Conformational dynamics of cathepsin D and binding to a small-molecule BACE1 inhibitor

Abstract

BACE1 is a major therapeutic target for prevention and treatment of Alzheimer’s disease. Developing inhibitors that can selectively target BACE1 in favor of other proteases, especially cathepsin D (CatD), has presented significant challenges. Here we investigate the conformational dynamics and protonation states of BACE1 and CatD using continuous constant pH molecular dynamics with pH replica-exchange sampling protocol. Despite similar structure, BACE1 and CatD exhibit markedly different active site dynamics. BACE1 displays pH-dependent flap dynamics that controls substrate accessibility, while the CatD flap is relatively rigid and remains open in the pH range 2.5–6. Interestingly, although each protease hydrolyzes peptide bonds, the protonation states of the catalytic dyads are different within the active pH range. The acidic and basic components of the BACE1 catalytic dyad are clear, while either aspartic acid of the CatD catalytic dyad could play the role of acid or base. Finally, we investigate binding of the inhibitor LY2811376 developed by Eli Lilly to BACE1 and CatD. Surprisingly, in the enzyme active pH range, LY2811376 forms a stronger salt bridge with the catalytic dyad in CatD than in BACE1, which might explain the retinal toxicity of the inhibitor related to off-target inhibition of CatD. This work highlights the complexity and challenge in structure-based drug design where receptor-ligand binding induces protonation state change in both the protein and the inhibitor.

Graphical abstract

BACE1 is a major therapeutic target for Alzheimer's disease. Developing selective inhibitors as presented significant challenges.

Here we use continuous constant pH molecular dynamics simulations to investigate the conformational dynamics of cathepsin D (CatD) and its binding to a BACE1 inhibitor as a function of pH.

Surprisingly, in the enzyme active pH range, the inhibitor forms a stronger salt bridge with the catalytic dyad in CatD than in BACE1, which might explain the off-target inhibition of CatD.

Our work highlights the importance of incorporating pH effects in structure-based drug design.

Introduction

Targeting β-secretase (BACE1) is a leading therapeutic route for prevention and treatment of Alzheimer’s disease (AD). BACE1 catalyzes the cleavage of β-amyloid peptide (Aβ) from amyloid precursor protein at the extracellular site. Subsequent oligomerization and fibrillation of Aβ lead to neuritic plaques which are linked to the onset of AD ¹. Hundreds of inhibitors have been discovered over the past decade, however, designing compounds with high selectivity has posed major challenges. A successful inhibitor must penetrate the central nervous system (CNS), cross the blood-brain barrier, and selectively target BACE1 in favor of other related proteases. Therefore, despite tremendous effort, no BACE1 inhibitors have passed the FDA required clinical trials and entered the market ^2,3.

Recently, the first small-molecule BACE1 inhibitor, a promising amino-thiazole based compound developed by Eli Lilly (LY2811376), entered early-stage clinical testing in humans but was discontinued due to retinal pathology observed in non-clinical toxicology studies with rats ⁴. The retinal toxicity was thought to be related to the off-target inhibition of cathepsin D (CatD) ⁵, an aspartyl protease involved in limited proteolysis ⁶. Mice deficient in CatD developed seizures and progressive retinal atrophy, which led to blindness ⁷. In human, CatD deficiency is associated with early blindness and progressive psychomotor disability ⁸. Therefore, effort has been made to optimize the inhibitor aspartyl binding motif (ABM) in order to increase BACE1/CatD selectivity ^5,9.

BACE1 is monomeric, while CatD is composed of two non-covalently associated chains generated from a post-translational cleavage and excision of an external loop ¹⁰. The sequences of BACE1 and CatD are also quite different. They share only 29% sequence identity and 46% sequence similarity (Figure 1a). Despite these differences, however, BACE1 and (active) CatD have very similar structure (Figure 1b and c) and function. Both proteins catalyze the hydrolysis of peptide bonds in acidic environment ¹¹. While BACE1 is primarily localized in the endosome and trans-Golgi apparatus with a narrow active pH range of 3.5–5.5 ^12,13, CatD is a major component of lysosome and functions in a broader pH range 2.5–6 ^12,14. The active site of BACE1 and CatD contains an aspartic dyad that acts as general acid and base in the catalysis. Substrates or inhibitors directly interact with the aspartic dyad via hydrogen bonds and/or salt bridges. Interestingly, both proteins contain a so-called flap, a ten-residue β-hairpin loop that lies directly over the catalytic dyad (Figure 1b and c). In BACE1, the flap is known to open and close at room temperature ¹⁵. In particular, the flap dynamics is mediated by the interaction between Tyr71, which is conserved among pepsin-like proteases, and the catalytic dyad controlling the access of substrate and inhibitor to the active site ^16,17.

