The flexible pocketome engine for structural chemogenomics

Abstract

Biological metabolites, substrates, cofactors, chemical probes, and drugs bind to flexible pockets in multiple biological macromolecules to exert their biological effect. The rapid growth of the structural databases and sequence data, including SNPs and disease-related genome modifications, complemented by the new cutting-edge 3D docking, scoring, and profiling methods has created a unique opportunity to develop a comprehensive structural map of interactions between any small molecule and biopolymers. Here we demonstrate that a comprehensive structural genomics engine can be built using multiple pocket conformations, experimentally determined or generated with a variety of modeling methods, and new efficient ensemble docking algorithms. In contrast to traditional ligand-activity-based engines trained on known chemical structures and their activities, the structural pocketome and docking engine will allow prediction of poses and activities for new, previously unknown, protein binding sites, and new, previously uncharacterized, chemical scaffolds. This de novo structure-based activity prediction engine may dramatically accelerate the discovery of potent and specific therapeutics with reduced side effects.

Figure 1

A general representation complete of chemogenomics matrix. Each column, P1, P2, … represent a conformational ensemble of a protein pocket. Different functional states (e.g., agonist bound and antagonist bound) and different locations on the same protein are considered separate pockets. SNPs and mutations may lead to variations of the same pocket. Each row represents a chemical compound. The chemicals are metabolic compounds, drug candidates and other chemical substances that are relevant for a biological system, including virtual compounds that have never been synthesized. The goal of this structural chemogenomics engine is to report, if experimental data is available, or predict the following: (i) the binding geometry of each compound to the pockets it can bind, and (ii) an estimate of the binding free energy, e_ij. While the screening application searching for potential binders among virtual or available chemicals is widely used, comparing e_ij for the same compound with different pockets (or proteins), a.k.a. specificity profiling, requires new approaches.

Figure 2

Multi-conformational Pocketome unit. Eight alternative conformers of MDM2 from apo and co-crystal structures are superimposed. The transparent surface represents the location of the known ligands.

Figure 3

A histogram of experimental structural variability of the 11,168 protein domains in the PDB. 25% of protein domains are represented by a single structure, and 5% are represented by more than 30 structures. Three quarters of the domains are represented by more than one conformation. The additional conformers are found in either different PDB entries or non-crystallographic symmetry related domains of the same entry.

Figure 4

Four levels of reliability of ligand positioning into the electron density. The coordinates are taken from the PDB and the density obtained from the Uppsala EDS server. Only the pose of the first (leftmost) ligand is unambiguously defined by the electron density. The last ligand represents a complete fantasy. More than a third of the ligands in pockets in the PDB need to be either ignored or re-positioned.

Figure 5

(a) Energy-based optimization of the ligand and the pocket side-chains often leads to a more energetically favorable conformation and improved electron density fit. (a) Unrestrained sampling of hydroxyl groups of β2AR Ser203 and Ser204 in the recently solved X-ray structure (PDB ID 2rh1) lead to improved energetics while preserving the electron density fit. (b) Effect of local heavy atom energy refinement / redocking on the pose and interactions of the pregnane X receptor bound to SRL (PDB ID 1nrl). Performed without any influence of the electron density, the ICM optimization shifted the ligand by 1.3 A and found a pose with better binding interactions and better fit to the electron density.

Figure 6

Flexibility of small molecule binding interfaces and induced fit. About one fifth of interface side-chains are displaced by more than 1.5 A when compared between different complex compositions. At least one interface residue backbone deviates by more than 1.5A in 33% of the cases, at least one side-chain – in 77% of the cases.

Figure 7

Pocketome entry for the kelch-like ECH-associated protein 1 (KEAP1). Four superimposed X-ray structures and the ICM PocketFinder envelope are shown. This protein was unsuccessfully targeted by a small-molecule inhibitor at Merck. The Pocketome analysis demonstrates that the pocket is too small and too flexible for a strong small molecule binder.

Figure 8

Pocket fumigation is a modeling technique based on torsional sampling in the presence of a repulsive density representing a generic ligand. (a) the original X-ray structure; (b) the result of Ala conversion: the “largest pocket” density is generated; (c) a “druggable” pocket conformation obtained by Monte Carlo simulation in the presence of the density. Coloring of the residues indicates the degree of their intrusion into the density (blue – low, red – high).

Figure 9

The accuracy of the ligand binding pose prediction for different ensemble sizes. The bars reflect the fraction of the ligands that dock correctly using traditional ensemble docking (orange) and 4D docking (blue) for varying ensemble size, compared to the accuracy of a single-receptor cross-docking (yellow).

Figure 10

Unlike the traditional ensemble docking, the 4D protocol docks the ligand into a set of receptor conformations in a single docking run.

Figure 11

The SCARE (SCan Alanines and REfine) cross-docking protocol produces a series of omission models by simultaneously mutating to alanines every pair of neighboring residues in the binding pocket. The ligand docking to the omitted models is followed by the refinement stage at which the omitted side-chains are rebuilt. The protocol successfully reproduces the ligand binding pose in ~~90% of the cases (compared to the 46.6% performance of the single rigid receptor cross docking).

Figure 12

Screening and selectivity profiling for the DOLPHIN models of DFG-out (inactive) states of kinases as an example a set of multi-conformer pocketome units for a group of related proteins in a certain functional state. Panel A shows the pose prediction by two models, Panel B shows the ligand screening benchmark by multiple conformers for the inactive state of MK14 kinase, and Panel C shows a comparison between predicted and experimental binding energies using the kinase-specific offset technique.