Mechanisms of recognition of amyloid-β (Aβ) monomer, oligomer, and fibril by homologous antibodies

Alzheimer's disease is one of the most devastating neurodegenerative diseases without effective therapies. Immunotherapy is a promising approach, but amyloid antibody structural information is limited. Here we simulate the recognition of monomeric, oligomeric, and fibril amyloid-β (Aβ) by three homologous antibodies (solanezumab, crenezumab, and their chimera, CreneFab). Solanezumab only binds the monomer, whereas crenezumab and CreneFab can recognize different oligomerization states; however, the structural basis for this observation is not understood. We successfully identified stable complexes of crenezumab with Aβ pentamer (oligomer model) and 16-mer (fibril model). It is noteworthy that solanezumab targets Aβ residues 16–26 preferentially in the monomeric state; conversely, crenezumab consistently targets residues 13–16 in different oligomeric states. Unlike the buried monomeric peptide in solanezumab's complementarity-determining region, crenezumab binds the oligomer's lateral and edge residues. Surprisingly, crenezumab's complementarity-determining region loops can effectively bind the Aβ fibril lateral surface around the same 13–16 region. The constant domain influences antigen recognition through entropy redistribution. Different constant domain residues in solanezumab/crenezumab/chimera influence the binding of Aβ aggregates. Collectively, we provide molecular insight into the recognition mechanisms facilitating antibody design.

amyloid-␤ (A␤) hypothesis (2) argues that accumulation of A␤ leads to tangles and plaques in AD-affected brains, which cause neuron loss. Several strategies have been developed to reduce the A␤ production, inhibit the A␤ aggregation, or directly enhance the A␤ clearance. Immunotherapy exploiting antibodies and antibody fragments is a major approach (3)(4)(5), and active and passive anti-A␤ immunotherapies can clear brain A␤ deposits. For example, aducanumab (7,8), a human monoclonal antibody, selectively targets aggregated A␤, reduces brain A␤ in a dose-and time-dependent manner, and slows clinical decline (9). Decreasing A␤ concentration in blood could also be an effective therapeutic strategy based on the peripheral sink hypothesis (10). The peripheral sink hypothesis proposed that drug molecules, which bind plasma A␤, can sequester brain A␤ into peripheral tissues without crossing the blood-brain barrier and reduce the accumulation of A␤ in the CNS. Based on this hypothesis, the A␤-binding properties of various biologics (11,12) (e.g. therapeutic antibodies, albumins (13)(14)(15)(16), and transferrins (17)) provide the basis for treating AD. Therefore, antibodies targeting various states of A␤ peptide have been actively studied. However, therapies aimed at reducing protein processing and clearance in AD have been unsuccessful in clinical trials, due to unfavorable pharmacokinetics, difficulty in crossing the blood-brain barrier, and potential neurotoxicity. The mechanisms underlying amyloid-based immunotherapy are complex (18,19) and not fully understood (20). Antibody and other proteins can bind A␤ peptide in different oligomerization states and on different A␤ regions. For example, the three domains of apo-albumin recognized the C-terminal hydrophobic residues (13,15). Despite the many crystal structures of monomeric amyloidogenic peptides in complex with antibodies, there is a lack of structural information on antibody recognition of any aggregated protein.
Among the several therapeutic antibodies for AD, solanezumab (Lilly) and crenezumab (Genentech) are two leading humanized monoclonal antibodies interacting with the midregion of the toxic A␤ aggregates. Solanezumab (21) used IgG1 as template and mainly recognizes monomeric soluble A␤ with picomolar affinity, whereas crenezumab (22) used IgG4 as template and can recognize monomers, oligomers, and fibrils. Solanezumab and crenezumab have the same number of residues in the CDR loops as well as in the fragment antigen-binding (Fab) region, with 11 residues on CDR loops (L1, H1, and H2) and 11 on the constant domain ( Table 1). The constant This work was supported in whole or in part by Federal funds from the NCI, National Institutes of Health, under Contract HHSN261200800001E. This work was also supported in part by the Intramural Research Program of the NCI, National Institutes of Health, Center for Cancer Research. The authors declare that they have no conflicts of interest with the contents of this article. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. This article contains supplemental Figs. S1-S6 and Tables S1-S9. 1  cro ARTICLE domains of the light chain of the two mAbs have identical amino acid sequences, whereas the constant domains of the heavy chain displayed different residues in the CH1-1 loop and in another loop, close to the C terminus. Sequence alignment showed that only Ͻ6% of the Fab residues are different between solanezumab and crenezumab (supplemental Table S1), making them an excellent pair for comparing antibodies with similar sequences but different specificities and raising the question of how these few residues (supplemental Fig. S1) differentially influence A␤ aggregate recognition.
The crystal structure of the complex between solanezumab and A␤12-28 has been resolved (11), revealing that A␤16 -26 forms extensive contacts and hydrogen bonds with solanezumab and that the Phe 19 -Phe 20 hydrophobic core is buried. Recently, Ultsch et al. (12) reported the structure of the complex of A␤11-25 with an engineered crenezumab (CreneFab), which has crenezumab's variable domain and solanezumab's constant domain (23). In the engineered CreneFab complex, like the solanezumab complex, A␤'s Phe 19 /Phe 20 form hydrophobic contacts with crenezumab. A broken essential salt bridge affects the organization of the A␤ oligomer. However, these co-crystal structures still provide limited information to answer why crenezumab recognized more A␤ species. Several recent studies suggested that the constant domain also plays an important role in the antigen recognition (24 -28). There is solid evidence for distant communication between the variable and constant domains (29). Evidence also suggests allosteric effects during antibody-antigen recognition (30).
Here we selected three homologous antibodies and compared theirstructureanddynamicsuponbindingtoA␤peptidesinmonomeric, oligomeric, and fibril forms. Based on crystal structures of solanezumab-A␤ and CreneFab⅐A␤ complexes (supplemental Fig.  S2), we systematically examined the structures of the solanezumab and crenezumab with the soluble, oligomeric, and fibril forms of A␤ using homology modeling, molecular docking, and molecular dynamics simulations. Using this protocol, we screened possible complexes between crenezumab and A␤ aggregates (monomeric, pentamer, and fibril-like structures). We investigated the structure, energy, and dynamics related to the roles of the CDR loops as well as the constant domain in amyloid recognition. Besides the CDR loops, the constant domain loops are also highly correlated with A␤, indicating their role in A␤ recognition. The flexibility of the constant domain CH1-1 and H4 loops (residues 400 -410) changed in response to the A␤ binding. We interpreted the response and correlation as reflecting entropy transfer and release from the antibody-antigen interface to the constant domains.
Collectively, our results provide atomic level structure and dynamics information which may facilitate antibody design in AD.