Figure 1. Sequence/structure comparison of BACE1 and CatD as well as the first small-molecule BACE1 inhibitor.

a. BLAST ⁴⁴ sequence alignment of BACE1 (top, blue) and CatD (bottom, orange). The BLAST algorithm covered 92% of the entire sequence. Sequence identity is shown by shaded grey. The catalytic dyad, conserved Tyr, and interacting Trp are highlighted by the red, blue, and green boxes, respectively. The vertical black line in the CatD sequence indicates the separation between the two noncovalently-linked N-terminal light chain (15 kDA) and C-terminal heavy chain (33 kDA). b and c. Cartoon representations of BACE1 (PDB ID 1SGZ) ²⁰ and low pH, active conformation of CatD (PDB ID 1LYA) ¹⁰. The catalytic dyad of BACE1 (Asp32/Asp228) and CatD (Asp33/Asp231) are represented by orange and green surfaces, respectively. The light and heavy chains of CatD are colored pink and grey, respectively. d. The first small-molecule BACE1 inhibitor LY2811376 developed by Eli Lilly. Note, the endocyclic amine of the aminothiazine ring remains charged in the enzyme active pH range (see Methods and Protocols). The titratable pyrimidine nitrogen is circled in red (see Methods and Protocols).

A major difference between BACE1 and CatD is the pH-dependent conformational activation. We recently characterized the pH-dependent conformational dynamics that regulates BACE1 activity using continuous constant pH molecular dynamics (CpHMD) ¹⁷. Consistent with experiment, BACE1 adopts distinct conformational states that relate to the protonation state of the catalytic dyad. At high pH when the catalytic dyad is di-deprotonated the conserved tyrosine of the flap forms a stable hydrogen bond with the dyad and prevents substrate access, while within the active pH range 3.5–5.5, protonation of one Asp residue destabilizes this interaction and a binding competent state becomes populated. Additionally, we demonstrated how subtle changes in the protonation state of the catalytic dyad as well as the inhibitor can significantly alter the strength of binding between inhibitor and the target enzyme ¹⁸.

In contrast to the subtle change in BACE1, crystal structures show CatD undergoes a dramatic conformational rearrangement induced by pH. At high pH, CatD is in an self-inhibited state, where the N-terminal residues (1–16) form random coil and are inserted into the active site cleft, blocking the substrate access ¹⁹. In particular, Lys8 forms a salt bridge with both catalytic aspartates ¹⁹. At low pH, the salt bridges are destabilized, and the N-terminal random coil relocates by nearly 30 Å to form the sixth strand of the N-terminal β sheet of the enzyme ¹⁹.

Towards the design of highly selective inhibitors for BACE1, here we explore conformational dynamics of the active form of CatD and contrast it with that of BACE1 using continuous constant pH molecular dynamics. We also examine CatD binding to the aforementioned BACE1 inhibitor LY2811376. According to IC₅₀ values of in vitro essays, LY2811376 is only 6 times more selective for BACE1 than CatD ⁵. In rats studies, LY2811376 showed retinal toxicity that was attributed to CatD inhibition ⁴. Our work links the selectivity issue with the pH-modulated inhibitor binding behavior.

Results and Discussion

Active site environment and catalytic dyad titration

Although sharing only 29% sequence identity and 46% sequence similarity (Figure 1a), the structures of BACE1 and active CatD are remarkably similar (Figure 1b and c). The root-mean-square deviation (RMSD) between the 273 aligned Cα coordinates of the two crystal structures (PDB ID 1SGZ for BACE1 ²⁰ and PDB ID 1LYA for CatD ¹⁰) is 1.33 Å . The active sites of BACE1 and CatD are also quite similar. The two catalytic aspartic acids, Asp32/Asp228 in BACE1 and Asp33/Asp231 in CatD, are within hydrogen bond distance in the crystal structure, and surrounded by hydrophobic and polar residues within 7 Å of either of the dyad (Figure 2a). The flap overlaying the dyad does not contain any charged/titratable residue that can be directed toward the interior of the protein. Extending out to a 10 Å radius, one can see a lysine and an arginine that remain 10±1 Å from the catalytic Asp228 of BACE1; however, the positions of these two residues are occupied by hydrophobic groups, Val and Leu, in the CatD structure (Figure 2a and b). Interestingly, the hydration pattern of the catalytic dyad in CatD is very similar to BACE1. Specifically, a water molecule bridges the dyad across all pH, while an additional water enters above pH 4 (Figure 2b).

