Identifying reliable binding sites based on three dimensional structures of proteins and other macromolecules is a key step in drug discovery. A good definition of known binding site and the detection of a novel site can provide valuable information for drug design efforts. CAVITY is developed for the detection and analysis of ligand-binding site(s) . It has the capability of detecting potential binding site as well as estimating both the ligandabilities and druggabilites of the detected binding sites.

CAVITY was originally used in the de novo drug design tool LigBuilder 2.0 to accurately reflect the key interactions within a binding site as well as to confine the ligand growth within a reasonable region; it was later developed into a stand-alone program for binding site detection and analysis. The CAVITY approach generates clear and accurate information about the shapes and boundaries of the ligand binding sites, which provide helpful information for drug discovery studies: 1) For cases where a protein-ligand complex of the target protein is available, CAVITY can be used to detect the binding site regions which are not covered by the known ligand(s) and provide clues for the improvement of ligand-binding affinity. In addition, the predicted ligandability and druggability of the binding site would tell the researchers whether further improvement of the known ligand is promising. 2) For cases where ligands are known, but the structural information of ligand-target interactions is not available, CAVITY can be used to detect the binding site and the binding mode of the known ligands could be predicted by using molecular docking technique. 3) For cases with no reported ligand, CAVITY can not only be used to detect potential binding sites, but also to provide qualitative estimations of ligandability and druggability for potential binding sites on the target protein, which is very important for making an early stage decision about whether the protein is a promising target for a drug discovery project. CAVITY has been used in many different projects to help generate such information and clues. We used the external NRDLD data set and Cheng’s data set to test Cavity, showing a satisfactory performance of 0.82 and 0.89 accuracy, respectively.

Schematic diagram of cavity detection in CAVITY. a) Protein (black-colored) in grid box (green-colored). b) Using the eraser ball to remove grid points outside protein. c) “Vacant” grid points after erasing. Four cavities were shown in different colors. d) Shrink each cavity until the depth reach the minimal depth. e) Recover cavities to obtain the final result.

Pharmacophores derived from three-dimensional structures of specific protein binding sites can provide useful information for analyzing protein-ligand interactions, which therefore will help guide computational drug design. Pharmacophore features and their spatial arrangements are usually used to describe a pharmacophore model. CavPharmer uses a receptor-based pharmacophore modeling program Pocket2 to automatically extract pharmacophore features within cavities. CavPharmer output 7 feature types as hydrophobic center, hydrogen bond donor, hydrogen bond acceptor, positive-charged center, negative-charged center, aromatic center and exclude volume, to make a pharmacophore model from a protein structure integral. Key features in the pharmacophore model are automatically reduced to a reasonable number. We used data from DUD database to test the efficiency of CavPharmer and a ligand-based pharmacophore modeling software LigandScout V2.02. Results show that for receptor-based pharmacophore modeling, CavPharmer outperforms LigandScout. (average AUC value 0.69 versus 0.63 in 38 cases) Overall, CavPharmer is an accurate pharmacophore modeling software and can be applied widely into different drug design cases.

CavPharmer can also identify hot spots in protein-protein interface using only an apo protein structure. Given similarities and differences between the essence of pharmacophore and hot spots for guiding design of chemical compounds, not only energetic but also spatial properties of protein-protein interface are used in CavPharmer for dealing with protein-protein interface.

CavPharmer has been applied to many studies and well reproduced previously published pharmacophore models in these cases. One notable feature of CavPharmer is that it can tolerate minor conformational changes on the protein side upon binding of different ligands to give a consistent pharmacophore model. For different proteins accommodating the same ligand, CavPharmer gives similar pharmacophore models, which opens the possibility to classify proteins with their binding features. The Pharmacophore models used in Pharmmapper 2017 server (http://lilab.ecust.edu.cn/pharmmapper) were generated using CavPharmer method.

The concept of allostery was proposed in early 1960’s, which mainly refers to the fact that the function of an orthosteric site can be affected by a ligand bound to an allosteric site that is topographically different from the orthosteric site. Allostery regulates numerous biological processes, and the dysregulation of allosteric communication networks between allosteric and orthosteric sites can lead to various human diseases. Compared to drugs targeting orthosteric sites, drugs targeting allosteric sites have higher specificity, fewer side effects and a variety of regulatory types. However, discovery of allosteric drugs remains challenging, as it is still difficult to accurately predict allosteric sites and to elucidate the mechanism of allosteric regulation.

CorrSite2.0 is a new method for predicting allosteric sites, which ranks the potential ligand binding sites based on calculated motion correlations using the slow and fast modes. The program first uses CAVITY to find all the potential ligand binding pockets on the surface of the protein. Then it selects an orthosteric pocket from the pockets found by CAVITY or imports an orthosteric pocket from the outside, such as an orthosteric pocket defined by the residues within 4Å around the orthosteric ligand. Next, the program calculates the correlations and Z-scores between the orthosteric pocket and the remaining pockets excluding the orthosteric sites in the top 10 fast modes and the top 3 slow modes, respectively. Finally, it takes the maximum of the two Z-Scores corresponding to each pocket, and ranks these pockets according to Z-Score. Pockets with Z-Score greater than 0.5 are predicted as potential allosteric sites.

Compared with CorrSite1.0, CorrSite2.0 has a great improvement in the applicability and prediction accuracy. First of all, CorrSite1.0 can only use one chain for prediction, while CorrSite2.0 can not only use one chain for prediction, but also can use multiple chains for prediction. Secondly, the prediction accuracy of CorrSite2.0 has been greatly improved. In the CorrSite2.0 dataset I containing 36 allosteric sites, CorrSite2.0 and CorrSite1.0 could successfully predict 35 and 25 of them, respectively. In the independent test set CorrSite2.0 dataset II, CorrSite2.0 and CorrSite1.0 could successfully predict 18 and 12 of 20 allosteric sites, respectively.

CovCys is developed based on a comprehensive statistical analysis on covalent modified cysteine residues in protein structures. Compared to unmodified cysteine residues in the same protein structure, the covalently modified cysteine residues showed lower pKa values and higher solvent exposure. Such cysteine residues are usually located within or near one of many pockets detected by a cavity-detection program.

The current application of computational procedure for targetable cysteine prediction only requires the protein three-dimensional structure coordinates. A number of descriptors will be calculated based on the structure, including surface pockets detected by using CAVITY, pKa by using PROPKA, SASA by using Pops as well as adjacent amino acid compositions calculated based on ProODY package. The generated descriptors will be used as the input for a pre-trained SVM model to predict whether a cysteine residues are suitable for druglike covalent ligand design. The external validation data set (positive/negative: 1377/5185) from Cysteinome database was tested by CovCys with prediction accuracy of 0.73.

Currently, if a cysteine is not within a pocket, it will not be used to make prediction. Such cysteine maybe an active cysteine, but it is hard to design a proper ligand without a binding cavity. Meanwhile, if the pKa calculation failed, it is also not considered. Such cysteine could be in a form of di-sulfur bond or inside the protein.

To use CovCys, please upload your protein structure information in the computing page. The output results contains all cysteine residues within the protein and their probability to be a druggable covalent reside. The pKa, percentage of exposure and the average pKd value of the associated pocket are also reported.