium from smaller and more toxic oligomers towards fibers. This shift in equilibrium reduces amyloid toxicity.

DOI: http://dx.doi.org/10.7554/eLife.00857.034

BAFs strengthen the hypothesis that Aβ16–21 fibers reflect essential features of full Aβ fibers

The identification of BAFs starts with the atomic structure of orange G bound within the fiber-like crystals of Aβ16–21, because as yet there is no high-resolution atomic structure available for ligands bound to full-length Aβ fibers. Nevertheless, we found that BAFs diminish toxicity of full-length Aβ fibers. This finding suggests that the steric zipper structure of Aβ16–21 fibers recapitulates some of the essential structural features of full-length Aβ fibers. We are currently attempting cocrystallization of BAFs with Aβ16–21 and other steric zipper structures. We speculate that coupled with computational methods, other steric zipper structures could enable the discovery of the lead compounds for inhibitors of other toxic amyloid entities.

Materials and methods

Computational procedures

Two choices of compound libraries for structure-based screening

We generated two sets of purchasable compounds to be screened via the computational docking:

  1. Cambridge Structure Database (CSD) set. 102,236 organic compounds, whose crystal structures have R-factor better than 0.1, were extracted from the Cambridge Structure Database (version 5.32 November 2010) using ConQuest. The SMILES string of each structure was then used to locate its purchasing information among the ZINC purchasable set (http://zinc.docking.org/) (Irwin and Shoichet, 2005) by OpenBabel package (http://openbabel.org/) (Guha et al., 2006). The fast index table of all SMILES strings of the ZINC purchasable set was generated to allow the fast search of each CSD structure against ZINC purchasable set. CSD structures that failed in locating their purchasing information (i.e., without any hit in searching against ZINC purchasable set) were omitted. A library of 13,918 structures from CSD representing 11,057 compounds were finally compiled, whose purchasing information is annotated by ZINC purchasable database. The complete list of CSD/ZINC entries of these compounds in this CSD set can be found in Supplementary file 1.

  2. Flat Compound (FC) set. A library of 6589 compounds containing phenol and less than three freely rotatable bonds were extracted from the ZINC database (http://zinc.docking.org/) (Irwin and Shoichet, 2005). Those compounds have a common feature of planar aromatic ring, resulting in a ‘flat’ compound. The flat compound library includes compounds with similar chemical structures to naturally fiber-binding molecules, for instance, Thioflavin-T (ThT), Congo red, Green tea epigallocatechin-3-gallate (EGCG), and Curcumin. It also includes many natural phenols, such as gallic acid, ferulic acid, coumaric acid, propyl gallate, epicatechin, epigallocatechin, etc. The complete list of ZINC entries of these compounds in this FC set can be found in Supplementary file 2.

Ligand ensemble preparation with near-‘native’ perturbation

Each molecule in our two compound libraries was prepared for the docking simulations. Hydrogen atoms of each molecule were added for the compounds lacking modeled hydrogens using the program Omega (v. 2.3.2, OpenEye) (Bostrom et al., 2003). Ligand atoms were represented by the most similar Rosetta atom type, their coordinates were re-centered to the origin, and their partial charges were assigned by OpenEye’s AM1-BCC implementation. We then generated the ligand perturbation ensemble near the crystal conformation (CSD set) or starting conformation (FC set) of each molecule. For each rotatable bond of the ligand, a small degree torsion angle deviation (±5°) was applied. K-mean clustering method was used to generate the ligand perturbation ensemble and similar/redundant conformations (rmsd to the selected conformation is less than 0.5 Å) were omitted. Finally, up to 100 conformations for each ligand were generated and made available for Rosetta LigandDock.