Figure 2. Environment, hydration and stepwise titration of the catalytic dyad in BACE1 and CatD.

a. Active site of BACE1 and CatD with hydrophobic and polar residues within 7 Å of either catalytic dyad shown as white and green surface, respectively. BACE1 contains two positively charged residues (Lys224 and Arg235) 10±1 Å away from the catalytic dyad, while CatD contains two hydrophobic residues (Val238 and Leu318) at these locations. b. Number of bridging water, defined as the average number of water molecules within 3.5 Å of both aspartic acids. The distance refers to that between water oxygen and Asp carboxylate oxygen. c. Total number of protons bound to the catalytic dyad at different pH. Curves represent the best fit to the stepwise titration model (Eq. 2). The first and second pK_a’s are respectively 4.1 and 1.8 for BACE1, and 4.7 and 2.9 for CatD.

Similarity of the active site environment is expected to give similar titration behavior for the catalytic dyad. Indeed, the first catalytic aspartate, Asp32 in BACE1 or Asp33 in CatD, has the calculated pK _a’s of 4.1 or 4.4 (Table 1 and Figure 3a). Interestingly, the calculated pK _a of the second catalytic aspartate, Asp228 in BACE1 (1.9) is 1.4 units lower than that of Asp231 in CatD (3.3). To understand this difference, we analyzed desolvation and hydrogen bonding, as charge-charge interactions are negligible (Figure 2a). Solvent accessibility of the two catalytic Asp residues was examined using fractional solvent accessible surface area (SASA) relative to the model Asp in solution. As seen in Figure 3b, the two aspartic acids are equally buried in BACE1 and CatD. The fractional SASA is about 6-8% in BACE1 and 11–15% in CatD; the difference is small. Interestingly, desolvation of the catalytic dyad is compensated by hydrogen bond formation of the charged aspartates with neighboring donors. To contrast the latter contribution in BACE1 and CatD, the number of hydrogen bonds (N_Hbonds) was counted in the entire simulated pH range (Figure 3c). Remarkably, the first catalytic aspartate in BACE1 as well as CatD and the second aspartate in CatD show nearly identical N_Hbonds; however, N_Hbonds for the second aspartate in BACE1 is larger by 1–1.5. A closer examination shows that Asp231 of CatD accepts hydrogen bonds from the neighboring backbone amide and sidechain hydroxyl groups (Gly233 and Thr234) (Figure 4). In contrast, Asp228 of BACE1 can accept additional hydrogen bonds from Thr33, Gly34 and Ser229. Thus, we suggest that the additional hydrogen bonds provide stabilization of the charged Asp228 in BACE1, causing its pK _a to shift 1.4 units lower than the counterpart in CatD.

Table 1.

Calculated pK_a’s of the catalytic dyad and the inhibitor

BACE1	apo	expt apo	holo	ΔpKa
Asp32	4.1	5.2	3.7	−0.4
Asp228	1.9	3.5	2.0	+0.1
Inhibitor	3.7	3.7	4.2	+0.5
CatD
Asp33	4.4	>5.0	2.4	−2.0
Asp231	3.3	4.1	3.0	−0.3
Inhibitor	3.7	3.7	4.9	+1.2

ΔpK_a refers to the calculated pK_a difference between the holo (bound) and apo (free) forms. Experimental apo pK_a’s for BACE1 were deduced from a kinetic measurement¹¹. Experimental apo pK_a’s for CatD were deduced from a kinetic measurement of short recombinant human pseudo-cathepsin D, which is believed to be kinetically similar to CatD¹⁴. Experimental apo pK_a for the inhibitor refers to the predicted pK_a of of the pyrimidine amine group (circled in red in Figure 1) by MoKa^36,37.

