NNScore

As high-throughput biochemical screens are both expensive and labor intensive, researchers in academia and industry are turning increasingly to virtual-screening methodologies. Virtual screening relies on scoring functions to quickly assess ligand potency. Unfortunately, these scoring functions generally have many false positives and negatives; indeed, a properly trained human being can often assess ligand potency by visual inspection with greater accuracy.

Given the success of the human mind at receptor-ligand complex characterization, we here present two scoring functions based on neural networks, computational models that simulate the microscopic organization of the brain. Computer-aided drug design depends on fast and accurate scoring functions to aid in the identification of small-molecule ligands. The scoring functions presented here, either on their own or used in conjunction with other more traditional functions, may prove useful in drug-discovery projects. Additional information about NNScore 1.0 can be found in the original paper. NNScore 2.0 is described in a separate publication. {LINK, REF}

Note that NNScore 2.0 is not necessarily superior to NNScore 1.0. The best scoring function to use is highly system dependent. Including positive controls (known inhibitors) in virtual screens is a useful way to identify which scoring function is best suited to your needs.

If you use NNScore in your research, please cite the appropriate reference:

NNScore: A Neural-Network-Based Scoring Function for the Characterization of Protein-Ligand Complexes. Jacob D. Durrant, J. Andrew McCammon. Journal of Chemical Information and Modeling, 2010, 50 (10), pp 1865-1871.

{REFERENCE FOR NNSCORE 2.0 HERE}

NNScore 1.0 has been implemented as a python script. The program accepts the following parameters:
     -receptor <pdbqt filename>
     -ligand <pdbqt filename>
     -network <network filename>
     -networks_dir <directory>

Note: It is best to use multiple neural networks to judge ligand binding by consensus. Commandline parameters can be used to add neural-network files to the list of those that will be used. To add a single neural network to the list, use the -network parameter to specify a single network file. To add mutliple networks to the list, create a directory containing only network files and specify the path to that directory using the -networks_dir parameter.

Note: Only pdbqt files of the receptor and ligand are accepted. Scripts to convert from pdb to pdbqt are included in the AutoDockTools package.

Examples:
     python NeuroScore.py -receptor neuraminidase.pdbqt
          -ligand oseltamivir.pdbqt
          -network ./networks/top_3_networks/12.net
     python NeuroScore.py -receptor integrase.pdbqt
          -ligand raltegravir.pdbqt
          -networks_dir ./networks/top_3_networks/
     python NeuroScore.py -receptor protease.pdbqt
          -ligand tipranavir.pdbqt
          -networks_dir ./networks/top_24_networks/
          -network ./networks/top_3_networks/16.net

NNScore 2.0 has also been implemented as a python script. As demonstrated in our paper {TITLE HERE}, NNScore 2.0 is not necessarily superior to NNScore 1.0. The best scoring function to use is highly system dependent. Including positive controls (known inhibitors) in virtual screens is a useful way to identify which scoring function is best suited to your needs.

REQUIREMENTS

Python: NNScore 2.0 has been tested using Python versions 2.6.5, 2.6.1, and 2.5.2 on Ubuntu 10.04.1 LTS, Mac OS X 10.6.8, and Windows XP Professional, respectively. A copy of the Python interpreter can be downloaded from http://www.python.org/getit/.

AutoDock Vina 1.1.2: NNScore 2.0 uses AutoDock Vina 1.1.2 to obtain some information about the receptor-ligand complex. Note that previous versions of AutoDock Vina are not suitble. AutoDock Vina 1.1.2 can be downloaded from http://vina.scripps.edu/download.html.

MGLTools: As receptor and ligand inputs, NNScore 2.0 accepts models in the PDBQT format. Files in the more common PDB format can be converted to the PDBQT format using scripts included in MGLTools (prepare_receptor4.py and prepare_ligand4.py). MGLTools can be obtained from http://mgltools.scripps.edu/downloads.

COMMAND-LINE PARAMETERS

PROGRAM OUTPUT

NNScore 2.0 evaluates each of the ligand poses contained in the file specified by the -ligand tag using 20 distinct neural-network scoring functions. The program then seeks to identify which of the poses has the highest predicted affinity using several metrics:

1) Each of the 20 networks are considered separately. The poses are ranked in 20 different ways by the scores assigned by each network.

2) The poses are ranked by the best score given by any of the 20 networks.

3) The poses are ranked by the average of the scores given by the 20 networks. This is the recommended metric.

EXAMPLE OF USAGE

python NNScore2.py -receptor myreceptor.pdbqt -ligand myligand.pdbqt -vina_executable /PATH/TO/VINA/1.1.2/vina

All versions of NNScore are released under the GNU General Public License. Your use of NNScore implies acceptance of the terms stipulated in that license.

Download any version of NNScore from sourceforge.net

If you have any questions, comments, or suggestions, please don't hesitate to contact me, Jacob Durrant, at jdurrant [at] ucsd [dot] edu. I'd be happy to help. :)