PySFD - Significant Feature Differences Analyzer for Python

PySFD computes and visualizes any significant differences in features, e.g., distances, contacts, angles, dihedrals, among different sets of molecular dynamics-simulated ensembles.

For an overview, pleas read the following documentation. For further details, please read

• the doc strings throughout the PySFD package
• example jupyter notebook and python files located in PySFD/docs/PySFD_example

Citation

If you find this tool useful, please cite:

Stolzenberg, S. "PySFD: Comprehensive Molecular Insights from Significant Feature Differences detected among many Simulated Ensembles." Bioinformatics (Oxford, England) (2018).

https://doi.org/10.1093/bioinformatics/bty818

Installation

PySFD is a Python package available for Python 2 and 3, should in principle run on all common platforms, but is currently only supported for Linux and MacOS.

Required Packages

With conda installed, e.g., via

Miniconda https://conda.io/miniconda.html

first make sure you have activated the conda-forge channel:

conda config --add channels conda-forge

and then install the folllowing python packages:

conda install jupyter numpy pathos pandas biopandas mdtraj scipy matplotlib seaborn

to install SHIFTX2 (only works for Python 3):

conda install -c omnia/label/dev shiftx2

For visualization of significant feature differences, please install the latest versions of:

• PyMOL (https://pymol.org), e.g., via

conda install -c schrodinger pymol
conda install Pmw

• VMD (http://www.ks.uiuc.edu/Research/vmd)

In case you would like to visualize your PySFD results directly from a Python/iPython/Jupyter session (via pysfd.view_feature_diffs()), you will need to install the Python module iPyMol (https://github.com/cxhernandez/ipymol.git) via:

pip install ipymol

Working Package System

PySFD was successfully tested, e.g., using the following python environment:

package	version		channel
biopandas	0.2.3	py36_0	conda-forge
ipymol	0.5	<pip>
jupyter	1.0.0	py_1	conda-forge
matplotlib	2.2.2	py36_1	conda-forge
mdtraj	1.9.1	py36_1	conda-forge
numpy	1.14.5	py36hcd700cb_0
pandas	0.23.1	py36_0	conda-forge
pathos	0.2.1	py36_1	conda-forge
pymol	2.1.1	py36_2	schrodinger
python	3.6.5	1	conda-forge
scipy	1.1.0	py36hfc37229_0
seaborn	0.8.1	py36_0	conda-forge

(created via conda list "^python$|^jupyter$|numpy|pathos|pandas|biopandas|mdtraj|scipy|matplotlib|seaborn|shiftx2|pymol|ipymol" )

The full conda environment is listed in ./pysfd.yaml

(created via conda env export > pysfd.yaml)

, which can be used to automatically install all required packages via:

conda env create -f pysfd.yaml

PySFD

Then just download the PySFD package, e.g., via GitHub

git clone https://github.com/markovmodel/PySFD.git

and from within the downloaded "PySFD" directory, install PySFD into the Python Path of your environment via:

python setup.py install

To uninstall, simply type:

pip uninstall pysfd

Features

Pre-defined features are stored in pysfd.features and currently include the following modules and feature classes:

srf.py : Single Residue Feature (SRF)

• CA_RMSF_VMD: CA atom root mean square flucatuations (RMSF), computed with VMD
• ChemicalShift: predicted NMR Chemical Shifts, computed via mdtraj and shiftx2
• Dihedral: dihedral angles computed with mdtraj, one of the following methods is passed to the Dihedral class:
  • mdtraj.compute_chi1
  • mdtraj.compute_chi2
  • mdtraj.compute_chi3
  • mdtraj.compute_chi4
  • mdtraj.compute_omega
  • mdtraj.compute_phi
  • mdtraj.compute_psi
• IsDSSP_mdtraj : binary secondary structure assignments (DSSP), e.g., whether or not a helix ("H") is formed in a particular residue
• Scalar_Coupling: scalar couplings computed with mdtraj, one of the following methods is passed to the Scalar_Coupling class:
  • mdtraj.compute_J3_HN_C
  • mdtraj.compute_J3_HN_CB
  • mdtraj.compute_J3_HN_HA
• SASA_sr : solvent accessibility surface areas (SASAs) via mdtraj.shrake_rupley
• RSASA_sr : relative SASA, i.e. normalized to the total SASA of a particular residue, computed via mdtraj.shrake_rupley

paf.py : Pairwise Atomic Features (PAF)

