Computational study of the inhibitory mechanism of kinase CDK5 hyperactivity by peptide p5 and derivation of a pharmacophore

Abstract

The hyperactivity of the cyclic dependent kinase 5 (CDK5) induced by the activator protein p25 has been linked to a number of pathologies of the brain. The CDK5-p25 complex has thus emerged as a major therapeutic target for Alzheimer’s disease (AD) and other neurodegenerative conditions. Experiments have shown that the peptide p5 reduces the CDK5-p25 activity without affecting the endogenous CDK5-p35 activity, whereas the peptide TFP5, obtained from p5, elicits similar inhibition, crosses the blood-brain barrier, and exhibits behavioral rescue of AD mice models with no toxic side effects. The molecular basis of the kinase inhibition is not currently known, and is here investigated by computer simulations. It is shown that p5 binds the kinase at the same CDK5/p25 and CDK5/p35 interfaces, and is thus a non-selective competitor of both activators, in agreement with available experimental data in vitro. Binding of p5 is enthalpically driven with an affinity estimated in the low μM range. A quantitative description of the binding site and pharmacophore is presented, and options are discussed to increase the binding affinity and selectivity in the design of drug-like compounds.

I. Introduction

The cyclin-dependent kinase 5 (CDK5) plays a central role in cognition, phosphorylation of cytoskeletal proteins, and other vital brain functions [1–3]. Under normal physiological conditions, CDK5 is regulated by neuron-specific activators p35 and p39. Under oxidative stress and other conditions, higher concentrations of Ca²⁺ activate the protease calpain, which cleaves p35 into fragments p25 and p10 [4–7]. The stable CDK5-p25 complex has a prominent role in the aberrant phosphorylation of tau proteins, which leads to the formation of β-amyloid plaques (BAP), neurofibrillary tangles (NFT), extracellular senile plaques of amyloid-42 (A), and intracellular NFT caused by accumulation of hyperphosphorylated proteins. Neuronal death is another consequence of CDK5 hyperactivity [4, 8, 9]. The CDK5-p25 complex has thus emerged as a major therapeutic target for Alzheimer’s disease (AD) and other neurodegenerative conditions. The hyperactivity of CDK5 has been the focus of numerous studies, although effective therapeutic protocols are lacking. Drugs targeting BAP clearance [10, 11], cholinesterase inhibitors [12], and antioxidants to relieve oxidative stress [13, 14] have been considered, but further studies are needed to assess their effectiveness and potential side effects. Inhibitors targeting ATP binding have also been considered, including aminothizole and roscovitine [15–17], but they lack specificity as both aberrant and physiological activities are inhibited, leading to side effects and reduced therapeutic effectiveness.

A promising approach for the selective inhibition of CDK5-p25 is the use of small peptides obtained by truncation of p35 [18]. In vitro and in vivo data have shown that the truncated 24-residue peptide p5 specifically reduces CDK5-p25 activity, without affecting the endogenous CDK5-p35 function in presence of certain neuron specific molecules [7, 19]. A peptide derived from p5, TFP5, has recently been shown to cross the blood brain barrier and rescue cortical neurons without causing side effects [20]. This experimental evidence shows the potential of p5-based compounds to reach the brain and selectively inhibit the CDK5-p25 activity. However, the molecular basis of inhibition is not presently known, a problem exacerbated by the scarcity of structural information of p5 and the CDK5-p5 complex. This information is crucial to rationally design small peptide-mimetic drug-like compounds that can bind selectively and with high affinity to the same site(s) in the kinase.

The molecular basis of the inhibitory mechanism of p5 is here investigated by computer simulations. The study builds upon a series of recent developments on ab initio complex structure prediction [21, 22] and force field optimization [23–25]. The predicted binding modes suggest that p5 is a non-selective competitive inhibitor of p25, in accordance with available experimental data. Binding of p5 to the kinase is enthalpically driven with an affinity estimated in the low μM range. A quantitative description of a pharmacophore is presented, on the basis of which options are discussed to increase the binding affinity and specificity.