Figure 3. Titration, solvation and hydrogen bonding of the catalytic dyad in BACE1 and CatD.

a. Unprotonated fractions at different pH for the first (upper, Asp32 of BACE1 and Asp33 of CatD) and second (lower, Asp228 of BACE1 and Asp231 of CatD) catalytic residues. Residues in BACE1 and CatD are colored black and red, respectively. Curves are the best fits to the Hill equation. b. Solvent accessibility at different pH. Fractional solvent accessible surface area (SASA) was calculated using the SASA of carboxyl oxygens in the solvated model Asp as a reference. SASA was based a probe radius of 1.4 Å . c. Number of hydrogen bonds formed by the two catalytic residues. Hydrogen bonds were defined using a donor-acceptor distance cutoff of 2.4 Å (recommended setting in CHARMM HBOND analysis ³¹) Panels b and c use the same color scheme as panel a.

Figure 4. Stabilization of CatD catalytic aspartates by hydrogen bonding.

Only hydrogen bonds with an average occupancy higher than 10% are shown.

Now we compare the calculated pK _a’s with experiment. The calculated microscopic pK _a’s of Asp32 and Asp228 in BACE1 are 4.1 and 1.9 (Table 1), 1.1 and 1.6 pH units lower than the respective values, 5.2 and 3.5, estimated from a kinetic experiment ¹¹. The underestimation in pK _a’s is due to the underestimation of desolvation by the GBSW model ^21,22,23. As to CatD, the only publication we could find reported a single ionization event with a pK _a of 4.1 in a recombinant variant of CatD, and attributed it to the second catalytic Asp ¹⁴. Consistent with the aforementioned underestimation, our calculated pK _a for Asp231 (3.3) is 0.8 units below this estimate (Table 1). With regards to the second ionization event, the authors suggested that it occurs above 5.0; however they were not able to collect the data above pH 5 ¹⁴. Our calculated pK _a for Asp 33 (4.4) is 0.6 units below the lower bound, consistent with the systematic error discussed above. For coupled titration sites, it is instructive to examine the macroscopic or apparent stepwise pK _a’s. Since the first ionization event imposes an electrostatic penalty for the second ionization event, the two pK _a’s are split, with the extent of splitting representative of the degree of electrostatic coupling. By fitting the number of bound protons to the dyad to the stepwise titration model (Eq. 2), the first/second pK _a’s were obtained as 4.1/1.8 for BACE1 and 4.7/2.9 for CatD (Figure 2c). The pK _a splitting is 2.4 for BACE1 and 1.8 for CatD.

We suggest the different dyad titration behavior may have important implications for the mechanism by which each protease catalyzes peptide hydrolysis. BACE1 functions from pH 3.5 to 5.5 ^12,13. Within this pH range only Asp32 is protonated, while Asp228 is deprotonated. Therefore, only Asp32 can act as the acid in peptide hydrolysis. In contrast, CatD has a broader active pH range and hydrolyzes different substrates from pH 2.5 to 6.0 ¹². Given the smaller difference in the pK _a’s of Asp33/Asp231, thus, either Asp residue may act as the acid or base across the pH range.

Flap conformational dynamics

Next, we compare the BACE1 and CatD flap sequence and conformation. The BACE1 flap (Val67–Tyr78) and CatD flap (Phe74–Leu85) are β-hairpin loops that can partially cover the catalytic dyad. Interestingly, unlike the amino acids adjacent to the catalytic dyad, which are very similar, the BACE1 and CatD flap sequences are very different. Aside from the conserved tyrosine, Tyr71 in BACE1 and Tyr75 in CatD (Figure 1a, blue box) and a glycine located at the β-hairpin turn, no other residues within the flap are identical, and only one residue has the same type. The different sequence and interaction with the active site lead to altered flap dynamics in BACE1 and CatD. The BACE1 flap is flexible and the configurations exhibit pH-dependent behavior, while the CatD flap is relatively rigid and remains “open” in the entire pH range.

Previously we defined the conformational dynamics of BACE1 as Tyr-Inhibited, Gln-Inhibited, and binding competent using two order parameters, R and ϕ ¹⁷. R represents the distance between the conserved Tyr71 and catalytic dyad, while ϕ characterizes the twist of the flap. Specifically, R is defined as the distance between Tyr71:OH and Asp32:C^γ , while ϕ is defined as the pseudo dihedral angle Trp76:C-Val69:N-Thr72:C^α-Gln73:C^α. We now apply these two order parameters to examine the flap conformation of CatD. For CatD, R refers to the distance between Tyr75:OH and Asp33:C^γ , while ϕ refers to the pseudo dihedral angle for Leu83:C-Ile76:N-Gly79:C^α-Ser80:C^α in CatD.

