Welcome to PyProteoNet’s documentation!

PyProteoNet is a Python package for imputation, (peptide-to-protein) aggregation/summarization as well as in silico data generation of proteomics data with a specific focus on the development and evaluation of imputation methods for proteomics datasets.

Datasets are given by a set of interrelated molecules with assigned (abundance) values for multiple samples. Values are structured in value columns allowing the representation and comparison of multiple measurements per molecule and sample (e.g. measured and imputed abundance values). While the focus is on datasets consisting of interrelated proteins and peptides as commonly used in proteomics, the underlying data structures allow for arbitrary molecule types such that additional measurements like mRNA can be incorporated.

Explicitely modeling the relations between different molecules (e.g. proteins and peptides) as well as providing a range of common aggregation as well as imputation methods together with proteomics specific evaluation metrics PyProteoNet is a one-stop shop for applying and benchmarking aggregation and imputation methods for proteomics datasets.

In addition, new imputation methods can be implemented and benchmarked in PyPoteoNet. To this end, PyProteoNet provides functions to transform proteomics dataset into a graph representation as defined by the Deep Graph Library (DGL). This allows for the development and integration of new proteomics-specific imputation methods based on graph neural networks (GNNs). While two reference GNN-based imputation methods are already implemented, PyProteoNet aim is to provide a plattform for the development of new methods.

Installation

First, clone the PyProteoNet repository to a local directory. Installation is best done within a Conda environment (you might also consider Mamba as a drop-in replacement for Conda with higher performance). You can install an already configured environment using the environment.yml included in the repository. To do so, just run the following command inside your cloned PyProteoNet directory. This creates a mamba environment called pyproteonet:

mamba env create --name pyproteonet --file environment.ym

Alternatively, you can create your own custom environment or use an existing one. In this case, it is advised to install the following requirements via Conda/Mamba because they are either not available via pip or using the pip version might lead to problems:

• r-base, conda install -c conda-forge r-base (you need an R installation because several provided imputation methods are wrappers around R packages)

• Pytorch, conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia (see the PyTorch website) for instructions for your system and Cuda version)

• DGL, conda install -c dglteam dgl (see the DGL website) for instructions for your system and Cuda version)

Afterward, you can install PyProteoNet in the created Conda/Mamba environment via pip by running the following command inside the root of your PyProtoNet directory:

pip install ./

Features

Data

Datastructures for proteomics data consisting of per-molecule values (e.g. abundance measurements) for several samples and different interrelated molecule types (e.g. proteins and peptides). Pandas interface for easy data handling and manipulation.

Simulation

Several functions for the creation of in-silico proteomics data. This includes the simulation of protein and peptide abundances, the simulation of fold changes and the simulation of missing values.

Aggregation

Native implementations of MaxLFQ, IBAQ, Top3 as well as simple peptide-to-protein aggregation methods like sum, mean, median.

Imputation

Several reference imputation methods are provided either implemented directly in Python or via a wrappers around common R packages.

Single Value Imputation Methods	Global Structure Imputation Methods	Local Similarity Imputation Methods
MinDet	PCA, PPCA, BPCA	MissForest
MInProb	ISVD	KNN
Mean, Median	Denoising Autoencoder, Variational Autoencoder	Collaborative Filtering
	MLE	Local Least Squares
		Iterative Imputation

Next to those established methods, PyProteoNet provides imputation methods based on homogeneous as well as heterogeneous graph neural networks (GNNs) specifically for the imputation of missing values in proteomics data. New DNN- and GNN-based methods can be easily implemented and integrated into the existing framework.

Evaluation

Several metrics for the evaluation of aggregation and imputation methods. Common evaluation metrics like abosulte error (AE) mean absolute error (MAE), mean squared error (MSE), Pearson correlation coefficient (PearsonR) can be computed for comparing aggregated/imputed against ground truth abundance values. In addition, catering for the specific needs of proteomics data fold-change-based evaluation methods are provided to evaluate based on molecule ratios between samples as well as discovered differential expressions.

Tutorials

• Getting Started
  • Molecules, Graphs, and MoleculeSets
  • Samples and Values
  • Creating a Dataset Directly from Pandas Dataframes
  • Protein Aggregation
  • Missing Value Imputation
  • Graph Neural Network Imputation
• Simulation of Protein-Peptide Datasets for Benchmarking Aggregatin and Imputation Methods
• Evaluating Imputation Against Ground Truth Abundance Values
  • Simulating a Dataset
  • MaxLFQ aggregation
  • Missing Value Imputation
• Evaluating Imputation Against Ground Truth Fold Change
  • Simulating a Dataset
  • MaxLFQ aggregation
  • Missing Value Imputation
  • Evaluation
• Developing new Imputation Methods
  • Developing Imputation Methods Working on One Molecule Type
  • Developing Imputation Methods Based on (Graph) Neural Networks

Python API

• pyproteonet.data
• pyproteonet.aggregation
• pyproteonet.simulation
• pyproteonet.imputation
• pyproteonet.metrics
• pyproteonet.io
• pyproteonet.masking
• pyproteonet.dgl

Indices and tables

• Index

• Module Index

• Search Page