Solanezumab is conformationally more flexible than crenezumab in the apo form
As can be seen in supplemental Table S1, solanezumab and crenezumab differ by 5, 2, and 4 residues on CDR-L1, H1, and H3 loops, respectively, whereas other CDR loops are identical. Besides CDRs, the light chains have three different N-terminal residues, and the heavy chains have 4 different residues on the CH1 loop and 5 different residues near the C-terminal, distant from the CDR loops. There is no crystal structure of "wildtype" crenezumab. To gain structural insight into A␤ recognition by crenezumab, we compare the crystal structures of the solanezumab⅐A␤ complex with CreneFab (models Smab1, CHmab0, and CHmab1; Table 2) (23). Superimposition of the three crystal structures (supplemental Fig. S2 and Table S3) on the V domain (RMSD Ͻ 0.8 Å) and C domain (RMSD Ͻ 0.65 Å) suggested that they are very similar locally, but the orientations of the C-V domains differ significantly (Ͼ1.3 Å) if the whole Fab structures are used, especially for 5VZX and 5VZY, considering that the V domain and C domain are only 0.382 and 0.575 Å locally but 4.045 Å with full Fab.

Amyloid recognition by homologous antibodies
were combined and analyzed. Apo-crenezumab and apo-Cren-eFab are homogeneous, with 99 and 98% of the population within 4 Å of RMSD, respectively. The structures of apo-solanezumab are more diverse, with three conformational clusters with 84, 12, and 4% of the population (Fig. 1b). Cross-talk between the antibody subdomains helps antibody-antigen recognition (28,(31)(32)(33). The dynamical covariance matrix plots of the apo structures of solanezumab and crenezumab showed that the motions of Fab subdomains are correlated with four major local correlations along the diagonal, representing the four subdomains, VL, CL, VH, and CH1 (Fig. 1c). There are couplings between distant parts of the complex in some systems. The motion correlation analysis suggested that apo-solanezumab has a much larger negative intersubdomain motion, especially between VH and CH-1 and between CL and CH1 (Fig. 1c), which resulted in a more diverse conformational population.
Second, we investigated the local residue fluctuations by using the RMSFs of individual residues (Fig. 1d). The residue RMSFs of the light chain loops (including CDRs) of solanezumab are slightly higher than crenezumab/CreneFab, whereas those of the heavy chain loops (including CDRs) of solanezumab are much larger than crenezumab/CreneFab. We compared the sequence and loop flexibility in their apo forms. There is an interchain disulfide bond between the light and heavy chains of crenezumab (Cmab0) but not in solanezumab/ CreneFab (Smab0 and CHmab0). This interchain disulfide bond, which constrains the loop fluctuation, is more flexible in the CH1-1 in apo-solanezumab/CreneFab than in crenezumab. On the heavy chain around residue 410, the chimera apo structure (CHmab0) showed an even higher peak than solanezumab. This analysis suggested that solanezumab and crenezumab/ CreneFab have different structural and dynamic properties in their apo forms. Without antigen binding, solanezumab is conformationally more flexible than crenezumab.