• Atm2Atm_Distance: atom-to-atom distance (CA atoms by default, i.e. if df_sel is None)
• AtmPos_Correlation: (partial) correlations between atom positions (CA atoms by default, i.e. if df_sel is None)

prf.py : Pairwise Residual Features (PRF)

• Ca2Ca_Distance: distance between CA atoms
• CaPos_Correlation: (partial) correlations between Ca positions
• Dihedral_Correlation: (partial) correlations between Dihedral angles (see srf module)
• Scalar_Coupling_Correlation: (partial) correlations between Dihedral angles (see srf module)

spbsf.py : sparse Pairwise Backbone Sidechain Features (sPBSF) (contact frequencies and dwell times)

• HBond_mdtraj: hydrogen bonds via mdtraj
• HBond_VMD: hydrogen bonds via VMD
• HBond_HBPLUS: hydrogen bonds via HBPLUS
  • link to HBPLUS program: https://www.ebi.ac.uk/thornton-srv/software/HBPLUS/
  • after installation, please update "export hbdir=..." in PySFD/features/scripts/compute_PI.hbplus.sh
• Hvvdwdist_VMD: heavy atom van der Waals radii contacts between residues that are > 4 residue positions apart, computed via VMD
• HvvdwHB: contact, if Hvvdwdist_VMD or HBond_HBPLUS contact in a simulation frame Hvvdwdist_VMD contacts can be transformed into a single residual feature (SRF, see above), e.g., by only considering the contact frequencies of any residual backbone or side-chain with water (Hvvdwdist_H2O).

pprf.py : Pairwise Pairwise Residual Features (PPRF)

• Ca2Ca_Distance_Correlation: pairwise (partial) correlations between Ca-to-Ca distances

pspbsf.py : Pairwise sparse Pairwise Backbone/Sidechain Features (PsPBSF)

• sPBSF_Correlation: (partial) correlations between sPBSF features (see spbsf module)

pff.py : Pairwise Feature Features (PFF)

• Feature_Correlation: pairwise (partial) correlations between (coarse-grained) features, each which are specified for individual frames, e.g. for correlations between srf.Dihedral and spbsf.HBond_mdtraj, but not with srf.CA_RMSF_VMD

Beware of spurious correlations in big data!

Each of these feature classes contained in each module is derived from pysfd.features._feature_agent.FeatureAgent. Each such feature class is instantiated and passed as a FeatureObj object into an pysfd.PySFD object, which in turn extracts the essential feature function from FeatureObj via FeatureObj.get_feature_func(). These features classes can be exented by adding further feature classes in each existing feature module, or adding further feature modules (make sure then to update PySFD/features/__init__.py).

An important functionality of PySFD is that it allows the histogramming of specific (or all) computed features of a particular type. The features to be histogrammed are initially listed in the df_hist_feats parameter, a pandas.DataFrame passed in a particular feature class (see docstrings). For each ensemble, the average feature histograms (with standard deviations) over ensemble trajectories are then computed, and stored into the PySFD.df_fhists dictionary.

Feature Coarse-Graining

Each of these features can be further coarse-grained for each simulation frame by "region" into regional features by providing df_rgn_seg_res_bb, a user-defined pandas Dataframe, and rgn_agg_func, user-defined function or string, to the specific feature class.

df_rgn_seg_res_bb maps from the residual (or sub-residual, i.e. backbone/sidechain) level to the regional level via the columns rgn (region ID), seg (SegID), res (ResID), and optionally bb (is backbone (1)? is sidechain (0)?).