Configurations with R < 5 Å are considered in a Tyr-inhibited state. At high and low pH, BACE1 is in the Tyr-inhibited state, as Tyr71 directly interacts with the catalytic Asp32, preventing substrate access (Figure 5a, solid black). In contrast, Tyr75 of CatD does not interact with the catalytic dyad in the entire active pH range 2–6 (Figure 5a, solid red). Within the BACE1 active pH range, protonation of Asp32 destabilizes the Tyr71-Asp32 hydrogen bond in favor of a binding competent state that has a shape complementary to the Tyr-Inhibited state (Figure 5a, dashed black). Within this pH range, BACE1 can also adopt a Gln-inhibited state, defined as R ≥ 5 Å and ϕ < −18°. In this state, the side chain of Gln73 can get close to the catalytic dyad, preventing substrate access, similar to the Tyr-inhibited state. However in CatD, the position of Gln is occupied by Ser which has a shorter side chain that does not engage the catalytic dyad. Therefore, the binding competent state is highly occupied across the entire active pH range 2–6 (Figure 5a, dashed red).

Figure 5. Flap conformation in BACE1 and CatD.

a. Occupancy of the Tyr-inhibited state (solid) and binding competent state (dashed) from BACE1 (black) and CatD (red) simulations. b. Occupancy of hydrogen bonding Tyr71… Trp76 in BACE1 (black) and Tyr75 … Trp40 in CatD (red). c/d. Snapshots of the Tyr… Trp hydrogen bonded configuration from the BACE1 and CatD simulations.

The different BACE1/CatD flap conformation can be largely attributed to the conformation dynamics of the conserved tyrosine. Tyr71 modulates the BACE1 flap conformation ^16,17, however, the Tyr75 conformations remains relatively rigid. The χ₁ angle of Tyr71 in BACE1 alternates between the three rotamer states, TyrUp (χ₁ ≈ +60°), TyrDown (χ₁ ≈ ±180°), and TyrBack (χ₁ ≈ −60°), while the χ₁ of Tyr75 in CatD remains in the TyrBack configuration over 90% of the time. The TyrBack configuration directs the side chain toward the base of the flap which increases hydrogen bonding with a tryptophan residue (Figure 5c and d). Notably, the tryptophan that interacts with the conserved flap tyrosine is located in different regions of each protease. In the binding competent state, Tyr71 of BACE1 points towards the base of the flap and forms a stable hydrogen bond with Trp76 (Figure 5c). Interestingly, the conserved flap tyrosine of CatD, Tyr75, also forms a hydrogen bond with a tryptophan, Trp40, which is located in a different part of the protein and the occupancy is much higher that in BACE1 (Figure 5d). In the TyrBack configurations, Tyr is no longer able to interact with the catalytic dyad, thus leaving the active site open (Figure 5c and d).

Since the flap tyrosine in CatD does not engage the active site as it does in BACE1, we use another order parameter Z (following Spronk and Carlson ¹⁶) to further characterize the openness of the active site. Z measures the distance between the tip of the flap (Thr72:C^α of BACE1 or Gly79:C^α of CatD) and the first catalytic aspartic acid (Asp32:C^γ of BACE1 or Asp33:C^γ of CatD) (Figure 6a). We then examine the free energy surface in terms of Z and aforementioned pseudodihedral angle ϕ (Figure 6b). In going from high to low pH, the FES for BACE1 clearly shows a conformational switch between pH 4 and 5 (Figure 6c). At pH 5, Z is small (less than 10 Å ), indicating a closed active site. The deep minimum on the FES (with Z around 8 Å ) corresponds to the aforementioned Tyr-inhibited state. As the pH is decreased to 4, the deep minimum disappears and regions in the FES with Z value greater than 10 Å become populated, indicating the Tyrinhibited state is destabilized in favor of the binding competent state (with the open active site).

Figure 6. Free energy surface analysis of BACE1 and CatD flap dynamics.

a/b. Illustration of the order parameter Z and ϕ. Definitions are given in the main text. c/d. Free energy surface for BACE1 and CatD at pH 3, 4, 5 and 6. Color scale labels are given in kcal/mol.

Comparing the FES of CatD with BACE1, the major difference is in the pH conditions above 4. In contrast to BACE1, the FES of CatD shows a very similar minimum region with Z in the range of 10–12 Å at all pH, which indicates the flap remains open within the CatD active pH range (pH 2.5–6) and no conformational switch is observed. Additionally, the minimum on the FES of CatD is ≈ 1 kcal/mol deeper than a majority of the minima on the FES of BACE1, indicating that the flap is less flexible than BACE1. Therefore, the increased active pH range of CatD can be attributed to a greater substrate accessibility across a wider pH range that results from increased openness of the flap, rigidity of Tyr75 conformation, and lack of Gln73-inhibited state.