The recognition mechanism of solanezumab/crenezumab/ chimera and A␤ monomer
Both solanezumab and crenezumab can bind soluble A␤ monomer with similar epitopes, but solanezumab showed stronger binding affinity than crenezumab (34). The crystal structure of the complex of solanezumab and A␤12-28 showed that A␤12-28 adopted a helical structure and formed extensive contacts with solanezumab (34). The crystal structure of the complex between the chimera and A␤11-25 also showed that A␤ formed extensive contacts with the chimera Fab with a similar helical structure (Fig. 2a). Therefore, we first mutated the C domain of the chimera Fab back to crenezumab sequence with V domain and A␤ (model Cmab1 is derived from model CHmab1).
Crenezumabs can recognize diverse A␤ conformation-During the 200-ns simulation of the solanezumab⅐A␤12-28 complex, the structure of A␤12-28, especially the helical structures between Ala 21 and Asn 27 in all (100%) conformers (RMSD Ͻ 5 Å), were well maintained (Fig. 2b). In the crenezumab system, the A␤12-28 structure largely fluctuates and leads to diverse conformations. Cluster analysis suggested that there is one major cluster accounting for 86% of the conformations, whereas the other three account for 10, 2, and 2%, respectively. In the CreneFab system, the A␤12-28 structures became more diverse. Besides the major cluster (76%), the other three account for 18, 3, and 3%, respectively. We measured the secondary structure distribution in the three complexes (Fig. 2c). The central region of the peptide A␤20 -25 mainly (Ͼ95%) adopted a helical structure in solanezumab, whereas in the system of crenezumab and CreneFab, A␤20 -25 adopted more diverse structures, including ϳ70% turn structure, ϳ20% helical structure, and ϳ10% random coil. Therefore, whereas solanezumab mainly recognizes helical A␤12-28, crenezumab and CreneFab can recognize other secondary structure features. As the helical structure is only highly populated in soluble A␤ species, this may explain the specificity of solanezumab recognition of soluble A␤ monomer and the diverse recognition of crenezumab of monomer, oligomers, and fibrils.
Fab structures show similar conformational trends. We clustered the Fab structures from the three complexes (Smab1, Cmab1, and CHmab1) by backbone RMSD of 4 Å. When in complex with A␤12-28, 100% of the population of solanezumab is in one cluster (Fig. 3a). For crenezumab, the conformations are in two clusters with one dominant (98%). For the CreneFab, the conformations become more diverse, with three clusters with populations of 79, 16, and 5%, respectively. In the apo form, solanezumab has higher conformational diversity than crenezumab/CreneFab while being more uniform in complex with A␤12-28.
We calculated the binding energies of the three antibodies binding the A␤12-28 monomer. That of the solanezumab⅐A␤12-28 complexes is Ϫ74.3 Ϯ 3.7 kcal/mol, whereas crenezumab⅐ A␤12-28 complexes are slightly less favorable (Ϫ66.0 Ϯ 3.7 kcal/mol) (Table 3), and CreneFab⅐A␤12-28 complexes are around Ϫ17.8 Ϯ 3.7 kcal/mol. Whereas the absolute energy difference from molecular mechanics is not directly comparable with experimental binding affinities, the relative energy binding preferences from the calculations agree with the experimental trend. Therefore, in terms of the energy landscape, all conformations of solanezumab are optimized for monomer recognition, whereas crenezumab can adapt to a diversified binding environment with some penalty for monomer-binding affinity.
Solanezumab and crenezumab have different epitope preferences-A␤ formed extensive contacts in the complex of solanezumab⅐A␤12-28 (Smab1), with an average contact area of 1889.7 Å 2 (supplemental Table S4). There are more interactions with heavy chain residues than light chain residues. In the crenezumab⅐A␤12-28 complex (Cmab1), the total contact area was 1785.2 Å 2 with VL and VH contributing similarly. The CreneFab⅐A␤12-28 interactions (CHmab1) are similar to the crenezumab⅐A␤12-28 complex. Essentially, the numbers of hydrogen bonds, hydrophobic and electrostatic interactions revealed that all three antibodies have similar contacts with the A␤12-28 monomer.
However, the contact frequency between residues from A␤ and Fabs during the simulation revealed that solanezumab and crenezumab have different epitope preferences (Fig. 2d). In the complex of solanezumab⅐A␤12-28 (Smab1), almost all of the residues from A␤12-28 made contacts, and Phe 19 , Lys 16 , and Phe 20 ranked as the top 3. The CDR-H1 and -H3 loops of solan-

Amyloid recognition by homologous antibodies
ezumab contributed most of the contacts. In the complex of crenezumab⅐A␤12-28 (Cmab1), A␤12-28 made considerable contacts; however, the preferred interacting residues shifted to the N terminus. Lys 16 , Phe 20 , and Leu 17 ranked as the top 3. Lys 16 becomes the most preferred residue, whereas the contacts from Phe 19/20 were reduced about 50%. In the complex of chimera⅐A␤12-28 (CHmab1), the preferred interacting residues further shifted to the N terminus, with Lys 16 becoming the most preferred residue and Phe 19/20 reduced to about 80%.
To further identify the essential residues of Fabs in the recognition, antibody-antigen contacts within 3 Å of each other were monitored during the simulation and normalized by the total contact time. A contact value of 1.0 means that a certain Fab residue always contacts an A␤ residue during the simulation, and a value Ͼ1.0 indicates that the residue contacts at least two A␤ residues.
The underlying driving force in epitope differentiation is embedded in the slight difference in the CDR region. There are more hydrophobic residues in contact with A␤ in solanezumab than in crenezumab/CreneFab. We identified the essential residues for the recognition by monitoring the change of accumulated antigen contacts at specific sites, especially where the residues of the two mAbs differ (supplemental Table S5). Four sites were identified with large contact change (⌬contact Ͼ 0.25) with mutations between the two mAbs. Arg 250 (31H), Figure 2. The V domains of crenezumab and CreneFab can recognize a more diverse A␤ ensemble than solanezumab. a, the two A␤-bound crystal structures superimposed on the A␤ and V domains. 4XXD and 4KNA are colored red and lime, respectively. b, clustered conformations of A␤12-28 in complex with solanezumab, crenezumab, and CreneFab (RSMD for clustering is 4 Å). c, secondary structure components of A␤12-28. Helical, ␤-strand, turn, and random coil structures are colored black, red, blue, and green, respectively. d, contact preference on the Fabs-A␤ interface from the A␤ and Fab sides. The Fab amino acid preference was obtained by summation of the contacts based on the Fab side contact preference.
The constant domains contribute to A␤-Fab recognition by entropy dissipation-The binding energy difference of crenezumab⅐A␤12-28 complex (Cmab1) and CreneFab⅐ A␤12-28 complex (CHmab1) suggested that even with the same sequence and initial conformation of the V domain, the . Solanezumab is less conformationally diverse than the other two Fabs when bound to A␤12-28. a, cluster analysis of the structures. Backbone RMSD ϭ 4 Å is used to define the cluster. Clusters are colored blue, red, and gray, respectively. b, motion correlation among the residues of the three Fabs. Residues with highly correlated and anti-correlated motion are red and blue, respectively. c and d, RMSFs (c) and order parameters S 2 (d) of the three Fabs in complex with A␤12-28. The location of CDRs and important constant domain loops are boxed. The values of solanezumab, crenezumab, and CreneFab are colored black, blue, and red, respectively. To identify important distant residues in A␤-Fab recognition, we compared the RMSF change of three Fabs upon binding to A␤ (Fig. 3c). For the V domains of those Fabs, the RMSFs are very similar, with only slight variations, whereas for the C domains, the flexibility varied upon binding A␤12-28. The L4 loop of the CL domain (around residue 205) of CreneFab showed significant flexibility (Fig. 3c). The distant CH1-1 loop region of solanezumab is highly flexible compared with crenezumab and CreneFab, with the chimera showing larger RMSF than solanezumab and crenezumab in the remainder of the CH1 domain. The conformational changes of solanezumab and A␤12-28 upon recognition indicate a larger entropy cost for stabilizing both solanezumab and A␤12-28, suggesting an entropy dissipation mechanism. The larger fluctuations of the distant loop may help in transferring vibrational entropy. Thus, the mutations in these distant loops appear important in antigen recognition.
There is a good correlation between backbone entropy and the order parameter obtained from MD simulations (35). To check all residues, including proline, we calculated the generalized order parameter S 2 of the CϭO bond of each individual residue of the antibodies (Fig. 3d). We found that on the heavy chain, which mainly interacts with A␤, solanezumab and crenezumab showed higher S 2 on the CDRs, whereas they showed lower S 2 than CreneFab on the constant domain CH1-1 loop. This suggests a transfer of entropy from CDRs to CH1-1 when A␤ and solanezumab/crenezumab form a stable complex. The diversified and longer communication pathways between the antigen epitopes and distant antibody residues may also reflect constant domain contributions. We examined the signals from the A␤ monomer to the constant domain of the three antibodies, especially the regions with strong response and bearing mutation sites. In most cases, there are longer optimal pathways from the Phe 20 /Lys 16 of A␤ to C domain residues in solanezumab than crenezumab (Table 4 and supplemental Table S8). There are no communications between A␤ and C domain residues in CreneFab. Therefore, the solanezumab constant domain has less effect on antigen recognition, whereas in crenezumab, the constant domain has a larger effect.

