Read more in D. Trapl, I. Horvaćanin, V. Mareška, F. Özçelik, G. Unal and V. Spiwok: anncolvar: Approximation of Complex Collective Variables by Artificial Neural Networks for Analysis and Biasing of Molecular Simulations Front. Mol. Biosci. 2019, 6, 25 (doi: 10.3389/fmolb.2019.00025)
Collective variables by artificial neural networks:
usage: anncolvar [-h] [-i INFILE] [-p INTOP] [-c COLVAR] [-col COL] [-boxx BOXX] [-boxy BOXY] [-boxz BOXZ] [-nofit NOFIT] [-testset TESTSET] [-shuffle SHUFFLE] [-layers LAYERS] [-layer1 LAYER1] [-layer2 LAYER2] [-layer3 LAYER3] [-actfun1 ACTFUN1] [-actfun2 ACTFUN2] [-actfun3 ACTFUN3] [-optim OPTIM] [-loss LOSS] [-epochs EPOCHS] [-batch BATCH] [-o OFILE] [-model MODELFILE] [-plumed PLUMEDFILE] Artificial neural network learning of collective variables of molecular systems, requires numpy, keras and mdtraj optional arguments: -h, --help show this help message and exit -i INFILE Input trajectory in pdb, xtc, trr, dcd, netcdf or mdcrd, WARNING: the trajectory must be 1. must contain only atoms to be analyzed, 2. must not contain any periodic boundary condition issues! -p INTOP Input topology in pdb, WARNING: the structure must be 1. centered in the PBC box and 2. must contain only atoms to be analyzed! -c COLVAR Input collective variable file in text format, must contain the same number of lines as frames in the trajectory -col COL The index of the column containing collective variables in the input collective variable file -boxx BOXX Size of x coordinate of PBC box (from 0 to set value in nm) -boxy BOXY Size of y coordinate of PBC box (from 0 to set value in nm) -boxz BOXZ Size of z coordinate of PBC box (from 0 to set value in nm) -nofit NOFIT Disable fitting, the trajectory must be properly fited (default False) -testset TESTSET Size of test set (fraction of the trajectory, default = 0.1) -shuffle SHUFFLE Shuffle trajectory frames to obtain training and test set (default True) -layers LAYERS Number of hidden layers (allowed values 1-3, default = 1) -layer1 LAYER1 Number of neurons in the first encoding layer (default = 256) -layer2 LAYER2 Number of neurons in the second encoding layer (default = 256) -layer3 LAYER3 Number of neurons in the third encoding layer (default = 256) -actfun1 ACTFUN1 Activation function of the first layer (default = sigmoid, for options see keras documentation) -actfun2 ACTFUN2 Activation function of the second layer (default = linear, for options see keras documentation) -actfun3 ACTFUN3 Activation function of the third layer (default = linear, for options see keras documentation) -optim OPTIM Optimizer (default = adam, for options see keras documentation) -loss LOSS Loss function (default = mean_squared_error, for options see keras documentation) -epochs EPOCHS Number of epochs (default = 100, >1000 may be necessary for real life applications) -batch BATCH Batch size (0 = no batches, default = 256) -o OFILE Output file with original and approximated collective variables (txt, default = no output) -model MODELFILE Prefix for output model files (experimental, default = no output) -plumed PLUMEDFILE Output file for Plumed (default = plumed.dat)
Biased simulations, such as metadynamics, use a predefined set of parameters known as collective variables. An artificial bias force is applied on collective variables to enhance sampling. There are two conditions for a parameter to be applied as a collective variable. First, the value of the collective variables can be calculated solely from atomic coordinates. Second, the force acting on collective variables can be converted to the force acting on individual atoms. In the other words, it is possible to calculate the first derivative of the collective variables with respect to atomic coordinates. Both calculations must be fast enough, because they must be evaluated in every step of the simulation.
There are many potential collective variables that cannot be easily calculated. It is possible to calculate the collective variable for hundreds or thousands of structures, but not for millions of structures (which is necessary for nanosecond long simulations). anncolvar can approximate such collective variables using a neural network.
You have to chose and install one of keras backends, such as Tensorflow, Theano or CNTK. For this follow one of these links:
Install numpy and cython by PIP:
pip install numpy cython
Next, install anncolvar by PIP:
pip install anncolvar
If you use Anaconda type:
conda install -c spiwokv anncolvar
A series of representative structures (hundreds or more) with pre-calculated values of the collective variable is used to train the neural network. The user can specify the input set of reference structures (-i) in the form of a trajectory in pdb, xtc, trr, dcd, netcdf or mdcrd. The trajectory must contain only atoms to be analyzed (for example only non-hydrogen atoms). The trajectory must not contain any periodic boundary condition issues. Both conversions can be made by molecular dynamics simulation packages, for example by gmx trjconv. It is not necessary to fit frames to a reference structure. It is possible to switch fitting off by -nofit True.
It is necessary to supply an input topology in PDB. This is a structure used as a template for fitting. It is also used to define a box. This box must be large enough to fit the molecule in all frames of the trajectory. It should not be too large because this suppresses non-linearity in the neural network. When the user decides to use a 3x3x3 nm box it is necessary to place the molecule to be centered at coordinates (1.5,1.5,1.5) nm. In Gromacs it is possible to use:
gmx editconf -f mol.pdb -o reference.pdb -c -box 3 3 3
It must also contain only atoms to be analyzed. Size of the box can be specified by parameters -boxx, -boxy and -boxz (in nm).
Last input file is the collective variable file. It is a space-separated text file with the same number of lines as the number of frames in the input trajectory. The index of the column can be specified by -col (e.g. -col 2 for the second column of the file.
The option -testset can control the fraction of the trajectory used as the test set. For example -testset 0.1 means that 10 % of input data is used as the test set and 90 % as the training set. The option -shuffle True causes that first 90 % is used as the training set and remaining 10 % as the test set. Otherwise frames are shuffled before separation to the training and test set.
The architecture of the neural network is controlled by multiple parameters. The input layer contains 3N neurons (where N is the number of atoms). The number of hidden layers is controlled by -layers. This can be 1, 2 or 3. For higher number of layers contact the authors. Number of neurons in the first, second and third layer is controlled by -layer1, -layer2 and -layer3. It is useful to use the number of layers equal to powers of 2 (32, 64, 128 etc.). Huge numbers of neurons can cause that the program is slow or run out of memory. Activation functions of neurons can be controlled by -actfun1, -actfun2 and -actfun3. Any activation function supported by keras can be used.
The optimizer used in the training process can be controlled by -optim. The default ADAM optimizer (-optim adam) works well. The loss function can be controlled by -loss. The default -loss mean_squared_error works well. The number of epochs can be controlled by -epochs. The default value (100) is quite little, usually >1000 is necessary for real life applications. The batch size can be controlled by -batch (-batch 0 for no batches, default is 256).
Output is written into the text file -o. It contains the approximated and the original values of collective variable. The model can be stored in the set of text files (try -model). The input file is printed into the file controlled by -plumed (by default plumed.dat). This file can be directly used to calculate the evolution of the collective variable by plumed driver or by Plumed-patched molecular dynamics engine. To use the collective variable in enhances sampling (for example metadynamics) it is necessary to add a suitable keyword (for example METAD).