Inhibitor binding to BACE1 and CatD

Finally, we compare the pH-dependent binding of inhibitor LY2811376 to BACE1 and CatD. In the presence of inhibitor binding, the pK _a’s of the catalytic dyad in BACE1, Asp32 and Asp228, do not change much. The shifts are -0.4 and 0.1, respectively. The former pK _a downshift of Asp32 is due to the salt bridge/hydrogen bond formation between Asp32 and the positively charged endocyclic amine on the inhibitor ABM (Figure 1). Surprisingly, despite the nearly identical binding mode (Figure 7a), inhibitor binding produces a larger effect on the catalytic pK _a’s of CatD. The respective shifts for Asp33 and Asp231 are -2.0 and -0.3 respectively. Thus, the salt bridge between Asp33 and the inhibitor ABM is much stronger in CatD. Interestingly, inhibitor binding also impacts the protonation state of the amine group on the pyrimidine ring of the inhibitor (Figure 1d). The pK _a is shifted up by 0.5 and 1.2 pH units to 4.2 and 4.9 in BACE1 and CatD, respectively. This indicates the protonation behavior is sensitive to subtle differences in the binding site.

Figure 7. pH-dependent binding of inhibitor LY2811376 to BACE1 and CatD.

a. Snapshot of the bound and unbound configurations. b. Unprotonated fractions of Asp32 in BACE1-inhibitor complex (black) and Asp33 in CatD-inhibitor complex (red) at different pH. The curves are best fits to the modified Hill equation. The resulting pK_a’s for Asp32 of BACE1 and Asp33 of CatD are 3.7 and 2.4, respectively. c. Occupancy of bound configurations at different pH. A bound configuration is defined as that with a distance between the inhibitor endocyclic amine of the ABM and the carboxylate carbon of Asp33 smaller than 4 Å . Data for BACE1 and CatD is colored in black and red, respectively. Grey shaded region indicates the active pH range of CatD. The standard deviation was calculated by separating the trajectory into three, 6 ns blocks and calculating the percent of bound configurations from each block.

In our previous study, we found that BACE1 is stably bound to LY2811376 above pH 6; however, in the enzyme active pH range (2.5–4.5 in the simulation), the inhibitor becomes dissociated (Figure 7c, black curve) ¹⁸. In the inhibitor bound state, the endocyclic amine of the ABM forms a salt bridge/charged hydrogen bond with Asp32. Protonation of Asp32 in this pH range (Figure 7b, black curve) leads to the breakage of this salt bridge and thus dissociation. Interestingly, inhibitor-CatD binding has a similar behavior, i.e., stable above pH 6; however, dissociation occurs below pH 5, although to a lesser degree in the pH range 2.5–4.5 (Figure 7c, red curve). CatD-inhibitor binding involves a stronger salt bridge, which shifts the pK _a of Asp33 down to 2.4. However, since Asp33 of CatD has a very broad transition (Hill coefficient of 0.70), partial deprotonation occurs readily just below pH 4. We suggest this may be the reason why the dissociation of CatD-inhibitor complex occurs in the pH range 2.5–4 but to a lesser degree.

Conclusion

BACE1 and CatD flap dynamics differ dramatically despite similar structure. The BACE1 flap is flexible and occupies pH-dependent conformational states, while CatD flap is rigid and remains in an open state across the entire simulated pH range. This is likely why CatD is active over a broader pH range compared to BACE1. Despite similar active site environments, the pK _a’s of BACE1 and CatD catalytic dyads are quite different. In BACE1 active pH range, Asp32 clearly acts as the acid in peptide hydrolysis, while Asp228 acts as the base. By contrast, in CatD the pK _a’s of Asp33 and Asp231 are very similar and either could act as acid or base. The protonation states of enzyme’s active site is crucial for understanding not only the pH-dependent enzyme activity but also inhibitor binding. Binding of LY2811376 to both BACE1 and CatD is stabilized by a salt bridge formed between the first catalytic residue and the endocyclic amine of the inhibitor ABM. Remarkably, while the pK _a of Asp32 in BACE1 is shifted lower by only 0.4 units, the corresponding pK _a of Asp33 in CatD is shifted lower by 2 units to 2.4. The lower pK _a, together with a broad transition, results in the partial protonation of Asp33 and destabilization of the Asp33-ABM salt bridge in the pH range 2.5–4. We suggest this may be a major contributing factor for the increased population of inhibitor-bound configurations in the enzyme active pH range, which might explain the retinal toxicity due to CatD inhibition, despite the IC₅₀ values showing six times selectivity for BACE1. We note that, in order to reach quantitative conclusion, pH-dependent binding free energy calculations and perhaps more extensive conformational sampling are needed, which are underway in our laboratory. Nevertheless, the data presented here highlights how subtle differences in structure can have large impact on protonation states, which in turn modulates the pH dependence of protein-ligand binding. To overcome the challenge of designing selective BACE1 inhibitors, structure-based design methods must incorporate the effects due to protonation state change in the enzyme active site as well as the inhibitor.