Rosetta LigandDock with additional near ‘native’ perturbation sampling

We adopted the docking algorithm based on the method previously described in the RosettaLigand docking paper (Meiler and Baker, 2006; Davis and Baker, 2009). In general, the algorithm includes three stages: coarse-grained stage, Monte Carlo minimization (MCM) stage and gradient-based minimization stage. Whereas the original RosettaLigand method performed a full sampling of torsional degrees of freedom in the internal ligands and protein side-chains, we made modifications to enable the fast run time required by the screening method. Specially, we sampled the ligand and protein side-chain torsion angles in near-‘native’ perturbation fashion, where only the near-‘native’ conformation of side-chain and ligand rotamers were allowed and any conformation far away from the starting conformation was omitted. For each protein side-chain, the deviations (±0.33, 0.67, 1 SD) around each input torsion were applied based on the standard deviation value of the same torsion bin from the backbone-dependent Dunbrack rotamer library. For each internal torsional angle of the ligand, the deviations (±5°) around the input torsion were applied as described above.

To optimize possible interactions (H-bonding or packing) between compound and fiber, we carried out random perturbations to the TS rigid-body degrees of freedom (5 Å for translational degrees of freedom; 360° for full rotational degrees of freedom) to explore different rigid body arrangements. For each rigid-body perturbation, different conformations of fiber sidechains, and compounds were explored to maximize the binding interactions. We next carried out simultaneous quasi-Newton optimization of the compound rigid body orientation and the sidechain torsion angles, and in some cases, the torsion angles of the compound and the backbone torsion angles in the binding site, using the complete Rosetta energy function.

Docking of molecules to KLVFFA and Aβ fibrillar structure

The structure of KLVFFA fiber was taken from the co-crystal structure of KLVFFA with orange G (pdb entry: 3OVJ) (Landau et al., 2011). After removing orange G, the sidechain torsion of KLVFFA was optimized to correct any conformational bias from the presence of orange G, and then the optimized structure were inspected to ensure that sidechain torsions are still within the original conformation of the co-crystal structure. The Aβ fibrillar structure was from ssNMR fiber structure of full-length Aβ (pdb entry: 2LMO) 40. The same optimization step was applied before docking. The comparison of docking onto both KLVFFA and Aβ fibrillar structure are discussed in Figure 13.

Post-docking analysis to rank the compounds

The docked compounds were filtered based on the following criteria: (1) The docking models with a compound-fiber van der Waals attractive energy > −7.0 kcal/mol were removed; (2) The docking models with a compound-fiber hydrogen-binding energy >−0.2 kcal/mol were eliminated. The remaining docked compounds were then ranked according to the energy of binding of compound to fiber. We used not only the total binding energy but also on each of the energy components separately (Lennard-Jones interactions, solvation, hydrogen bonding, and electrostatics) (Lazaridis and Karplus, 1999; Kuhlman and Baker, 2000; Kortemme et al., 2003) for ranking. The compounds ranked in the top 40% according to all of these measures were selected. Finally, the compounds were ranked by tightest binding energy (Meiler and Baker, 2006) and best shape complementarity (Lawrence and Colman, 1993).

Description of geometrical parameters of the interactions between BAFs and Aβ fiber defined based on structure-based screening of Aβ toxicity inhibitor

Based on the rounds of computing search and experimental test, general rules of the essential interactions of BAF binding to Aβ fibers are summarized here. As illustrated in Figure 10, the geometrical parameters of those key interactions are specified as followings:

  1. H-bond acceptor (or negative charge) of the inhibitor makes either hydrogen bond or salt bridge to sidechain nitrogen atoms (NZ) of at least two Lysine residues from adjacent Aβ strands along the fiber axis. Our data suggest that the BAFs need to have good contacts across 2 to 4 adjacent Aβ strands, in order to effectively bind to Aβ fiber and reduce Aβ toxicity.

  2. The hydrogen bond or salt bridge described in 1) follows the general rule of H-bond geometry, which are:

    1. Distance (d1, as shown in the figure) between the NZ atom of Lys16 and H-bond acceptor atoms of BAFs: 2.8∼3.5 angstrom;

    2. Angle (Θ1) at BAF H-bond acceptor atoms:100∼150°;

    3. Angle (Θ2) at the NZ atom of Lys16: 130∼180°.

  3. Hydrophobic interactions between the apolar residues (phenylalanine18 and valine 20) and the planar aromatic portion of the compounds. The aromatic portion of compounds should be planar or semi-planar to pack against the flat surface of Aβ which spans across at least two adjacent Aβ strands.

