____ ___________________ __ __ ______
\ \ / /___ ___| _ \| | | | _ \
\ \ / / | | | |_) | | | | |_) |
\ V / | | | ___/| | | | /
\ / | | | | | |_| | \
\ / __| |__| | | | |\ \
\_/ |________|__| \_____/|_ | \__\
Code for running VIPUR can be found here.
The VIPUR training set files (feature set and original structures) and third party executables (built for 64-bit linux) can be found at https://drive.google.com/a/nyu.edu/folderview?id=0B6exAih8BuuUVXBHRUlFRVZKd1U&usp=drive_web.
What is VIPUR?
The VIPUR pipeline analyzes protein variants by considering conservation scores
and structural scores to identify variants that are likely to disrupt protein
function
Using Rosetta, these structural features allow you to interpret what is causing
the variant to be identified as deleterious
Conservation scores are derived from a PSSM of similar sequences found
using PSIBLAST against the NCBI nr database
Structural analysis is done using Rosetta to:
consider variant structures by rapid structure optmization, allowing
fast evaluation of approximate variant ddG values (Rosetta ddg_monomer)
and
refine variant structures and consider the distribution of energies and
structural scores (rms, gdtmm) across several low energy conformations
(physically near the input conformation, Rosetta relax)
Additional features are provided by an internal "aminochange" classification
(crude similarity of amino acid properties) and the variant position
surface area, evaluated using PROBE
These analyses are combined with a learned Logistic Regression model to
classify input variants as "neutral" or "deleterious" and provide a
structurally-informed hypothesis as to why variants are likely to
disrupt the protein
##################
(rewrite all this, copied from VIPUR.py for now)
VIPUR INSTALLATION
- download the VIPUR module
- verify that you have the necessary software
running VIPUR requires
PSIBLAST
PROBE
Rosetta
PyMOL (or PyRosetta)
-add paths to settings.py
DOWNLOAD PSIBLAST
PSIBLAST is a sequence search tool based on iterative alignment profile scans
PSIBLAST is part of the free NCBI BLAST+ distribution that can be found at:
ftp://ftp.ncbi.nlm.nih.gov/blast/executables/LATEST/
There are many tutorials on downloading, installing, and running PSIBLAST,
we suggest the NCBI usage book:
http://www.ncbi.nlm.nih.gov/books/NBK52640/
VIPUR has been benchmarked using BLAST 2.2.25+
with the NCBI nr database (downloaded Feb 2012)
Camacho, C. et al. BLAST+: architecture and applications.
BMC Bioinformatics 10(421) (2009).
DOWNLOAD PROBE
PROBE is a tool for analyzing contacts within protein structures (PDB format)
Here, we use PROBE as an accurate method for calculating the
ACCessible surface area of Protein amino acids (ACCP), the fraction of
potential surface area in contact with other residues
PROBE is freely available for download:
http://kinemage.biochem.duke.edu/software/probe.php
VIPUR has been benchmarked using probe.2.12.071128
Word, et. al. Visualizing and Quantifying Molecular Goodness-of-Fit:
Small-probe Contact Dots with Explicit Hydrogens.
J. Mol. Biol. 285, 1711-1733 (1999).
DOWNLOAD ROSETTA
Rosetta is a software suite capable of modeling and designing proteins
and other biomacromolecules
Here, we use the well established Rosetta protocols "relax" and "ddg_monomer"
for refining variant protein structures and evaluating protein energetics
Rosetta is free for academic use with the license and download tutorial at:
https://www.rosettacommons.org/software/license-and-download
Compiling Rosetta may introduce additional dependencies depending on your system
Here, we will use the default "release" compilation settings
(e.g. no "mode=MPI" etc.)
Please consult the Rosetta user documenation and forums if you encounter
complications during setup:
https://www.rosettacommons.org/docs/latest/Build-Documentation.html
Note: both relax and ddg_monomer perform stochastic searches and may output
slightly different energy values on evaluation with different random seeds
VIPUR has been benchmarked using the Rosetta 3.4 release version (54167)
Leaver-Fay, A. et al. ROSETTA3: An object-oriented software suite for
the simulation and design of macromolecules.
Methods in Enzymology 487, 548-574 (2011).
DOWNLOAD PYMOL
PyMOL is a tool for molecular visualization and analysis
PyMOL has several versions with advanced features, however we only require
simple PyMOL functionality
Please consult the PyMOL website to determine which license works best for you:
http://pymol.org/educational/
PyMOL can be downloaded at:
http://pymol.org/dsc/
Note: PyMOL will be used to create variant structures, PyRosetta can
alternatively be used for this task, however you only need one of these programs
The PyMOL Molecular Graphics System, Version 1.5.0.4 Schrodinger, LLC.