'''
* df_rgn_seg_res_bb : optional pandas.DataFrame for coarse-graining that defines
                      regions by segIDs and resIDs, and optionally backbone/sidechain, e.g.
  df_rgn_seg_res_bb = _pd.DataFrame({'rgn' : ["a1", "a2", "b1", "b2", "c"],
                                     'seg' : ["A", "A", "B", "B", "C"],
                                     'res' : [range(4,83), range(83,185), range(4,95), range(95,191), range(102,121)]})
                      if None, no coarse-graining is performed
'''

rgn_agg_func defines the function used in the coarse-graining, e.g. mean, sum, ...

'''
* rgn_agg_func  : function or str for coarse-graining
                  function that defines how to aggregate from residues (backbone/sidechain) to regions in each frame
                  this function uses the coarse-graining mapping defined in `df_rgn_seg_res_bb`
                  - if a function, it has to be vectorized (i.e. able to be used by, e.g., a 1-dimensional numpy array)
                  - if a string, it has to be readable by the aggregate function of a pandas Data.Frame,
                    such as "mean", "std"
'''

Each feature class then coarse-grains these features

• in each simulation frame, if feature values are computed by frame
• otherwise over feature values, e.g., if the underlying feature type describes an entire trajectory (correlation coefficients, RMSF, ...)

Each feature type has a default for rgn_agg_func (usually "mean" or "sum", see doc strings)

For help on how to semi-automatically define df_rgn_seg_res_bb see:

PySFD/docs/notes/how_to_define_rgn_to_seg_res_ranges.txt

For example definitions of df_rgn_seg_res_bb data frames, see

PySFD/docs/PySFD_example/scripts

Feature and Significant Feature Differences

These are computed by the following methods (see example jupyter notebooks in PySFD/docs/PySFD_example):

1. PySFD.comp_features(),
2. PySFD.comp_feature_diffs(), or PySFD.comp_feature_diffs_with_dwells(), and
3. PySFD.comp_and_write_common_feature_diffs()

(further notes below)

Multiple features types and differences can be computed within the same PySFD instance and are stored, respectively, into the dictionaries PySFD.df_features[PySFD.feature_func_name] and PySFD.df_feature_diffs[PySFD.feature_func_name], where PySFD.feature_func_name is the name of the currently selected feature function PySFD.feature_func.

Further Notes:

1. PySFD.comp_features()

Features are computed as means over ensemble trajectory means (and optionally higher statistical moments with respect to the mean, see parameter max_mom_ord in the docstrings), i.e. each feature as a mean of the feature's mean (or higher statistical moment with respect to the mean) along each trajectory (circular/arithmetic means for circular/linear feature types). The uncertainties of these feature means (and optionally higher moments) can be computed either as

"standard errors" (PySFD.error_type[PySFD.feature_func_name] = "std_err") (statistical uncertainty)

or

"standard deviations" (PySFD.error_type[PySFD.feature_func_name] = "std_dev") ("exclusive"(/"effect size") uncertainty).

(optional higher moments, i.e. max_mom_ord>1, are defined only for "standard errors")

If these uncertainties are computed as "standard errors", then each as a (circular) standard deviation over ensemble trajectory means.

If these uncertainties are computed as "standard deviations", then each as a mean of (circular) standard deviations over ensemble trajectories.

PySFD.error_type[PySFD.feature_func_name] is implicitly defined in each FeatureObj object passed into PySFD.

(

PySFD.error_type[PySFD.feature_func_name] gets updated in the beginning of PySFD.run_ens():

self.error_type[self._feature_func_name] = dataflags.get("error_type")

Alternatively, one could directly define error_type as a universal PySFD.error_type and make PySFD.PyASDA.feat_func() read it, but then looping through a list of feature functions items would require updating self.error_type for every iteration, e.g. via an external dictionary.

)

Reloading of already computed features (not differences) is invoked via PySFD.reload_features() and writing features to disk via PySFD.write_features().