  4. The hydrophobic interactions described in 3) follow the pi-pi stacking geometry, which are:

    1. Distance (d2) between the center of the apolar sidechains and the center of BAF aromatic rings: 4.0∼5.0 angstrom;

    2. Dihedral angle (Φ) between the surface plane defined by Phe18 and Val20 and the aromatic ring of the BAFs: 0∼40°.

Experimental procedures

Chemicals and reagents

Chemicals were obtained from a variety of companies (Table 1) and were of the highest purity available.

Source of KLVFFA(Aβ16–21) and Aβ1–42 peptide

N-terminal acetylated and C-terminal amidated KLVFFA(Aβ16–21) peptide was synthesized by Celtek Bioscience Peptides (Nashville, TN). Aβ1–42 peptide was overexpressed through Escherichia coli recombinant expression system and was purified as reported previously (Finder et al., 2010). The fusion construct for Aβ1–42 expression contains an N-terminal His tags, followed by 19 repeats of Asn-Ala-Asn-Pro, TEV protease site and the human Aβ1–42 sequence. Briefly, the fusion construct was expressed into inclusion bodies in E.coli BL21(DE3) cells. 8 M urea was used to solubilize the inclusion bodies. Fusion proteins were purified through HisTrap HP Columns, followed by Reversed-phase high-performance liquid chromatography (RP-HPLC). After TEV cleavage, Aβ1–42 peptide was purified from the cleavage solution by RP–HPLC followed by lyophilization. To disrupt preformed aggregation, lyophilized Aβ1–42 was resuspended in 100% Hexafluoroisopropanol (HFIP) which was finally removed by evaporation.

Preparation of KLVFFA (Aβ16–21) and Aβ1–42 fiber samples for 1D 1H NMR titration measurement

KLVFFA (Aβ16–21) peptide was dissolved in PBS buffer, pH 7.4 at the concentration of 1 mM and incubated at 37° with continuing shaking for 3 months. Pre-disaggregated Aβ1–42 was dissolved in PBS buffer, pH 7.4 at the concentration of 200 μM and incubated at 37° with continuing shaking for 2 months. For NMR titration samples preparation, KLVFFA (Aβ16–21) or Aβ1–42 fiber stocks were diluted in the PBS buffer solution at the indicated concentrations, followed by adding the small molecules from 100 mM stock solutions in DMSO into fibrillar solution. The final concentration of the small molecule was 50uM or 100 μM. The final volume of NMR samples was 500 μL containing 5% D2O. Prior to NMR spectra collection, samples were incubated at room temperature for 0.5 hr. 500 MHz 1H NMR spectra were collected on a Bruker DRX500 at 283 K with either 256 or 1024 scans collected depending on the intensity of the small molecule signal. H2O resonance was suppressed via excitation sculpting (Hwang and Shaka, 1995); DMSO resonance was suppressed via a frequency shifted presaturation of the DMSO peak. Spectra were processed with XWINNMR 3.6.

Dissociation constant (Kd) of small molecules to fibers calculated from NMR data

NMR data were analyzed to estimate the binding constant for the interaction between the BAF compounds and KLVFFA fibers. We monitored the decrease in the 1H aromatic resonance of the compounds as a function of increasing concentrations of KLVFFA fibers. The general equation for deriving the apparent dissociation constant (Kd) is as follows:

For a general reaction of a ligand binding to fibers (containing N monomers):F(ibril)N+L(igand)FNL.

We estimated the concentration of fibers at any given monomer concentration as:[F(iber)N]=[Fmomomer](1fiber/N monomers),and then we could get:

[FN]=[F ]TN[FNL], [L]=[L]T[FNL], where [F]T is the total monomer concentration, [L]T is the total ligand concentration and [FNL] is the concentration of bound fiber;Kd=[FN][L][FNL]andKd=([F ]TN[FNL])([L]T[FNL])[FNL],and thus[FNL]2([F ]TN+[L]T+Kd)[FNL]+[F ]T[L]TN=0.

