The LigTMap server provides a fully automated workflow to identify targets from 17 target classes with >6000 proteins. It is a hybrid approach, combining ligand similarity search with docking and binding similarity analysis, to predict putative targets. Besides its outstanding performance, LigTMap provides straightaway the PDB of a predicted target, the optimal ligand binding mode, and the protein-ligand interaction profile (PLIP). These were designed to meet the needs of structure-based drug design and to shed lights on the function of the ligand.
The prediction workflow is illustrated in the following figure. It consists of five steps:
1. For a query compound, a set of potential targets is selected based on fingerprint similarities to the crystal ligands. Multiple fingerprints (Morgan, MACCS, Daylight, MMD) are generated, and the ligand similarity score (TL) is taken as an average of either three MMD or two MM fingerprint Tanimoto coefficients, depending on whether the run is for all-target class prediction or class-specific prediction. A cutoff value for TL is used as the target selection criteria; the default is 0.4.
2. For each potential target, molecular docking is performed using PSOVina2 to predict the most optimal binding mode of the compound in the ligand-binding pocket.
3. A binding interaction fingerprint (IFP) of the compound is generated based on the predicted binding mode. The established IFP is compared with the IFP of the crystal ligand using the Tanimoto coefficient to obtain the binding similarity score (TB).
4. For each potential target class, the compound binding activity is predicted using the class-specific random forest (RF) model based on the Avalon fingerprint.
5. Finally, all prediction results are consolidated, and the protein targets are ranked based on the combined score LigTMap score = 0.7 TL + 0.3 TB.
In the validation and benchmark experiments, LigTMap achieved a top-10 success rate of about 70%, with an average precision rate of 0.34 and recall of 0.26. The class-specific prediction method improved the success rate further to 80% with enhanced precision to 0.54 and recall to 0.52. The benchmark result showed that LigTMap compares favorably to the state-of-the-arts prediction servers.
Source CodeThe LigTMap method and its prediction workflow can be setup for local use. The code is freely available under the BSD 3-Clause open source license at https://github.com/ShirleyWISiu/LigTMap.
Copyright © 2021 University of Macau