EPIFANY is a protein inference engine based on a Bayesian network. Currently a similar model to
Fido is used with the main parameters alpha (pep_emission), beta (pep_spurious_emission) and gamma (prot_prior).
If not specified, these parameters are trained based on their classification performance and calibration
via a grid search by simply running with several possible combinations and evaluating. Unless you see very extreme output
probabilities (e.g. many close to 1.0) or you know good parameters (e.g. from an earlier run),
grid search is recommended, although slower. The tool will merge multiple idXML files (union of proteins
and concatenation of PSMs) when given more than one. This is useful for fractions or replicates.
It assumes one search engine run per input file but might work on more with undefined behaviour.
Proteins need to be indexed by OpenMS’s PeptideIndexer but this is usually done before
Percolator/IDPEP since target/decoy associations are needed there already. Make sure that the input PSM
probabilities are not too extreme already (garbage in – garbage out). After merging, the input probabilities
are preprocessed with a low posterior probability cutoff to neglect very unreliable matches. Then
the probabilities are aggregated with the maximum per peptide and the graph is built and split into
connected components. When compiled with the OpenMP flag (default enabled in the release binaries) the tool is multithreaded which can
be activated at runtime by the threads parameter. Note that peak memory requirements
may rise when processing multiple components of the graph at the same time. The tool offers an option for a greedy group resolution
which is helpful for reducing false positive identifications in the presence of many shared peptides between present and non-present proteins.
Note that it might reduce true positives, too and introduces a fake certainty by removing proteins without evidence after retaining only associations
between a peptide and its best producing group (after probabilistic inference). The option is mainly suggested when protein inference
is the last step of the analysis. A middle-ground between greedy resolution and standard (sum-product) inference is a regularized
max-product inference (for this, set the regularize flag and choose a p-norm smaller than 0 [implying infinity]).
1) Download the installer for your platform via the button above
2) Install OpenMS following the Installation instructions
3) After installation and adding the TOPP tools (including EPIFANY) to your global Path environment,
you can start scripting. Follow the instructions here for some introduction to scripting with OpenMS
1) Download the KNIME installer for your platform via the button above
2) Install KNIME
3) After installation follow this small video tutorial to install the OpenMS plugin. Use the following update site URL for
the copy-paste step in the video: https://abibuilder.informatik.uni-tuebingen.de/archive/openms/knime-plugin/updateSite/experimental/feature/proteomic_lfq
You can give it any name (e.g. OpenMS EPIFANY Update Site).
4) You will then find EPIFANY (together with the other OpenMS nodes) in the node repository in the lower left of the program.
5) You can create a new workflow with it or import the workflow for our example data. Don’t forget to configure Input File(s) to point to the files on your computer.
Install or build OpenMS (see OpenMS’ requirements at the bottom of the page). You can do it either standalone or as a KNIME
Command line documentation can be found here in the UTILS section of the OpenMS documentation under the tool’s name.
Contradicting connected components may dictate runtime in certain parameter settings.
Extensive parameter grid search is computationally expensive (may be reduced if there are good estimates).
Direct input for the tool: An idXML file with indexed and target-decoy annotated proteins and peptides that carry probabilities (e.g. from OpenMS’ IDPosteriorProbability, Percolator, Peptide/iProphet…)
General input for a complete protein inference workflow: Centroided mzMLs from a bottom-up/shotgun data-dependent acquisition experiment. If your data is in “.raw” format you can do the conversion and centroiding with e.g. proteo-wizard (included in every OpenMS installation, too).
Inputs can be replicates and/or fractions. Depending on your goals you could also merge conditions. We advise to calculate peptide probabilities on a global (i.e. merged) level, as done in the example workflow.
Download sources and example data:
Experimental installer downloads until it is merged into the OpenMS Release: OpenMS experimental archive
Example Data: You can use the converted mzMLs on our server as direct input to the KNIME workflow that is also linked there. The data comes from the iPRG2016 study (sample “B”). You can also use the idXML on the server for a direct input to EPIFANY.
Example KNIME Workflow: Workflows implemented in KNIME are also found on the server. The outputs of the workflow are compatible with the evaluation script of the iPRG2016 study. Experimental update site for the plugin installation in KNIME (copy-paste when adding a new update site under Help->Install New Software): https://abibuilder.informatik.uni-tuebingen.de/archive/openms/knime-plugin/updateSite/experimental/feature/proteomic_lfq
OpenMS implementation sources: github.com (Epifany.cpp and BayesianProteinInferenceAlgorithm class)
Inference library source: Evergreen