zfpkm

Overview

This package performs zFPKM on RNA-seq FPKM data. This implementation is adapted from ronammar/zFPKM, which was originally based on Hart et al. 2013 (PMID24215113). The original article recommends selecting an active/inactive cutoff at -3; this value was selected based on experimental data that indicated the number of active promoters becomes more than the number of repressed promoters Figure 1.

Installation

This package can be installed from GitHub or PyPi

PyPi

python3 -m pip install zfpkm

Latest GitHub Version

python3 -m pip install git+https://github.com/JoshLoecker/zfpkm.git

Usage

To calculate zFPKM, simply import the zfpkm function and provide the raw, non-normalized FPKM values as a pandas DataFrame. The row names of the input dataframe should be the genomic identifier (Entrez IDs, Ensembl IDs, Gene Symbols, etc.) and the column names should be the sample name. The returned DataFrame will have the same number of rows and columns (in the same order provided) with a modified z-transformation applied.

import pandas as pd

from zfpkm import zFPKM, zfpkm_plot


def main():
    fpkm = pd.read_csv("fpkm.csv", index_col=0, header=0)
    zfpkm_df, zfpkm_results = zFPKM(fpkm)
    zfpkm_df.to_csv("zfpkm.csv", index=True)
    zfpkm_plot(zfpkm_results, save_filepath="zfpkm_density.png")


if __name__ == "__main__":
    main()

Results

Expected zFPKM Distribution

The following figure shows the expected FPKM ('fpkm_density', in teal) and zFPKM ('fpkm_density_scaled', in salmon). The Gaussian curve is fit to the peak of the FPKM density distribution. Values > -3 can be marked as active.

Figure 1: The expected zFPKM Gaussian distribution overlaid on the FPKM distribution

Actual zFPKM Distribution (from this package)

The following figure shows the calculated zFPKM from this package. Like Figure 1, the FPKM ('fpkm_density', in teal) and zFPKM ('fpkm_density_scaled', in salmon) are overlaid on the highest FPKM peak.

Example code showing how this graph was generated can be found in examples/example_zfpkm.py.

Figure 2: The actual zFPKM Gaussian distribution from this package overlaid on the FPKM distribution

Comparison with scikit-learn and scipy

Figure 3 and 4, below, were generated with scikit-learn==1.7.2 and scipy==1.16.3, respsectively. These figures, while very similar to the expected and actual zFPKM distributions, have several noteable differences:

1. The maximum density is ~37% greater than the expected and actual zFPKM distributions (~0.131 expected vs ~0.180 for scikit-learn and scipy)
2. The left-hand of the FPKM density distribution is much smoother than the expected results; this comes specifically from a larger bandwidth value than the original R source.

Taken together, this results in the final zFPKM scores using scikit-learn or scipy being similar to, but distinct enough, from the expected values to potentially cause problems in downstream analysis.

Figure 3: scikit-learn FPKM ('fpkm_density', in teal) & zFPKM ('fitted_density_scaled', in salmon) distribution

Figure 4: scipy FPKM ('fpkm_density', in teal) & zFPKM ('fitted_density_scaled', in salmon) distribution