Note: you only need a copy of PyMOL or PyRosetta to run VIPUR
(for creating variant structures)
DOWNLOAD PYROSETTA
PyRosetta is an interactive Python interface to Rosetta
PyRosetta is free for academic use but on a separate license from Rosetta:
http://c4c.uwc4c.com/express_license_technologies/pyrosetta
The PyRosetta software and a download tutorial can be found at:
http://www.pyrosetta.org/dow
Chaudhury, S. et al. PyRosetta: a script-based interface for implementing
molecular modeling algorithms using Rosetta.
Bioinformatics 26(5), 689-691 (2010).
###########
USING VIPUR
VIPUR is written as a simple Python module
Once you have VIPUR and the required software setup, you can use VIPUR.py as
a Python script or a library (from VIPUR import run_VIPUR)
other scripts that make up VIPUR outline the specific feature generation steps,
including input file parsing and output file analysis
Currently, VIPUR supports a commandline interface to the options necessary to
run a single protein variant, specifying the input structure file (-p)
<pdb_filename> and variants (-v) <variant_filename>
ex.
python VIPUR.py -p example_input/2C35.pdb -v example_input/2C35.txt
VIPUR can also run on a directory of (PDB, .txt) file pairs by inputting a path
to the directory containing the files (-p), defaults to the current directory
ex.
python VIPUR.py -p example_input/
Please consult the help (-h) for more details on run options, including:
-d filename of the output predictions file
-o path for output to be written (several intermediate files)
-c chain of the input PDB that contains the native sequence (if not A)
-s filename of the intermediate (native) sequence file (FASTA format)
-w option to write out the numbering map (PDB to 1-indexed)
-q option to run in "sequence only" mode, no structural analysis
An example of VIPUR input files is provided in "example_input"
and their expected output is provided in "example_output_reference"
You can automatically run VIPUR on this demo with the --demo option
ex.
python VIPUR.py --demo
VARIANT INPUT FORMAT
VIPUR currently takes variant input files as plain text files with one variant
per line, expecting a reference amino acid for the native protein
ex.
E14R
R84P
A101W
the native reference amino acid is necessary to ensure the input numbering
matches the input protein structure (so you don't have to worry about constantly
renumbering your indices or PDB files)
currently, any input variants that do not have correct positions or native
amino acids will be skipped
INTERPRETING RESULTS
VIPUR outputs a ".predictions" file containing a summary of the predictions and
analysis of input variants. The default output file is tab delimited ('\t')
with one line per variant indicating:
variant the variant (ex. P335R)
predicted label "deleterious" or "neutral"*
prediction confidence P(deleterious|analysis), deleterious score*
structure-only label prediction of the structure-only classifier
structure-only confidence structure-only confidence score
sequence-only label prediction of the sequence-only classifier
sequence-only confidence sequence-only confidence score
exposure "surface" or "interior" for the position**
"essential" score can indicate conserved positions***
ddG prediction prediction of ddG in kcal/mol (approx)
interpretation simple interpretation of the effect****
explanation top contributions to the interpretation****
*VIPUR was trained on a dataset of natural variants, pseudomutations, and
protein variants from mutagenesis experiments. Variants are curated as
"deleterious" if they have literature or UniProt annotations indicating loss of
an essential protein function. Pseudomutations are from HumDiv and assumed to be
"neutral" (though not all neutral examples come from pseudomutations).
Please see the <main text> for a full description of the VIPUR training set.
Note: this binary label refers to PROTEIN function which may, or may not,
indicate disease association (a more complicated phenotype)
variants that are predicted "deleterious" are expected to lack an essential
protein function (in many cases, the protein variant is misfolded or
too unstable to maintain its native fold, as indicated by the structure-based
features), however they may display a severely reduced function or
maintain other functions
variants that are predicted "neutral" are expected to function effectively as
the native protein, however they may have slightly reduced function
(nearly-neutral, not predicted as not well-curated data exists)
Note: our "deleterious" predictions are NOT synonymous with the impact
on stability, deleterious variants may minimally change or even stabilize
a protein (ex. can make an enzyme too rigid) and neutral variants can
reasonably destabilize a protein as long as it still folds,
please consult the ddG prediction provided by the ddg_monomer protocol to
interpret changes in stability
Note: our "deleterious" predictions are NOT synonymous with "disease assocition"
while many disease associated/causal variants are loss-of-function changes,
the disruption of protein function is itself insufficient to indicate the cause
of a disease (this comes from knowledge of the protein's function or prior
variant association)
the goal of VIPUR is to provide a clear prediction of variants that disrupt
protein function to help INTERPRET variants that already have some correlated
label (e.g. disease or other phenotype)
not all deleterious predictions of VIPUR necessarily influence disease, however
confident deleterious predictions of variants known to be disease associated
suggests a strong effect worth investigating
Since VIPUR contains a binary classifier, the output confidence metric (the
learned conditional probability) indicates confidence of the binary prediction.