II. Results

The prediction of the CDK5-p5 complex structure in aqueous solution and at physiological conditions follows the general method described in [21, 22]. The method consists of three stages, in this case: i) prediction of p5 and CDK5 conformers in solution; ii) identification of CDK5-p5 pre-relaxation binding modes; and iii) structural relaxation of the complexes at physiological conditions. The method makes no assumptions regarding geometric complementarity prior to binding, as both molecules can undergo structural changes upon contact. The method is thus designed to account for conformational selection of both molecules and for mutually induced fit upon association. The method is also designed to identify weak and ultra-weak modes of association, which have been shown to play a role in molecular recognition [26, 27]. The method yields a set of conformational families of the complex, from which a discrete set of conformations (the “binding modes”) are identified. These modes are used to define the binding site(s) and possible pharmacophore(s). Application to the CDK5-p5 system makes use of the amino acid sequence of p5 since its three dimensional structure in solution is unknown, and the atomic coordinates of CDK5 in the active state, as obtained from the CDK5-p25 crystal [28].

II.1 Prediction of p5 conformers in solution

The amino acid sequence of p5 is ¹KEAFWDRCLSVINLMSSKMLQINA²⁴, with both termini uncapped. All the residues are assumed to be neutral at pH 7 except K⁺, R⁺, E⁻ and D⁻. The technique used to predict the conformers of p5 in solution has been described [23, 29]: first, a stochastic conformational optimization based on Monte Carlo Minimization/Annealing (MCMA) [29] simulations is performed to find a set of N structurally distinct low-energy conformations. Biased movements of side-chain or main-chain dihedral angles are generated and energy-minimized before a decision on acceptance is made. The N optimized conformations are used as initial structures in a series of Replica-Exchange (REx) Langevin Dynamic simulations. Replicas are distributed over temperatures in ascending order of energies, with the lowest energy assigned to the lowest temperature where data are collected (here, 37 °C). Simulations are carried out with the all-atom CHARMM force field [30] and the SCP implicit solvent model [23, 25, 31, 32]. The converged canonical distribution contains a large number of structures, which are clustered to obtain a discrete set of conformational families [22]. The clustering algorithm [33] assumes a maximum intra-cluster C_α-rmsd variance of 2 Å, which determines the number of families. Cutoff values used for clustering should in principle be chosen based the expected structural fluctuations around a given conformational state, and are thus system dependent. Order parameters and B-factors from proteins in the PDB suggest that these fluctuations are typically less than 1.5 Å. Molecular dynamics simulations, however, tend to yield larger structural fluctuations, partially due to the accuracy of current force fields to reproduce properties of the folded state. Therefore, slightly larger cutoffs are used here to avoid potentially spurious sub-states that would increase the computational cost unwarrantedly. All clusters with populations > 5% are considered. A representative member of each family is then selected and identified as a conformer of p5. The calculation for p5 yields four conformers (Fig. 1), each with relatively small populations but all sharing common structural motifs, similar to those of the same sequence in p25: a stable α-helix spanning residues K1-S10 and a less stable helix between I12 and A24; a kink around V11 may provide flexibility for fast interconversion. Each of the conformers has the potential to bind the kinase in a variety of modes and affinities and elicit specific biological responses.

Figure 1.

Four predicted conformers of peptide p5 in aqueous solution at 35 °C and pH 7. Common structural motifs: K1-S10 α-helix (red), I12-A24 α-helix (blue), and a kink around V11 (grey) may provide flexibility for fast interconversion between conformers. Absolute populations indicated in parentheses.