The recognition mechanism of crenezumab and A␤ oligomer
Identification of stable structures of antibody⅐A␤ oligomer complexes-Based on the populations of the crenezumab Fabs, ϳ90 and ϳ10% of the conformations fall into two clusters, and the V domains are very similar in all. This ensemble led us to select the crenezumab Fab based on the crystal structure of the chimera for molecular docking with A␤12-28.
The polymorphic nature of the A␤ oligomers and fibrils makes it difficult to obtain their crystal structures in complex

Amyloid recognition by homologous antibodies
with antibodies. A␤1-42/1-40 fibril structures have been resolved by ssNMR (36 -39) and by computations (40 -51). In this work, a pentamer and 16-mer of A␤11-42 obtained from ssNMR (PDB code 2MXU) (51) were selected to represent the oligomers and fibrils, respectively. In the isolated states, without antibody interactions, the overall structure and the ␤ structure of A␤11-42 5-mer and 16-mer were well maintained during the 200-ns simulation, whereas the A␤12-28 lost its initial helical structure (Fig. 4). Therefore, we used the A␤11-42 5and 16-mer for searching a potential antibody recognition pattern by antibody docking and subsequent MD simulations of the docked complexes. Initial docking between A␤ pentamer and crenezumab pointed to two possible patches, His 13 -His 14 -Gln 15 and Ala 21 -Glu 22 -Asp 23 , of A␤ as antibody-binding sites. We selected five models (three with the 21-23 and two with the 13-15 at the interface; supplemental Fig. S3) for further refinement. The A␤ pentamer dissociated from crenezumab within 30 -70 ns for models with the 21-23 patch as epitope, whereas complexes with the 13-15 epitope were stable through the 200-ns simulation (supplemental Fig. S4). In the stable complexes, crenezumab mainly recognizes A␤ N-terminal residues 11-22 with some contact with C-terminal residue Ile 32 as well. The RMSDs of the two stable complexes with the 13-15 epitope were ϳ5 Å throughout the 200-ns simulation, whereas the three complexes with the 21-23 epitope quickly reached Ͼ8 Å in ϳ50 ns (supplemental Fig. S4a). The binding energy (supplemental Table S6) also indicated that crenezumab prefers to bind the
The recognition mechanism of crenezumab and A␤ oligomer-Using the most stable crenezumab-A␤ oligomer model Cmab5 (Ϫ39.8 kcal/mol) as representative, we also similarly examined a possible solanezumab-A␤ oligomer (Smab5) and a chimera-A␤ oligomer (CHmab5) recognition. After 200 ns, the A␤ pentamer interaction with solanezumab rotated from the 13-16 patch to the 19 FFAE 22 and 32 IGL 34 regions, with a small binding energy (Ϫ3.9 Ϯ 4.5 kcal/mol). Our simulations confirmed that solanezumab prefers to bind monomeric A␤, whereas crenezumab prefers oligomeric A␤, agreeing with the experimental results. Crenezumab can maintain a consistent binding epitope around A␤13-15, although with different conformation, whereas solanezumab is only able to stay around A␤21-23. Two binding modes have similar contact surface areas; the stable crenezumab⅐A␤11-42 oligomer complex has a contact area of 1505.5 Å 2 . VL (46.7%) and VH (53.3%) bind to the A␤ pentamer with a similar contribution. The corresponding solanezumab⅐A␤11-42 5-mer complex has a contact area of 1456.8 Å 2 . A␤ pentamer preferred to bind VH (61.4%) rather than VL (38.6%).
What are the factors responsible for the binding preference switch between monomer/oligomer for crenezumab and solanezumab? Fig. 5a highlights residues with cumulative contacts Ͼ 1, and Fig. 5b shows the residue components on the interface. Ser, Tyr, Thr, and Asn are the top ranking residues on the crenezumab-A␤ interfaces, whereas solanezumab uses Arg, Asn, and Val for the A␤ binding. Arg contributes to the binding in most of the solanezumab⅐A␤ pentamer complexes, whereas in the crenezumab⅐A␤ pentamer complex, there is only a small contribution from Arg. Hydrophobic residues and Leu contribute to the contacts in the solanezumab⅐A␤ pentamer complex, whereas in crenezumab⅐A␤ pentamer complexes, most of the preferred contacts are from hydrophilic and electrostatic interactions. Overall, crenezumab CDR-L1 and H1 contribute to the binding more than other CDRs, with very minor hydrophobic contributions (0.6 hydrophobic interactions, 13.8 hydrogen bonds, and 2.1 electrostatic interactions on the interface). The hydrogen bonds were reduced to 9.2, whereas hydrophobic interactions increased to 8.2.