Methods and Protocols

Continuous constant pH molecular dynamics

This work is based on continuous constant pH molecular dynamics (CpHMD) simulations ^24,25,26,23, in which conformational state of the system is evolved simultaneously with proton titration of all ionizable sites of the solute at a specified pH condition. Importantly, CpHMD enables protonation states to change in response to the change in environment and solution pH, which is critical for investigating enzymes that function in a pH-dependent manner and require proper protonation states for optimal function. Specifically, we applied the CpHMD method with a hybrid-solvent scheme and pH replica-exchange (REX) sampling protocol ²³. Hybrid-solvent CpHMD calculates solvation forces on titratable residues using the generalized Born (GB) implicit-solvent model, while sampling conformational states using explicit solvent ²³. Combined with the pH REX protocol, this method provides converged pK _a’s based on 1 ns sampling per pH replica with an average absolute error of 0.5 pH unit for most protein residues ²³. For deeply buried and enzyme active-site residues, the sampling time requirement is higher, about 5 ns, and the error is up to 1–1.5 unit ^23,27. Significantly, we recently demonstrated hybrid-solvent CpHMD is able to predict the active-site pK _a’s of BACE1 enzyme with the correct order and a systematic error of about 1–1.5 pH units ¹⁷.

Structure preparation and force field parameters

The apo (free) BACE1, holo (inhibitor-bound) BACE1, and small molecule parametrization have been previously described ^17,18. The X-ray crystal structure (PDB ID 1LYA) of the active CatD ¹⁰ was used as the starting structure for apo CatD simulations. Inhibitor LY2811376 was docked to CatD by aligning the apo crystal structure with the crystal structure of BACE1 bound to the same inhibitor (PDB ID 4YBI) ⁴, using Multiprot Sever ²⁸. Each system was solvated in an octahedral box with at least 10 Å between the box edge and the protein. In all simulations, the protein was represented by the CHARMM22/CMAP all-atom force field ^29,30, and water was represented by the CHARMM modified TIP3P model ³¹. The force field parameters for the inhibitor were obtained in house following the protocol of CHARMM General Force Field (CGenFF) ³² . The partial charges are included in Supporting Information. The topology and parameter files are available upon request.

Equilibration and production runs

All simulations and energy minimization were performed using CHARMM program (version c37b1) ³¹. Functionalities related to CpHMD and pH REX were implemented in the PHMD and REPSTR modules, respectively. First, the system was gradually heated over the course of 120 ps from 100 K to 300 K with the protein heavy atoms harmonically restrained using a force constant of 5 kcal/mol/Ų. Following heating, the system underwent 180 ps restrained equilibration, where the force constant was gradually reduced from 5 (40 ps), to 1 (40 ps), and 0.1 kcal/mol/Ų (100 ps). Lastly, the system underwent 100 ps unrestrained equilibration. In the energy minimization and equilibration steps, the hybrid-solvent CpHMD function was turned on with pH set to 7.

Production simulations were performed with the pH REX hybrid-solvent CpHMD method ²³. Each pH replica was simulated in the NPT ensemble at a temperature of 300 K and pressure of 1 atm. Temperature and pressure were maintained using a modified Hoover thermostat ³³ and Langevin piston coupling method ³⁴, respectively. The van der Waals interactions were smoothly switched to zero between 10 and 12 Å . The particle mesh Ewald method ³⁵ was used to calculate long-range electrostatics, with a real-space cutoff of 12 Å and a sixth-order interpolation with 1.4 Å grid spacing. The SHAKE algorithm was used to constrain bonds involving hydrogen to enable a 2-fs timestep.