II.2 Prediction of CDK5-p5 pre-relaxation binding modes

All the conformers of p5 are treated as independent structures coexisting in solution with conformers of CDK5. Unlike p5, for which a canonical ensemble could be obtained, thorough sampling of the conformational space of CDK5 is not feasible. However, the crystal structure of the CDK5-p25 complex (PDB ID: 1UNL) can be used to obtain a family of conformations in the vicinity of the active state of CDK5 through molecular dynamics simulations. The kinase was thus isolated from the crystal, immersed in water, and subjected to a 50-ns MD simulation at 37 °C and 1 atm (protocol described in Section II.4). The trajectory was analyzed with the same clustering algorithm used for p5, using an intra-cluster C_α-rmsd variance of 2.5 Å. Two main conformational families were identified, each with about 50% of the population. One family contains the active state, and its representative member is similar to the conformation of CDK5 in the complex. The conformer of the second family differs from the active conformer mainly in the activation loop. Further analysis of its dynamics indicates that this conformer does not represent a true (locally stable) sub-state, but results merely from transient fluctuations of the loop through the dynamics, probably on its path towards the transition state connecting the active and the inactive conformer of the kinase. In the latter case the activation loop adopts a stable retracted helical motif [34, 35], the structure of which is not known and is beyond the reach of the present simulations. The focus in this study is on the inhibition of the active kinase and thus only the first conformer is used. The pharmacophore described in Section II.5 is thus intended for targeting this active state. Other pharmacophores may eventually be developed for the inactive state.

The method used to predict the binding modes consists of two consecutive stages [21, 22]: a prescreening of favorable CDK5-p5 interactions aimed to identify physically relevant first-encounter modes, followed by a configurational biased sampling to identify statistically relevant conformational families of the complex. Both stages were described in detail previously [21, 22]; briefly, the prescreening consists of a simulated-annealing MC optimization of an electrostatic norm e and a hydrophobic norm h [21]. The functional forms of e and h are suitable simplifications of the physical effects that each one intends to describe, and they were designed specifically for computational efficiency: optimization of e represents electrostatic complementarity between the protein surfaces, whereas optimization of h represents the degree of burial of the local hydrophobic surfaces. No assumptions are made about surface complementarity because both proteins may undergo post-binding structural relaxations, and this is addressed in the final stage.

The prescreened modes are subsequently used to construct a biasing function for efficient conformational sampling using adaptive biased MC simulations with the complete force field. The resulting canonical distribution obtained upon convergence is post-processed by the clustering algorithm to identify a discrete set of binding modes using a cutoff of 2 Å. Simulations are carried out with the all-atom CHARMM force field [30] and the SCP implicit solvent model [23, 25, 31, 32]. The distributions of first-encounter modes (Fig. 2, upper row) are relatively broad but confined to specific regions of the kinase, most notably on the small domain. The calculations yield five structurally distinct CDK5-p5 binding modes (Fig. 2, middle row), with one conformer (c₃) binding to two sites of the kinase. None of the modes appear to interfere directly with the ATP or the substrate binding sites. One of the modes (indicated by arrow) deserves special consideration since it binds the kinase at the known CDK5/p25 interface, and also at the predicted binding site of inhibitors p6 and CIP [22, 35, 36]. This site lies in the vicinity of three critical structural motifs implicated in both p25 and p35 binding and in the CDK5 activation and function [9, 18, 19, 28, 34, 35]: the PSAALRE helix (blue in Fig. 3), a loop rich in acidic residues (purple), and the activation loop (red). Unlike the other predicted binding sites, which may or may not be occluded by other proteins in vivo (e.g., the synaptic, vesicle-associated protein p67), this site is fully accessible to the solvent, hence to p5 or other inhibitors or activators. These observations suggest that inhibition of the aberrant CDK5-p25 hyperactivity by p5 (and presumably by TFP5 as well; cf. Section II.3) can be explained as a competition for binding to the same site of the kinase.

Figure 2.

Predicted CDK5–p5 complexes in aqueous solution at 35 °C: distribution of first-encounter modes (top row); pre-relaxation binding modes (middle row); and post-relaxation binding modes at physiological conditions (bottom row). The arrow indicates the predicted inhibitory mode.