Several CDR residues, which differed in solanezumab and crenezumab, dramatically changed the contact behavior ( Fig.  5b and supplemental Table S7). In the crenezumab⅐A␤ pentamer complexes, Ser 250 (31H) interacts with the A␤ cationic region of 13 HHQKL 17 , whereas in the solanezumab⅐A␤ pentamer complex, this residue was mutated to cationic Arg, causing electrostatic repulsion on the interface and a shift to 19 FFAE 22 . Clearly, The Arg 250 (31H) in solanezumab prevents interaction with the highly positive charged 13 HHQKL 17 patch and leads solanezumab to the negatively charged A␤ 19 FFAE 22 patch. Other residue differences enhanced the hydrophobic interactions or hydrogen bonds at the solanezumab interface. For example, in the crenezumab⅐A␤ complexes, Asn 273 (53H) and Gly 275 (55H) made no contacts with the antigen; however, in the solanezumab⅐A␤ complex, the corresponding Val 273 (53H) and Asn 275 (55H) interact with A␤ 32 IGL 34 . Although these two mutations increase the hydrophobic interactions and hydrogen bonds, they potentially weaken the electrostatic interactions with A␤ 13 HHQKL 17 . The change of cumulative contacts at specific sites between solanezumab and crenezumab and the A␤ oligomer is shown in supplemental Table S7. Many residues have large contact change (⌬contact Ͼ 0.25). For example, Arg 250 (31H) and Asp 33 (33H) in solanezumab were mutated to Ser and Asn in crenezumab, leading to 1.82 and 1.75, respectively. These mutations also resulted in neighboring residue contact change (e.g. Thr 247 , Tyr 251 , Tyr 31 , and Gly 34 ).
Constant domain contribution-We simulated possible structures of the CreneFab⅐A␤ pentamer complex, starting with a conformation identical to that of the crenezumab⅐A␤ pentamer complex. It was unchanged during the 100-ns simulation. The CreneFab⅐A␤11-42 5-mer complex has a contact area of 1724.5 Å 2 . VH binding to the A␤ pentamer is preferred (60.7%) over that of VL (39.3%). Hydrogen bonds were reduced to 9.4, whereas hydrophobic interactions increased to 7.6 ( Table 4). Even with identical variable domains, the binding energy of CreneFab⅐A␤11-42 oligomer complex was reduced significantly compared with the wild-type crenezumab (Ϫ6.4 Ϯ 4.5 kcal/mol; Table 3), suggesting that constant domain mutations can influence the A␤ recognition.
We clustered the Fab structures from the three complexes (Smab5, Cmab5, and CHmab5) by backbone RMSD of 4 Å. As shown in Fig. 6a, both solanezumab and CreneFab (which share the constant domains) become relatively rigid. The conformations of solanezumab form five clusters with one dominant (86%). For CreneFab, the conformations are categorized into four clusters with one dominant (94%). The conformations become more diverse for crenezumab, which has six clusters with one dominant (56%). Fig. 6c compares the RMSFs of the complexes of the three Fabs. Residues 1-110 (especially CDR-L1) in crenezumab light chain are stabilized, whereas the 120 -178 fragment is more flexible compared with solanezumab. All RMSFs of the heavy chain CH1 domain are higher for crenezumab than solanezumab. We observed that crenezumab and CreneFab have lower S 2 in the constant domain of light chain (Fig. 6d). These features indicated that crenezumab has tight variable domain binding and higher constant domain motion for entropy dissipation, as in the case of monomer binding.
Is the motion correlation pattern found in monomer recognition conserved in oligomer binding? Compared with the solanezumab/crenezumab⅐A␤ monomer complexes, the Fab subdomains in solanezumab/crenezumab⅐A␤ oligomer complexes showed much stronger positive or negative correlation (Fig. 6b). In the crenezumab⅐A␤ complexes (Cmab5), there are strong negative motion correlations among VL, CL, and VH domains, whereas these correlations are reduced in the solanezumab⅐A␤ and CreneFab⅐A␤ complexes. As crenezumab has the most favorable binding energy with the A␤ oligomer, it seems that the subdomain fluctuations or flexibility is necessary for Fab to recognize larger A␤ aggregates. Moreover, in the complex of A␤ pentamer with crenezumab, there are strong positive correlations between crenezumab CDRs and the contacting residues from A␤, whereas those A␤ residues that are far from the interface showed negative correlation with CDRs. Figure 5. The epitope of A␤ oligomer shifted to the N-terminal hydrophilic and cationic residues when in complex with crenezumab and CreneFab compared with solanezumab. a, molecular details of the A␤ oligomer-crenezumab complex. Residues with cumulative contacts Ͼ1.0 are represented by sticks. Crenezumab residues Tyr 251 H and Asp 319 H (alanine scanning using SPR (23)) are highlighted by beads, whereas other important residues, Ser 250 H and Asn 33 L, which differ between crenezumab and solanezumab, are also underlined. Residues from A␤, light chain, and heavy chain are indicated by A␤, L, and H, respectively. Light chain, heavy chain, and A␤ oligomer are colored pink, lime, and ice blue, respectively. Hydrophobic, hydrophilic, cationic, and anionic residues are colored white, green, blue, and red, respectively. b, contact preference on the Fabs-A␤ interface from A␤ side and Fab side. Fab amino acids preference was obtained by summation of the contacts based on the Fab side contact preference.

Amyloid recognition by homologous antibodies
However, in the complex of A␤ pentamer and CreneFab, both the positive and negative motion correlations were greatly reduced. We further examine the signals from the A␤ oligomer to the constant domain, especially the regions with strong response and mutations, in the three antibodies (supplemental Table S7). In most cases, there are shorter optimal pathways from A␤ Lys 16 to C domain residues in crenezumab than solanezumab or CreneFab.