Finally, we could get the concentration of bound complex [FNL]:[FNL]=([F ]TN+[L]T+Kd)([F ]TN+[L]T+Kd)24[F ]T[L]TN2(1)

We then applied this equation (1) to our NMR experiments, where we monitored the integrated area of each NMR peak (A) of the compounds over a range of KLVFFA fiber concentrations. Assuming the complex of the BAF compound with fiber is in fast exchange, the peak area is the average of the peak signals for free and bound states, weighted by the fraction of the observed molecule in each state:A=fLAL+fFNLAFNL.

And the change in NMR peak area (ΔA),ΔA=ALA=fFNL(ALAFNL)ΔA(ALAFNL)=ΔAΔAmax=ΔA/ALΔAmax/AL=%ΔA%ΔAmax=fFNL=[FNL][L]TΔA/AL=ΔAmax/AL([F ]TN+[L]T+Kd)([F ]TN+[L]T+Kd)24[F ]T[L]TN2[L]T.

Hence, the observed fraction of peak area change during the titration of increasing fiber concentration against fixed small compound,fobs=fmax([F ]TN+[L]T+Kd)([F ]TN+[L]T+Kd)24[F ]T[L]TN2[L]T.

Our structural model suggests that one BAF compound binds three fiber monomers. To obtain the Kd, we fit the equation for 1:3 (small molecule:fiber) binding to the NMR titration curve (N = 3), with fobs defined as the fraction of peak area decrease (ΔAAL) for each titration experiment, and fmax defined as the fraction maximum of peak area decrease (AmaxAL) for the saturated complex.

MTT cell viability assay

We performed MTT-based cell viability assay to assess the cytotoxicity of Aβ1–42 with or without the addition of BAFs and orange G. A CellTiter 96 aqueous non-radioactive cell proliferation assay kit (MTT) (Promega cat. #G4100, Madison, WI) was used. HeLa and PC-12 (ATCC; cat. # CRL-1721, Manassas, VA) cell lines were used for measuring the toxicity of Aβ1–42. Prior to toxicity test, both HeLa and PC-12 cell lines were plated at 10,000 cells per well in 96-well plates (Costar cat. # 3596, Washington, DC). HeLa cells were cultured in DMEM medium with 10% fetal bovine serum, PC-12 cells were cultured in ATCC-formulated RPMI 1640 medium (ATCC; cat.# 30–2001) with 10% heat-inactivated horse serum and 5% fetal bovine serum. Cells were cultured in 96-well plates for 20 hr at 37°C in 5% CO2. For Aβ1–42 and BAFs samples preparation, purified Aβ1–42 was dissolved in PBS at the final concentration of 5 μM, followed by the addition of BAFs at indicated concentrations. The mixtures were filtered with a 0.2-μm filter and further incubated for 16 hr at 37°C without shaking for fiber formation. To start the MTT assay, 10 μl of pre-incubated mixture was added to each well containing 90 μl medium. After 24 hr incubation at 37°C in 5% CO2, 15 μl Dye solution (Promega cat. #G4102) was added into each well. After incubation for 4 hr at 37°C, 100 μl solubilization Solution/Stop Mix (Promega cat. #G4101) was added to each well. After 12 hr incubation at room temperature, the absorbance was measured at 570 nm with background absorbance recorded at 700 nm. Four replicates were measured for each of the samples. The MTT cell viability assay measured the percentage of survival cell upon the treatment of the mixture of Aβ1–42 and BAFs. The toxicity inhibition (%) or rescuing percentage (%) of each BAF compound was calculated by normalizing the cell survival rate using the PBS buffer-treated cells as 100% and 0.5 μM (final concentration) Aβ1–42 fiber alone-treated cell as 0% viability.