Figure 3.

A (X-ray structure) and B (snapshot from the dynamics): Ribbon representations of CDK5 showing key structural motifs involved in binding and regulation: acidic loop (purple), PSAALRE helix (blue), activation loop (red), and β-sheet structure of the small lobe (green); ATP- and substrate-binding pockets indicated. C: molecular surface of the CDK5 binding interface for the proposed inhibitory mode (same orientation as in B). Interfacial residues in contact with p5 obtained from six independent dynamics simulations shown in green and yellow (see text): D41, E42, P45, S46, L49, R50, C53, L54, K56, E57, R120, N121, L123, R125, L147, R149, A150, F151, G152, I153, P154, V155, R156, Y158, S159. The interfacial residues corresponding to the other binding modes are: for c₁: D39, G43, S46, R125, R149, V155, R156, C157, Y158, S159, E161; for c₂: K213, R214, R217, L218, Y242, P243, T245, T246, N250, V251; for the non-inhibitory c₃: K3, T26, E28, K59, H60, K61, R65; for c4: K59, H60, K61, N62, R65, K112, F116, C290.

To evaluate the strength of the CDK5-p5 interaction in this pre-relaxation inhibitory mode, the binding enthalpy (ΔH) and binding entropy (ΔS) are calculated separately. The dissociation constant is given by K_d = c^Øexp(ΔG/kT), where c^Ø = 1 M is the standard concentration and ΔG = ΔH − T ΔS is the free energy of binding. Calculations are carried out at 37 °C using the all-atom CHARMM force field and the SCP solvent model. The enthalpy is calculated from the canonical distribution of the complex, obtained from a MC simulation [21, 22] as ΔH ≈ <E_b> − <E_∞>, where <E_b> = Z⁻¹Σ_i E_i exp(−E_i/kT) is the non-bonded energy of the bound state, Z is the partition function, k is the Boltzmann constant, and T is the absolute temperature; <E_∞> is the non-bonded energy of the unbound complex. Trial moves consist of rigid-body rotations and/or translations, or rotations of a single side-chain torsion angle, and were carried out with a non-adaptive configurational-bias sampling [21]. The calculation yields ΔH ~ – 10 kcal/mol, the largest value of the five modes. The entropy is given by ΔS = ΔS_rt + ΔS_c, where ΔS_rt is the change in rotational-translational entropy upon binding, and ΔS_c is the corresponding change in configurational entropy, which include vibrational and (bond) rotational components. Based on previous studies [37, 38] the former is here assumed to be TΔS_rt ~ −2 kcal/mol, whereas ΔS_c is calculated as the difference ΔS_c = S_b − S_∞, where S_b and S_∞ are the absolute configurational entropies of the complex in the bound and unbound states, respectively. These components to S are here estimated from the atomic fluctuations using the approximation [39] S = k + k/2 ln[det (kT/ħ² Mσ + I/12)] where M and I are the mass and unit matrices, and ħ is the reduced Planck constant. The covariance matrix σ is obtained from Langevin dynamics simulations. Convergence of the cross-correlation terms is rather slow, so to speed up convergence the C_α atoms were fixed, and only the contribution of p5 were evaluated; even in this case simulations of 5–10 μs length were required for convergence. The calculation yielded an unfavorable configurational entropy of about TΔS_c ~ −1 kcal/mol. The free energy of the solvent is incorporated into the energy function of the implicit model, so the solvent entropy is effectively embedded in the enthalpic contribution. In the SCP model the entropy is absorbed mainly in the cavity term and in the screening function, and is expected to be small; however, other solvent contributions not amenable to a mean field approximation may be important [21] and would require further analysis. The calculations then suggest that the initial binding of p5 to CDK5 is enthalpically driven and has an affinity in the μM range; only a modest increase in affinity could be achieved by reducing the internal flexibility of the ligand. A post-relaxation calculation, not conducted here, might yield a higher affinity due to structural adaptations at the surface and increased complementarity (cf. Section II.4), although enthalpy-entropy compensation is always present and its effects are difficult to predict.