Amyloid recognition by homologous antibodies Which antibody structural features recognize the A␤ fibril?
We screened the binding modes of the A␤ 16-mer/crenezumab complex with the epitope of the A␤13-15 region based on the simulations of crenezumab⅐A␤ pentamer. Considering the linear morphology of the A␤ fibril, three different binding orientations were tested (supplemental Fig. S5). Our simulations indicated that in models a and b, the A␤ 16-mer dissociated from crenezumab within 160 and ϳ20 ns, respectively, whereas in model c, A␤ 16-mer and crenezumab form a stable complex throughout the 200-ns simulation. This indicated that the A␤13-15 region disfavored the light chain of crenezumab. The net charge of the crenezumab light and heavy chain is ϩ2 and 0, respectively, whereas the net charge of A␤13-15 is ϩ1 (His 13 and His 14 might also be protonated); thus, repulsion might prevent the binding between the light chain and A␤13-15. Through extensive screening of possible complexes of A␤ fibril and crenezumab, we obtained a stable complex structure of A␤ 16-mer⅐crenezumab (supplemental Fig. S5), with a binding energy of Ϫ195.8 Ϯ 6.3 kcal/mol.
The recognition mechanism of crenezumab and A␤ fibril-The stable crenezumab⅐A␤11-42 16-mer complex has a contact area of 1667 Å 2 , and VL and VH contributed 34.1 and 65.9% contact area to the interface, mostly from CDR-H1 and H2. In this model, Crenezumab can bind the A␤ fibril on the lateral surface around residues 10 -16, which constitute also the epitopes in oligomer and monomer binding. The interactions are mostly salt bridge and hydrogen bonds (Table 4 and Fig. 7a This most stable conformation was used to test whether solanezumab Fab can bind A␤ fibrils. After 200-ns simulations, solanezumab is indeed able to maintain interaction with the A␤ fibril. Hydrophobic residues (e.g. Val) contribute to the contacts in solanezumab⅐A␤ monomer complex, whereas in crenezumab⅐A␤ oligomer complexes, most of the preferred contacts were from hydrophilic interactions (Fig. 7b). Although the contact area between solanezumab⅐A␤11-42 16-mer (1818.7 Å 2 ) is larger than between crenezumab⅐A␤ 16-mer, the binding energy is only Ϫ41.2 Ϯ 6.3 kcal/mol, much less than the corresponding binding energy of Ϫ195.8 Ϯ 6.3 kcal/mol of crenezumab ( Table 3). The binding energy indicated that crenezumab can bind A␤ fibril better, agreeing with experimental results. Three mutations D33N(28L), V273N(53H), and N35D(30L) from solanezumab to crenezumab are potentially important in differentiating between the recognition of the A␤ fibril by the two Fabs (Table 6). Clearly, these mutations enhanced antibody-A␤ fibril interactions for crenezumab.
The expected increasing flexibility of the H1 and H2 loops due to more Gly residues in crenezumab is reflected in the A␤ fibril recognition. As shown in Fig. 8c, the RMSFs of CDR H1 and H2 loops as well as L2, L3, and L4 are larger in crenezumab than in solanezumab. The higher RMSFs in most loops may imply that higher flexibility helps in binding the amyloid fibril.

Constant domain contribution in amyloid fibril recognition-
We simulated possible structures of CreneFab⅐A␤ 16-mer complex, with a starting conformation identical to that of the crenezumab⅐A␤ 16-mer complex. The complex is stable throughout the 100-ns simulation, with a much smaller contact area (887.1 Å 2 ) and fewer hydrogen bonds (Table 4). VH binding to the A␤ 16-mer is dominantly preferred to VL's (94% versus 6%). The binding energy is lower than crenezumab but higher than solanezumab (Ϫ85.2 Ϯ 6.0 kcal/mol; Table 3), indicating an interplay between the variable and constant domains in antigen recognition. When forming a complex with A␤ 16-mer, the population of solanezumab conformations is stably distributed, and all are in one cluster (Fig. 8a). For crenezumab, the conformations are in four clusters with one dominant (80%). For the CreneFab, the conformations fall into three clusters with populations of 81, 11, and 8%, respectively.
The motion correlations between Fabs and A␤ aggregates were also studied (Fig. 8b). The interdomain motion correlation  (23)) are further highlighted by beads, whereas Asn 273 H, Asp 35 L, and Asn 33 L, which differ between crenezumab and solanezumab, are also underlined. Residues from A␤, light chain, and heavy chain are indicated by A␤, L, and H, respectively. Light chain, heavy chain, and A␤ oligomer are colored pink, lime, and ice blue, respectively. Hydrophobic, hydrophilic, cationic, and anionic residues are colored white, green, blue, and red, respectively. b, contact preference on the Fabs-A␤ interface from the A␤ side and Fab side. Fab amino acid preference was obtained by summation of the contacts based on the Fab side contact preference.

Amyloid recognition by homologous antibodies
is enhanced in the A␤ 5-mer complex and is stronger in crenezumab than in solanezumab and CreneFab in complex with A␤ 16-mer. There is a stronger A␤-Fab correlation in crenezumab than in solanezumab/CreneFab. The order parameter S 2 showed that when binding to A␤ 16-mer, crenezumab showed higher S 2 on CDRs and a lower one on constant domain loops, especially on the light chain. As can be seen in Fig. 9b, crenezumab has a regular shift of increasing S 2 in CDR loops and decreasing S 2 in light chain constant domain from the apo form to complexes with A␤ monomer, A␤ 5-mer, and A␤ 16-mer.
We finally examined the signals from the A␤ fibril to the constant domain, especially the regions with strong response and mutations (Table 4). In most cases, like the A␤ oligomer, there are shorter optimal pathways from A␤ Lys 16 to C domain residues in crenezumab than in solanezumab and CreneFab (Table 4).
To elucidate the recognition mechanism of different oligomerization states in atomic detail, here we search for potential complexes between A␤ oligomers/fibrils and solanezumab/ crenezumab and systematically evaluate their stabilities. We found that crenezumab consistently recognizes exposed A␤11-16 epitopes in different oligomerization states. The sequence similarity between solanezumab and crenezumab suggested that both exploit similar epitopes in the recognition of A␤ monomers and oligomers (11,23). The new structure of the CreneFab⅐A␤ monomer complex indicated that crenezumab binds to A␤13-24, whereas solanezumab prefers A␤16 -26 (23). Consistently, our docking, modeling, and energy evaluation indicated that crenezumab recognized A␤11-16, especially in complex with A␤ oligomers/fibrils rather than 16 -26 as in solanezumab. Experimentally, crenezumab binds with aggregated A␤ species with ϳ10 times lower K D than with the monomer (23). The secondary structure and exposed residues in aggregated states differ from the mono- Among the available ssNMR structures of A␤ fibrils (36 -39), A␤ residues (Phe 19 , Phe 20 , Ser 26 , and Asn 27 ) were almost buried inside the hydrophobic core (PDB code 2NAO) or in the turn region (PDB codes 2M4J and 2LMP) and thus cannot be recognized by antibodies. The A␤ 13 HHQK 16 region is usually exposed to bulk solution with higher flexibility compared with A␤ 19 FFAEDVGSN 27 . Thus, the nature of the A␤ fibrils also determines the recognizable regions. Moreover, cationic residues on the antigen (e.g. Lys and His) are easier to recognize than anionic residues on the antibody-antigen interface. 4 The preference of cationic residues further favored A␤ 13 HHQK 16 .