This confidence score is effectively a "deleterious" score, with higher values
indicating increased probability of being deleterious.
Lower values similarly indicate an increased probability of being neutral
(since the labels are binary).
You can filter your results to contain the most confident deleterious
predictions (ex. identify the 5 protein variants with the highest
deleterious confidence etc.).
**"exposure" indicates the local environment of the variant amino acid position
Here, we restrict "exposure" to "surface" and "interior", indicating if the
position is (approximately) on the protein surface or inside the protein.
The VIPUR training set does not suggest there is a significant difference
in performance for positions on the surface or interior of the protein,
although available data is imbalanced (many surface variants are annotated
"neutral", many interior variants are annotated "deleterious").
***Disagreement between the overall VIPUR classifier (sequence+structure) and
the structure-only classifier indicates a strong sequence signal without
apparent destabilization of the monomer structure. In some cases, this is due to
the inadequacy of the monomeric structure to capture the protein energetics,
suggesting this amino acid is important of interactions (can be with other
proteins, metals, ligands, nucleic acids etc.).
When analyzing VIPUR output, consider surface variants with
high (>=.8) deleterious scores and low structure-only scores (difference >=.2)
as "potential interaction sites".
Please see the <main text> for more information on VIPUR analysis and scores.
****VIPUR automatically identifies which features contribute to
deleterious predictions. The structure-based features are directly interpretable
since they represent specific physical interactions (hydrogen bonding,
solvent favorability, dilsufide bond strain, etc.) and expected distributions
based on other proteins (backbone conformational strain,
side chain interactions, etc.).
The top structure features are included as an "explanation" (3 by default,
can be set to include more).
We also include a reduced "interpretation" of the variant deleterious effect as:
structural conservation
The variant destabilizes the protein structure, likely due to improper size or
surface area. This includes unfit deviations in backbone conformation
(e.g. G and P positions) and packing configurations
(e.g. V, L, I, M differences). In some cases, it is seen that binding sites
have more restricted conformations, detectable by destabilizing variations even
in the absence of the binding partner (e.g. structurally conserved and the
variant cannot attain the necessary conformation).
disrupted chemical interactions
This variant eliminates an important and/or stabilizing chemical interaction of
the protein. This includes hydrogen bonding partners, conserved hydroxyl groups,
salt bridges, and disulfide bonds.
potential interaction site
As noted above*** high deleterious scores with low confidence structure-only
scores can indicate conservation where there is no structural evidence and
can indicate potential interaction sites on the protein surface.
other
The remaining features are difficult to directly interpret. They indicate
sequence conservation and structural conservation, but cannot clearly suggest
why.
PREPARING STRUCTURES FOR ROSETTA
Rosetta requires structural models that are cleanly readable. For
many applications, this includeds removal of waters, ligands, nucleic acids,
and metals. Rosetta has methods for handling all of these inputs, however
our benchmark uses protein structures stripped of all these additional
coordinates (note: while Rosetta can handle these cases, many proteins lack
models of docked conformations, ligands, etc. and we wanted to ensure the same
information was available for all samples). There are several methods available
for cleaning structures for Rosetta, including simply removing all HETATM and
nucleic acid ATOM coordinates (note: Rosetta can handle input structures with
missing densities or regions ex. removed a non-canonical amino acid).
The script available at:
https://github.com/Olvikon/miscellaneous_scripts/blob/master/process_pdb.py
outputs a directory containing the monomeric structure cleaned for use with
Rosetta (either ".clean.pdb" or ".protein.pdb" if nucleic acid lines are
removed).
More suggestions on how to clean structures for Rosetta is available at:
https://www.rosettacommons.org/manuals/archive/rosetta3.5_user_guide/dd/da1/preparing_structures.html
note: additional refinement steps of initial models are unnecessary since
Rosetta relax is run as part of feature generation, and the benchmark
performance is using structures that have not been pre-relaxed