II.3 Experimental evidence of selective and non-selective inhibition

In vivo and in vitro experiments [7, 19], have uncovered important aspects of AD-related CDK5 activity and hyperactivity. These include the selective inhibition of CDK5 hyperactivity by small peptides, such as p5, in the presence of neuron specific molecules. The results of the previous section imply that, in the absence of other proteins that could affect the predicted binding mode, p5 should inhibit both CDK5-p25 and CDK5-p35 activities similarly (non-selective inhibition). This conclusion is indeed supported by experimental data in vitro: in the absence of cytoskeletal or neuron specific molecules, such as p67, p5 displays similar inhibitory effects on the pathological and physiological activities. On the other hand, in the presence of p67, p5 selectively inhibits p25-induced deregulation and hyperactivity in vivo, showing only marginal effect on p35 activation (selective inhibition) [19]. Data were also obtained in vivo for TFP5, a modified, longer peptide derived from p5: intraperitoneal injections of TFP5 on 5XFAD AD model mice show that TFP5 crosses the blood-brain barrier, whereas mice injected with TFP5 exhibited behavioral rescue with no toxic side effects [7]. Evidence of selective inhibition of abnormal CDK5 hyperactivity was also found. Additional in vitro kinase assays in presence of p67 and in vivo experiments on transfected Human Embryonic Kidney 293 (HEK 293) cells have confirmed the selective inhibition of CDK5 hyper-activity by TFP5 (to be published).

The available experimental evidence thus supports the notion that inhibition is non-selective in the absence of p67, and the inhibitory mode predicted here provides the molecular rationale for such non-selective behavior. On the other hand, in the presence of p67, both p5 and TFP5 inhibit the CDK5-p25 activity but not the CDK5-p35 activity in a dose-dependent manner. On the basis of these data it is proposed here that the 10-KDa N-terminal domain of p35 acts as a protector in situ/in vivo, possibly by shifting the thermodynamic equilibrium in favor of CDK5-p35 by interacting with p67 and/or microtubules. The molecular mechanism of this selectivity cannot be elucidated from the present study, as it requires an extension of the method to multi-protein systems [22].

II.4 Structural relaxation and mutually induced fit

The five predicted binding modes were subjected to dynamics simulations to analyze structural relaxation following the initial binding. The simulations were performed in the isothermal–isobaric ensemble at 37 °C and 1 atm; temperature and pressure were maintained with the Langevin and the Nosé–Hoover methods. The TIP3P water model and the all-atom CHARMM force field were used. The system was neutralized with Cl⁻ ions, and ~150 mM of K⁺ and Cl⁻ ions were added to mimic the intracellular ionic strength; residues R⁺, K⁺, E⁻ and D⁻ were assumed to be charged at pH 7, and protonation states were fixed throughout the simulations. Particle mesh Ewald summations and orthorhombic periodic boundary conditions were used; the simulation cells consisted of cuboids with dimensions chosen so that the minimum distance between protein images was ~4 nm. Additional specifications of the simulations were the same as reported previously [35]. The total simulation times varied between 30 and 50 ns, depending on the complex, and included an initial period of structural adaptation and at least 10 ns of structural stability, which were used for analysis. Six simulations were carried out for each binding mode for statistical purposes.