Transmission electron microscopy (TEM)

TEM was performed to visualize the fibrillation of Aβ1–42 in presence of BAFs. The samples of Aβ1–42 and BAFs mixture for TEM measurement were the same as those for MTT assay. For specimen preparation, 5 μl solution was spotted onto freshly glow-discharged carbon-coated electron microscopy grids (Ted Pella, Redding, CA). Grids were rinsed twice with 5 μl distilled water after 3 min incubation, followed by staining with 1% uranyl acetate for 1 min. A CM120 electron microscope at an accelerating voltage of 120 kV was used to examine the specimens. Images were recorded digitally by TIETZ F224HD CCD camera.

ThT fibrillation assay

Purified Aβ1–42 was dissolved in 10 mM NaOH at the concentration of 200 μM, followed by sonication for further solubilizing Aβ1–42. Aβ1–42 was diluted into PBS buffer at the final concentration of 20 μM, and was mixed with 20 μM Thioflavin T (ThT) and different concentrations of BAFs. The reaction mixture was filtered with a 0.2 μm filter, split into four replicates and placed in a 96-well plate (black with flat optic bottom). The ThT fluorescence signal was measured every 5 min using the Varioskan plate reader (Thermo Fisher Scientific, Inc) with excitation and emission wavelengths of 444 and 484 nm, respectively, at 37°C.

References

Acknowledgements

We thank N Wu for help with initial computational work, R Peterson and E Hartman for help with NMR experiments, A Soragni, MR Sawaya, H Chang, D Li, D Anderson, P Bajaj, J Bowie, T Yeates, and F Guo for discussion, and HHMI, NIH, NSF, and DOE for support.

Decision letter

John Kuriyan, Reviewing editor, Howard Hughes Medical Institute, University of California, Berkeley, United States

eLife posts the editorial decision letter and author response on a selection of the published articles (subject to the approval of the authors). An edited version of the letter sent to the authors after peer review is shown, indicating the substantive concerns or comments; minor concerns are not usually shown. Reviewers have the opportunity to discuss the decision before the letter is sent (see review process). Similarly, the author response typically shows only responses to the major concerns raised by the reviewers.

Thank you for sending your work entitled “Structure-based discovery of fiber-binding compounds that reduce the cytotoxicity of amyloid beta” for consideration at eLife. Your article has been favorably peer reviewed by two reviewers, one of whom is a Senior editor.

The Senior editor has assembled the following comments to help you prepare a revised submission in which you should respond to the issues raised by the reviewers.

In this paper, David Eisenberg and colleagues use computational docking to discover small molecules that bind to and stabilize a variety of peptide fibers formed by fragments of amylogenic proteins, and in this way suppress the toxicity of amyloid fibrils of the Alzheimer’s Abeta peptide. They validate the binding of the peptides by NMR, followed up by assays for fibril formation, and also show that the small molecules alleviate the toxicity of the amylogenic peptides to mammalian cells. This work builds on earlier results from the Eisenberg group that established high-resolution crystal structures for many different kinds of peptide fibrils and revealed how small molecules such as Orange G bind to them. The work also follows on earlier studies from others (for example, Pastore and Wanker) on the effects of small molecules on amylogenic peptides. What is new in the present work, however, is the use of computational docking to come up with novel pharmacophores. This is interesting because the binding sites on the fibrils are quite different from the deep hydrophobic cavities that are the targets of conventional docking efforts. Here, the binding sites are flat, and although hydrophobicity is still important, the small molecules lie along the flat surfaces of the fibrils and are anchored by polar residues in a way that stabilizes the fibrils.

The paper is generally clearly written and illustrated. The great interest in developing drugs against amyloid diseases as well as the creative use of computational methods towards that end makes this paper suitable for publication in eLife.

Major comment:

A stronger case would be made if the authors tested whether their small molecule compounds bind to and reduce toxicity of oligomers. This would be a good test of their hypothesis that their compounds target fibrils (not oligomers) and prevent toxicity by reducing the release of toxic oligomers from fibrils (i.e., shifting the equilibrium toward fibrils). Do the authors have data that can address this question, or can such data be obtained relatively easily?

DOI: http://dx.doi.org/10.7554/eLife.00857.039

Author response