Pathway-Regularized Matrix Factorization (PRMF)

Slides from ISMB2018

Installation

git clone https://github.com/gitter-lab/prmf.git
cd prmf
pip install -r requirements.txt
python setup.py install

Pathway Data Installation

A script to download pathways from KEGG and transform them to the expected graph format is script/download_pathways.sh. This script runs R and has separate (installation) dependencies from the core PRMF package. Its usage is below.

# install the installation dependencies
Rscript script/install_pathway_dependencies.R # TODO

# from the prmf root directory
mkdir pathways
script/download_pathways.sh pathways

recount2 Example

Follow these steps if you would like to run PRMF to learn patterns from recount2 and apply those patterns to a rare disease dataset such as Neurofibromatosis. Skip to the next section if you would like to run PRMF on your own data.

0. Install R dependencies and download a recent harmonized dataset from the NF data portal at https://www.synapse.org/#!Synapse:syn18137070 and put the file 2019-02-01-bioBank_glioma_cNF_pnftidiedData.csv.gz in the PRMF directory. Decompress the file.

script/prep_instance.R
# visit the above URL in your browser or use the synapse client below
synapse get syn18137070
gunzip 2019-02-01-bioBank_glioma_cNF_pnftidiedData.csv.gz

1. Download the recount2 data archive, unzip the archive, and make the files available in ~/data/recount2_PLIER_data by running prep_recount.sh

script/prep_recount.sh

2. Transform the data with some light pre-processing for PRMF

script/recount_rds_to_tsv.R

3. Run PRMF on the transformed data to discover the latent patterns hidden in the data.

export OUTDIR=~/data/prmf-results ; script/prmf.py -k 10 --data ~/data/recount2_PLIER_data/recount_rpkm.tsv --manifolds pathways/kegg_ensg/*.graphml --outdir $OUTDIR > $OUTDIR/prmf.out 2> $OUTDIR/prmf.err ; echo "Exit Status: $?" >> $OUTDIR/prmf.out

4. PRMF decomposes the sample x gene data into a sample x latent matrix and a gene x latent matrix. These matrices and the decomposition is denoted X = U V^T and PRMF produces files U.csv and V.csv in the output directory. We next apply the gene x latent matrix in a transfer learning context: we use the latent gene patterns to quantify latent variable activity in NF.

transfer_learning.R -y 2019-02-01-bioBank_glioma_cNF_pnftidiedData.csv -z $OUTDIR/V.csv --outdir $OUTDIR

5. TODO kolmogorov_smirnov.R

Running

With the software and necessary data installed, you can now run PRMF.

export OUTDIR=prmf_out
prmf_runner.py -k 10 --data ~/data/recount2_PLIER_data/recount_rpkm.tsv --normalize --delimiter '\t' --manifolds pathways/kegg_ensg/*.graphml --outdir $OUTDIR > $OUTDIR/prmf.out 2> $OUTDIR/prmf.err
echo "Exit Status: $?" >> $OUTDIR/prmf.out

Graph Format

The pathways used in PRMF are assumed to be undirected graphs in the .graphml file format where the node attribute. A sample is given below.

<?xml version="1.0" encoding="UTF-8"?>
<graphml xmlns="http://graphml.graphdrawing.org/xmlns"
         xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
         xsi:schemaLocation="http://graphml.graphdrawing.org/xmlns
         http://graphml.graphdrawing.org/xmlns/1.0/graphml.xsd">
<!-- Created by igraph -->
  <key id="v_name" for="node" attr.name="name" attr.type="string"/>
  <key id="e_weight" for="edge" attr.name="weight" attr.type="double"/>
  <graph id="G" edgedefault="directed">

    <!-- Node data sample -->
    <node id="n0">
      <data key="v_name">ENSG00000141837</data>
    </node>
    <node id="n1">
      <data key="v_name">ENSG00000148408</data>
    </node>
    ...

    <!-- Edge data sample -->
    <edge source="n2" target="n13">
      <data key="e_weight">1</data>
    </edge>
    <edge source="n2" target="n12">
      <data key="e_weight">1</data>
    </edge>
    ...

Matrix Data Format

PRMF expects CSV-like data which describes gene expression measurements for a set of samples. A generated dataset which illustrates this is available in the test directory at test/script/PRMF/test_cv/data_header.tsv. Samples can also be named. The gene names must correspond to node names as described in the Graph Format section.

ENSP0	ENSP1	ENSP2	ENSP3	...
sample1	198.45537672404348	130.95172604578474	94.30322859309669	120.43851578854357	...
sample2	170.9483727406653	103.88265492990088	88.86178431789227	116.51811746441824	...
...

Tests

# from the prmf root directory
python test/run_tests.py

Dependencies

• Python (3.7.0)
  • numpy
  • scipy
  • matplotlib
  • sklearn
  • networkx

Installation Dependencies

• R (3.5.1)
  • KEGGREST
  • KEGGgraph
  • graph
  • igraph
  • argparse