All the binding modes undergo conformational changes in various degrees. Figure 2 (bottom row) shows snapshots of the relaxed modes at the end of the simulations. The changes occur mainly in p5, and to a lesser extent in unstructured segments of the small domain of CDK5; the large domain of the kinase remains largely unaffected, as observed previously in the dynamics of the CDK5-p25 and CDK5-CIP complexes [35]. The changes in p5 originate mainly in rearrangements of the helices, although partial unfolding is observed in some cases. The major changes in the complexes occur in the relative positions and orientations of p5 and CDK5, resulting in an increase in surface complementarity at the CDK5/p5 interface. This is more apparent in the inhibitory mode, which is here analyzed in more detail. Figure 3 shows the region on the surface of CDK5 corresponding to the CDK5/p5 binding interface for this binding mode. The surface includes all the atoms that are in contact with p5 at any point of the trajectories; a distance criterion is used to determine whether two atoms, i and j, are in contact, namely, if their separation r is such that no water molecule could fit in between (i.e., r < 2.8 Å + R_vdw,i + R_vdw,j). The interface includes residues (green) D41, E42, P45, S46, L49, R50, C53, L54, K56, and E57, all belonging to the acidic loop-PSAALRE helix sequence. In particular, residues E42 and R50 appear to be critical since they interact directly with p5 through hydrogen bonds (cf. Section II.5). Other interfacial residues (green) belong to the larger lobe, namely R120, N121, L123, and R125; and yet others (yellow) to the activation loop. Mutations of these interfacial residues are expected to change either the binding affinity, the selectivity, or the inhibitory activity of the complex in the presence of p5, and should then be the main focus of systematic mutagenesis studies to experimentally map the binding site. Mutations of CDK5/p5 interfacial residues of the other modes (listed in the caption of Fig. 3) would probe other possible inhibitory mechanisms, such as allosteric modulation of the kinase dynamics with potentially deleterious effect on ATP or substrate binding, as previously conjectured from results of dynamic simulations [35].

II.5 Characterization of the pharmacophore

The six independent dynamic simulations of the complex in the inhibitory mode were considered to propose a pharmacophore. Figure 4 shows the molecular-surface of a representative CDK5/p5 interface on both molecules taken from the equilibrated complex. Inspection of the surface electric field suggests strong polar interactions, indicated by circles on the molecules: two regions of negative potential on the surface of p5, formed mainly by the unprotonated carboxyl groups of the C-terminus (circle A) and the side chain of Glu2 (circle E), interact with regions of positive potential on the surface of CDK5, which involves the PSAALRE helix and the end portion of the activation loop. In addition, a region of positive potential on p5 formed mainly by the side chain of Lys1 (circle K), interact with a negative region on the surface of CDK5 formed by the acidic loop. A potentially relevant hydrophobic interaction (dashed circle) is also observed, which involves the same nonpolar region of the PSAARLE helix that stabilizes the CDK5-p25 complex. Equally important are the pattern of hydrogen bonds: one donor (–NH₃⁺ of the Lys1 side chain) and one acceptor (–O₂⁻ of the C-terminus and the Glu2 side chain) group in p5 interact with a reduced set of residues on the CDK5 surface (cf. caption of Fig. 4). To quantify these observations statistically, the hydrogen bonds and the contacts between the polar and between the non-polar moieties of p5 and CDK5 were calculated over the last 20 ns of the trajectories. In addition, the structural stability of p5 was confirmed through several internal RMSD measurements. Figure 4 (right most panels) shows the lengths a, b, and c in the proposed pharmacophore (lower middle panel) and the contact distances between the polar regions as a function of time in a typical simulation (see caption for details); the time evolution of the hydrophobic contacts is also shown (1 indicates hydrophobic interaction; 0 otherwise). These results are well reproduced in four out of the six simulations, thus statistically robust. Of the two outlier simulations, one still shows the same geometric and electrostatic pattern of the proposed pharmacophore but with longer sides a and c, whereas the second outlier shows larger structural distortions. These differences reflect the non-specific nature of p5 binding to CDK5, and the pharmacophore proposed here is expected to render the binding specific. Although the hydrophobic contact is observed in all the simulations, it is not included in the pharmacophore since the electrostatic and the H-bond interactions appears to be strong enough for the stabilization of a prospective drug; other pharmacophores could, however, be proposed with a more hydrophobic character. Figure 5 shows schematically the pharmacophore in its three dimensional context, with an atomic representation of relevant interactions. Reproducing the pattern of interactions proposed in this pharmacophore appears to be feasible within the requirements of the Lipinski’s guidelines (or variants thereof) for improved drug likeness, so the potential exists for replacing p5 (and TFP5) by a small, drug-like molecule that could elicit similar inhibitory effect on CDK5-p25 hyperactivity, thereby reversing β-amyloid deposition in AD.