Amyloid recognition by homologous antibodies
There are other antibodies and biologics specifically recognizing C-terminal residues. For example, anti-A␤42 oligomeric VIA antibody was raised against VIAVIA peptide (61), which corresponds to the A␤40 -42 residues.
A␤ oligomer and fibril structures are highly polymorphic (46), and antibodies should have the ability to adjust their conformation to recognize them. Our analysis of A␤12-28 in complex with three different Fabs indicated that whereas crenezumab and CreneFab can bind heterogeneous A␤12-28 conformations, this is not the case for solanezumab. This is in line with our observation that solanezumab has stronger binding affinity to monomeric A␤12-28 in helical conformation and fails to adapt to other conformations in A␤ oligomers and fibrils.
Essential variable domain residues shift crenezumab's epitope preference. Alanine-scanning mutagenesis of crenezumab suggested that Y251(32H)A, G318(95H)A, and D319(101H)A abolish binding (23). The three residues are the same in solanezumab and crenezumab. This suggests that the three residues are important for monomer binding, but not other A␤ species. However, several mutations (R250S(H) and S252G(H)) shift A␤ recognition from the middle to the N-terminal region. As to A␤ oligomer/ fibril⅐crenezumab complexes, Asp 35 (L) and Asn 273 (H) make considerable contacts with A␤ 13 HHQK 16 , whereas the corresponding residue Asn and Val in solanezumab undermines the interaction with A␤. These residues were anionic (Asp) or hydrogen bond-forming (Asn, Ser, and Tyr) in crenezumab recognizing A␤ 13 HHQK 16 , whereas they were cationic (Arg) or hydrophobic (Val and Phe) in solanezumab recognizing A␤ 19 FFAEDVGSN 27 . These residues, which differ in solanezumab and crenezumab, directly interact with the A␤ aggregates.
The constant domain may affect antigen recognition through an entropy redistribution mechanism. The recognition is associated with structural transitions of the inherently flexible antibody (32,62,63). The variable domains, especially CDRs, mainly modulate the specificity and affinity (64), whereas the constant domains control the isotype/effector (65). However, recent studies indicated that besides the variable domains, the constant domain is also implicated in antigen binding (24 -28). Communications between the variable domains of the light and heavy chains (66) as well as between the variable and constant domains (29) were unraveled, consistent with allostery-elicited conformational changes in antibodies (67). In our previous work involving a short prion peptide and an antibody with IgG2 template (31), we found that oxidation of the intermolecular disulfide bond dramatically undermines the binding affinity. In this work, we observed that CreneFab, which combines the crenezumab V domain and solanezumab C domain, showed decreased binding affinity to all A␤ species compared with wild-type crenezumab, demonstrating that the C domain influences antigen recognition. In crenezumab, there is an interchain disulfide bond between the CH1-1 loop and C terminus of the light chain but not in solanezumab and CreneFab. This disulfide bond allosterically influences prion recognition (31), suggesting its essential role in antigen conformer recognition. Two H4 loop residues, which differ between solanezumab and crenezumab (Asn-417 3 Asp(H) and Gln-410 3 Lys(H)), are negatively correlated (reverse synchronous motion) with the A␤ antigen and may act in the recognition.
Although there is no uniform structural change upon binding to A␤ species by solanezumab or crenezumab, there are significant light chain 160 -180 and 210 -220 and heavy chain 415-425 changes. Considering the RMSF, overall conformational diversities, and order parameter change, it is evident that upon the binding of A␤, the Fab entropy is redistributed. The high entropy in the flexible CDR loops and peptide needs to be transferred to the solvent or another antibody region to avoid entropy penalty upon ligand binding. In Fabs or in the full antibody, the entropy might be transferred to the loops in the constant domain or further to the Fc region facilitating receptor binding.
The correlated variations of mutations, conformational diversities, and communication pathways suggest that binding to larger A␤ oligomers requires balance of paratope-epitope interaction, allosteric loop response, and antibody conformational dynamics.

Conclusions
Complementary to the known antibody⅐A␤ monomer crystal structures, for the first time we identified interaction patterns of therapeutic antibodies in recognizing A␤ oligomers/fibrils. In A␤ oligomers/fibrils, the antibodies prefer epitope residues 13-16. Crenezumab anionic and hydrogen bond-forming residues in the CDR loop are responsible for the interaction. For the larger and more organized A␤ aggregates, a conformational reorganization takes place among Fab subdomains. In solanezumab and crenezumab, constant domain loops' residues respond differently, indicating a role in A␤ aggregate recognition. Entropy transfer to these constant domain loops upon antigen binding might trigger a change in flexibility. Understanding how an antibody can recognize simultaneously a peptide monomer and oligomer has been puzzling. Unlike a previously suggested mechanism that monomers are sequestered from oligomers to enable oligomer recognition, our results suggest that the flexibility of the antibody CDR region is the key in recognition of the same epitope region with different antibody conformations (i.e. conformational selection (18, 68 -71) is a decisive factor).