2. PySFD.comp_feature_diffs(), or PySFD.comp_feature_diffs_with_dwells()

To determine significant differences among pairs of ensembles, the individual ensemble uncertainties are added geometrically (to form "sigma", i.e. the sqrt(...) in the manuscript) and scaled by num_sigma. An individual feature is significantly different between two ensembles, if its mean (optionally a higher moment with respect to the mean) differs in absolute value by more than both num_funit and num_sigma * sigma:

(f_a, f_b could either be a mean or optionally a higher statistical moment with respect to the mean)

If significant differences are computed also in regard to higher statistical moments, then significant differences in the m-th moment (for all 1 < m <= max_mom_ord) are selected, for which all n-th moments (n<m, "mean" for n=1) are not significantly different.

Significant feature difference (d_s in the manuscript) are written to disk via

PySFD.write_feature_diffs()

PySFD.comp_feature_diffs_with_dwells() computes significant differences in sPBSF features and dwell times (for this, PySFD.is_with_dwell_times must be set to True) sPBSF dwell times are defined as the average simulation time (in units of frames) a sPBSF spends in the "on" ("off") state before switching to the "off" ("on") state. Currently, for sPBSF dwell time computations, the corresponding ensemble trajectories have to be plain MD simulation trajectories (i.e. time-dependent, not "samplebatches" trajectories).

!!!!!!

Currently, no significant difference in dwell time has been observed in real MD trajectories that cannot already be explained by a significant difference in interaction frequency.

If you are the first to find such a significant difference, please inform the author to claim your complimentary beer / chocolate bar ! ;-)

To scan your own MD simulations for such peculiar differences, just type in, e.g.:

mySDA = PySFD(...)
mySDA.comp_features(...)
...
mySDA.comp_feature_diffs_with_dwells(num_sigma=2)
abc = mySDA.df_feature_diffs['pbsi.HBond_VMD.std_err'][('bN82A.pcca1', 'WT.pcca3')]
# significant differences in 'ton' ("on"  dwell time) that is NOT significant in 'f' (interaction frequency)
print(abc[(-1, 1, 0)])
# significant differences in 'tof' ("off" dwell time) that is NOT significant in 'f' (interaction frequency)
print(abc[(-1, 0, 1)])

!!!!!!

3. PySFD.comp_and_write_common_feature_diffs()

PySFD.comp_and_write_common_feature_diffs() computes and writes out significant feature differences that are common among different pairwise ensemble comparisons, and optionally that are NOT significantly different among a different set of pairwise ensemble comparisons. These common differences may indicate general mechanistic elements that trigger conformational changes in the simulated system.

Multiprocessed Feature Computation

PySFD.comp_features() uses a two-layer multiprocessing (using the pathos module) to compute features. In this multiprocessing, i*j CPU cores on a single node simultaneously compute features of i ensembles and j trajectories (replica). If the features of more than i ensembles and/or j replica are to be computed, these features will be computed subsequently on the respectice level (of i or of j). In particular, PySFD.comp_features() spawns i non-deamon processes each executing PySFD.run_ens(). Each PySFD.run_ens() then spawns j deamom processes, each executing PySFD.feature_func(), where the latter is the current feature function.

Simulation Input

Features are computed from input simulation trajectories that are to be organized with respect to your current working directory as:

'input/%s/r_%05d/%s.r_%05d.prot.%s' % (myens, r, myens, r, intrajtype)

each with a PDB File containing the topology information:

'input/%s/r_%05d/%s.r_%05d.prot.pdb' % (myens, r, myens, r), where

• myens is the name of the simulated ensemble
• r is the replica index running from 0 to PySFD.num_bs
• intrajtype is the trajectory format

These input trajectories can be one of two different types:

PySFD.intrajdatatype : string, default="samplebatches"
        * 'samplebatches'   : trajectories each containing frames sampled from
                              a stationary distribution of, e.g.,
                              a trajectory-bootstrapped MSM or a bayesian MSM sample
                              of bootstrapped frames drawn, e.g., from
                              a meta-stable set of a Markov State Model
        * 'raw'             : plain simulation trajectories, whose feature statistics 
                              - means or (means and standard deviations) - 
                              are further bootstrapped on the trajectory level
                              (with *num_bs* bootstraps)
        * 'convcheck'       : for convergence checks of the significant feature differences:
                              samples a single bootstrap (with replacement) from
                              the input trajectories containing num_bs (<= numreplica)
                              bootstrapped trajectories;
                              for the actual convergence check, you need to run PySFD
                              with the parameter value 'convcheck' multiple times with
                              varying parameter values of *num_bs*
                              (see Fig. S5 in the manuscript and
                              PySFD/docs/PySFD_example/convcheck folder for example scripts)

and one of two different formats:

PySFD.intrajformat : string, default = "xtc"
        Input trajectory format
        |  "xtc" : gromacs xtc format (additional pdb file for topolgy information required, see example)
        |  "dcd" :         dcd format (additional pdb file for topolgy information required, see example)
        |  "pdb" :         pdb format

The user is responsible for the correct naming and numbering of residues and atoms among different ensembles trajectories (e.g. among different mutants)

Significant Feature Difference Visualization

(Common) significant feature differences can be visualized with PyMOL and VMD scripts provided in $PYSFDPATH/VisFeatDiffs , where $PYSFDPATH, e.g., can be the path to your GitHub-downloaded (or already locally installed) PySFD directory. These scripts are executed directly either within a PyMOL or VMD session, respectively, and read in the (common) significant feature difference tables. You should copy $PYSFDPATH/VisFeatDiffs into your current working directory (see PySFD/docs/PySFD_example).

PyMOL

PyMOLVisFeatDiffs.py: contains the parent PyMOLVisFeatDiffs class

Each of

PyMOLVisFeatDiffs_prf.py,

PyMOLVisFeatDiffs_spbsf.py, and

PyMOLVisFeatDiffs_srf.py

contains all the feature-specific parameter values and an _add_vis() method that is executed by PyMOLVisFeatDiffs.vis_feature_diffs

These feature-specific python files are executed within PyMOL, e.g., via

run $CURRENT_WORKDIR/VisFeatDiffs/PyMOL/PyMOLVisFeatDiffs_spbsf.py

(also see VisFeatDiffs/PyMOL/readme.txt).

Before visualizing coarse-grained significant differences, the coresponding "seg,res"->"rgn" mappings have to be defined in $CURRENT_WORKDIR/scripts/df_rgn_seg_res_bb.dat see PySFD/docs/notes/how_to_define_rgn_to_seg_res_ranges.txt for help on how to create this file from "df_rgn_seg_res_bb" defined in PySFD

Currently, PyMOL visualizations for significant feature differences are implemented only for:

• Single Residual Features (SRF)
• Pairwise Residual Features (PRF)
• sparse Pairwise Backbone/Sidechain Features (sPBSF)

In case you would like to visualize your PySFD results in PyMOL directly from a Python/iPython/Jupyter session, you can use the PySFD.view_feature_diffs() method (see docstring for further documentation).

VMD

VMD_Vis.tcl : main VMD Visualizer function VMD_visualize_feature_diffs

Each of

VMD_Vis.prf.tcl,

VMD_Vis.spbsf.tcl, and

VMD_Vis.srf.tcl contains all the feature-specific parameter values and an add_vis() function that is executed by VMD_visualize_feature_diffs (contained in VMD_Vis.tcl).

These feature-specific tcl files are executed within VMD, e.g., via

set VisFeatDiffsDir $CURRENT_WORKDIR/VisFeatDiffs
source $VisFeatDiffsDir/VMD/VMD_Vis.spbsf.tcl

(also see VisFeatDiffs/VMD/readme.txt)

You can "cursor" through the different simulated ensembles (and their significant feature differences) using the "a" and "d" keys.

Before visualizing coarse-grained significant differences, the coresponding "seg,res"->"rgn" mappings have to be defined in $CURRENT_WORKDIR/scripts/rgn2segres.tcl see PySFD/docs/notes/how_to_define_rgn_to_seg_res_ranges.txt for help on how to create this file from df_rgn_seg_res_bb defined in PySFD

Currently, VMD visualizations for significant feature differences are implemented only for:

• Single Residual Features (SRF)
• Pairwise Residual Features (PRF)
• sparse Pairwise Backbone/Sidechain Features (sPBSF)