Figure 4.

Electrostatic representation on the molecular surface of the CDK5/p5 interface of the inhibitory mode (negative potential: red; positive: blue) indicating the regions of strong polar interactions inferred from the simulations (orientation similar to middle panel of Fig. 3). The view is obtained by rotating p5 in the direction indicated by the arrow. The main regions of electrostatic complementarity are denoted with solid circles A, E and K; a potentially relevant hydrophobic contact is shown with a dashed circle. A schematic representation of the main characteristics of the proposed pharmacophore is shown, in the same orientation as in p5. This pattern of interactions is reproduced in four out of six independent dynamics simulations; the average length among the four simulations are in the ranges (in Å) 10 < a < 14, 7 < b < 9, and 6 < c < 8; the right upper panel shows the time evolution in one of the four trajectories. The CO₂⁻ and NH₃⁺ groups form persistent HB with CDK5 residues R50, R125, R149, R156, E42 and S159. The right middle panel shows the corresponding HB distances; the larger variations in K (NH₃⁺) are due to the mobility of the lysine side chain, which transiently binds CDK5 by a water bridge; a more rigid molecule with the characteristics of the proposed pharmacophore should decrease these fluctuations. The frequency of hydrophobic contacts (dashed circles) is illustrated in the plot at the right lower panel (1, hydrophobic interaction; 0, otherwise).

Figure 5.

Representation of the relative position of the pharmacophore of Fig. 4, showing p5 residues A, E and K hydrogen bonded with the corresponding partner residues in the kinase (obtained from a snapshot at the end of one of the simulation for schematic purpose). Only residues interacting with p5 are shown in atomic detail. Due to conformational fluctuations, several other residues are found to transiently bind to p5, with statistically meaningful frequencies, as indicated in the caption of Fig. 4 (see text).

IV. Conclusions

The inhibitory mechanism of kinase CDK5 by the peptide p5 was investigated by computer simulations. A pharmacophore was described that can be used for the design of a small peptide-mimetic drug-like compound to target CDK5 and selectively inhibit the pathological CDK5-p25 hyperactivity. The study builds upon a series of recent developments on ab initio complex-structure prediction and force field optimization that allows detection and characterization of non-specific, sparsely populated binding modes of multi-conformer binding partners. The predicted binding modes suggest that p5 is a non-selective competitor of p25, and imply that both the physiological and the pathological activities are similarly inhibited in the absence of other proteins in the medium. This is consistent with available experimental data in vitro. The predicted binding site is located at the CDK5/p25 interface, and in the vicinity of three critical structural motifs implicated in p25 and p35 binding and in CDK5 activation and function, namely, the PSAALRE helix, the acidic loop, and the activation loop. It was shown that binding is non-specific, with a (pre-relaxation) affinity estimated in the low μM range. The proposed pharmacophore is expected to make binding of a drug candidate more specific and increase its affinity. It was shown that enthalpy is the main contribution to binding (~70%), whereas configurational contributions to the binding entropy are small and unfavorable (~10%). Significant structural relaxations were observed upon association. Electrostatic and hydrogen-bond interactions are the main driving force to association and stability of the complex, whereas hydrophobicity seems to play a more modest role. Reproducing the predicted pattern of electrostatic/hydrogen-bond interactions may be feasible within the general guidelines for improved druglikeness, so p5 and the longer peptide TFP5 can potentially be replaced by a small drug-like molecule that elicits similar inhibitory effect on CDK5 hyperactivity, thereby reversing β-amyloid deposition in AD. The current study provides the basis for systematic structure-activity studies.