System preparation
The structures of the bound and apo form of the CreneFab were directly obtained from the crystal structures of 5VZX and 5VZY (23). For solanezumab, as there are two complexes packed in the asymmetric unit of the crystal (PDB code 4XXD) of solanezumab⅐A␤ complex with slight structure variation (11), we considered both complexes. The unbound (apo) form of solanezumab was obtained by manually removing the A␤ peptide in the bound (PDB code 4XXD) structure.
Although there is no structural information available for crenezumab, as it showed high sequence similarity (Ͼ97%) ( Table 2) compared with CreneFab, the structures of the bound and apo form of the crenezumab were modeled by templatebased homology modeling using the SWISS-MODEL server with the CreneFab structures as templates. The missing residues were modeled by template-based homology modeling using the SWISS-MODEL server (72).

Amyloid recognition by homologous antibodies
Solid-state NMR structures (PDB code 2MXU) (51) were used as two representative fibril (16-mer)/oligomeric (5-mer) structures of A␤. To search the potential interfaces between A␤ oligomers/fibrils and the three Fabs, we performed automated molecular docking of A␤ oligomers/fibrils and Fabs with the program HADDOCK version 2.1 web server (73). HADDOCK can use various kinds of ambiguous interaction restraints to guide the docking process. In this work, the epitope of A␤ oligomers is 12-28, whereas the paratopes of solanezumab/crenezumab were set as CDR loops. In the HADDOCK docking, 12-28 of A␤ and CDR loops of Fabs were input as active residues; A␤ oligomers/fibrils were set rigid, whereas the residues of CDR loops of Fabs, which were in the loop conformation according to the crystal structures (i.e. position 26 -37, 55-58, 95-102, 245-253, 270 -277, and 318 -320) were set flexible.
For the crenezumab⅐A␤ oligomer/fibril complex, the analysis of the final 200 HADDOCK models of complexes resulted in five and three clusters for crenezumab⅐A␤-5-mer and crenezumab⅐A␤-16-mer, respectively. To further refine these complexes between Fabs and A␤ in different states, the initial poses from HADDOCK were locally perturbed by Rosetta docking (74 -77). The poses with larger cluster size, lower Z-score and total_score/I_sc were considered promising candidates. The complexes between A␤ oligomers/fibrils and solanezumab or CreneFab were obtained by mutations of crenezumab back to the solanezumab or CreneFab sequences using the crenezumab⅐A␤ oligomer/fibril complex.

MD simulation protocols
The conserved disulfide bonds were constructed according to the specific IgG subtypes. As the non-sequential Kabat numbering scheme is used in the crystal structures, we renumber the residues for convenience in the simulation (see supplemental Table S2). For the light chain, the heavy chain, and the peptide, the N termini and C termini were charged as NH 3 ϩ and COO Ϫ groups, respectively. The crystal water molecules in the crystal structures were kept. The systems were then solvated by TIP3P water molecules, and sodium and chlorides were added to neutralize the system and to achieve a total concentration of ϳ150 mM. The resulting solvated systems were energy-minimized for 5000 conjugate gradient steps, where the protein was fixed and water molecules and counterions could move, followed by an additional 5000 conjugate gradient steps, where all atoms could move. In the equilibration stage, each system was gradually relaxed by performing a series of dynamic cycles, in which the harmonic restraints on proteins were gradually removed to optimize the protein-water interactions. In the production stage, all simulations were performed using the NPT ensemble at 310 K. All MD simulations were performed using the NAMD software (78) with CHARMM36 force field (79). MD trajectories were saved by every 2 ps for analysis. A summary of all simulation systems is given in supplemental Table S6 and supplemental Table S9.

MD simulation analysis
To identify the essential interactions between amyloid and Fabs, all atoms within 3 Å between amyloid and the Fabs during the last 100-ns simulation were considered as input into PROTMAP2D (80), which can calculate the accumulated contact map by summing up all of the frames during simulations.
To evaluate the binding energy between Fab and the prion peptide, the trajectory for each bound and apo system was extracted from the last 20 ns of explicit solvent MD without water molecules and ions. The solvation energies of all systems were calculated using the generalized Born method with molecular volume (GBMV) (81) after 500 steps of energy minimization to relax the local geometries caused by the thermal fluctuations that occurred in the MD simulations. In the GBMV calculation, the dielectric constant of water is set to 80, and no distance cutoff is used. The binding energy between the two Fabs and the A␤ species was calculated by the equation, ͗E bind ͘ ϭ ͗E complex ͘ Ϫ ͗E Fab ͘ Ϫ ͗E A␤ ͘. Binding energy is summarized in supplemental Fig. S6.
Correlations between all the residues in the 12 systems were analyzed for the entire 100-ns MD trajectory (25,000 frames) using the normalized covariance to characterize the correlation in motion of protein residues (82)(83)(84)(85), ranging from Ϫ1 to 1. If two residues move in the same (opposite) direction in most the frames, the motion is considered as (anti-)correlated, and the correlation value is close to Ϫ1 or 1. If the correlation value between two residues is close to zero, they are generally uncorrelated. The correlation evaluation was performed by using CARMA (52). The weighted network, optimal/suboptimal paths in Fab/peptide systems were analyzed using the Net-workView (6) module in VMD.
Generalized order parameter S 2 of the carbonyl group of each individual antibody residue in the 12 systems was analyzed for the last 20-ns MD trajectory (5000 frames) using the CHARMM NMR analysis module, ranging from 0 to 1. Higher value indicated a more ordered structure.
Author contributions-J. Z. performed experiments and wrote the paper. R. N. wrote the paper. B. M. conceived and coordinated the study and wrote the paper. All authors reviewed the results and approved the final version of the manuscript.