CpHMD specific protocols

The titratable sites in CpHMD included all Asp, Glu, and His residues on the protein and the pyrimidine nitrogen of the inhibitor (Figure 1d). The endocyclic amine group within the aspartyl binding motif (ABM) was kept in the charged state, because its pK _a is expected to be higher than the apo (solution) pK _a (measured as 8.7 ¹⁸) due to the stabilizing salt-bridge interaction with the catalytic aspartate (Asp32 in BACE1 or Asp33 in CatD). For titrating the pyrimidine amine, the apo or solution form of the inhibitor was used as a model. The model pK _a was calculated as 3.7 by the widely used, QSPR-based program MoKa ^36,37. MoKa makes use of a training set of ~25,000 experimentally measured pK _a values (including 685 pyrimidine derivatives) and atomic descriptors from 3D molecular interaction fields generated on a very large set of fragments ^36,37. The internal and external validations gave RMSE of about 0.5 and 1 pH unit, respectively ^36,37. Admittedly, the use of a predicted model pK _a may introduce an unwanted error. An alternative approach is to use pyrimidine alone as a model compound which has a known experimental pK _a. However, with this model, the electronic effects due to substitution at the 5 position are not accounted for. Further, the bonded parameters as well as the partial charges of pyrimidine are likely different from those of the pyrimidine ring in the inhibitor, which is a situation that cannot be handled by the current CpHMD framework ^25,23. The extent of the electronic effects and the best way to construct model compounds for small molecules are important topics that will require our attention in the future.

In CpHMD, an electrostatic solvation free energy calculation using the generalized-Born implicitsolvent model GBSW ²¹ was invoked every 5 MD steps to update the titration coordinates. In the GBSW calculation, the default settings were used, consistent with our previous work ²³. The GBSW input radii for the protein were taken from Chen et al. ³⁸. The input radii for carbon and sulfur atoms in the inhibitor were set to 2.0 and 2.3 Å following Nina et al ³⁹. All CpHMD settings (for both protein and inhibitor) were identical to our previous work ²³. The apo CatD system was simulated for 31 ns per replica with the fist 1 ns discarded in analysis. The inhibitor-bound system was simulated for 36 ns per replica, with the first 18 ns discarded in analysis. All simulations used 24 pH replicas in the pH range 1-8. The pH conditions of the 24 replicas of each simulation are presented in SI. The average replica-exchange acceptance ratio was 49±8%. Thus, the aggregate production sampling time for apo and holo CatD is 744 and 864 ns, respectively.

pK_a calculations

The λ (representing proton titration) and x (representing tautomer intercon-version) values ²⁵ were collected at each REX exchange attempt. The protonated state was defined as λ < 0.1 and (x < 0.1 or x > 0.9). The unprotonated state was defined as λ > 0.9 and (x < 0.1 or x > 0.9). The residue-specific pK_a’s were calculated by fitting the fraction of unprotonated states S at different pH to the modified Hill equation,

$$S=\frac{1}{1+{10}^{n(\text{p}{K}_a-\text{pH})}},$$ (1)

where n is the Hill coefficient that represents the slope of the transition region in the titration curve.

For the catalytic dyad, the macroscopic stepwise or sequential pK_a’s are also of interest. The two sequential pK_a’s can be obtained by fitting the total number of bound protons (to the dyad), N_prot, to the following statistical mechanics based model ^40,41,

$$N_{\text{prot}}=\frac{{10}^{\text{p}{K}_2-\text{pH}}+2\times {10}^{\text{p}{K}_1+\text{p}{K}_2-2\text{pH}}}{1+{10}^{\text{p}{K}_2-\text{pH}}+{10}^{\text{p}{K}_1+\text{p}{K}_2-2\text{pH}}},$$ (2)

where pK₁ and pK₂ are the two macroscopic pK_a’s, and the denominator represents the partition function.

Alternatively, the total unprotonated fractions can be fit to the decoupled, biphasic titration equation ^42,43,

$$S_1+S_2=\frac{1}{1+{10}^{\text{p}{K}_1-\text{pH}}}+\frac{1}{1+{10}^{\text{p}{K}_2-\text{pH}}},$$ (3)

where S₁ and S₂ are the unprotonated fractions of the two catalytic residues. N_prot from Eq. 2 are linked to S₁ and S₂ via the following equation,

$$N_{\text{prot}}=2-S_1-S_2.$$ (4)

We verified that both models (Eq. 2 and Eq. 3) gave identical stepwise pK_a’s for BACE1 and CatD.