Work with trained factor models in Python.

This library provides convenience functions to load and visualize factor models trained with MOFA+ in Python. For more information on the multi-omics factor analysis v2 framework please see mofapy2 and MOFA2 GitHub repositories as well as the website.

Getting started

Installation

pip install git+https://github.com/gtca/mofax
# or
pip install mofax

Training a factor model

Please see the MOFA+ GitHub repository for more information on training the factor models with MOFA+.

Loading the model

Import the module and create a connection to the HDF5 file with the trained model:

importmofaxasmfxmodel=mfx.mofa_model("trained_mofaplus_model.hdf5")

The connection is created in the readonly mode by default and can be terminated by calling the close() method on the model object at the end of the working session:

model.close()

Model object

Model object is an instance of a mofa_model class that wraps around the HDF5 connection and provides a simple way to address the parts of the trained model such as expectations for factors and for their loadings (weights) eliminating the need to traverse the HDF5 file manually. The original connection to the HDF5 file is exposed via the model.model attribute.

Model methods

Simple data structures (e.g. lists or dictionaries) are typically returned upon accessing the properties of the mofa model, e.g. model.shape:

model.shape# returns (10138, 1124)# samples^ ^features# (cells)

More complex structures are typically returned when calling methods such as model.get_samples() to get sample -> group assignment as a pandas.DataFrame while also providing the way to only get this information for specific groups or views of the model. model.get_cells() works the same way.

model.get_cells().head()
# returns a pandas.DataFrame object:# group	cell# 0	T_CD4	AATCCTGCACATCGCC-1# 1	T_CD4	AAGACGTGTGATGCCC-1# 2	T_CD4	AAGGAGCGTCGGCATG-1# 3	T_CD4	AATCCGTCACGAGACG-1# 4	T_CD4	ACACCGAGGAGGTTGA-1

Use model.metadata to get the metadata table — it's a shorhand for samples_metadata, there's also features_metadata available.

model.metadata.head()
# returns a pandas.DataFrame object:# group n_genes# sample# AATCCTGCACATCGCC-1 T_CD4 1087# AAGACGTGTGATGCCC-1 T_CD4 1836# AAGGAGCGTCGGCATG-1 T_CD4 2216# AATCCGTCACGAGACG-1 T_CD4 1615# ACACCGAGGAGGTTGA-1 T_CD4 1800

To get expectations of W (weights) and Z (factors) matrices, use get_weights() and get_factors(), respectively. There's also a df=True option to get expectations as a Pandas data frame rather than a NumPy 2D array.

model.get_factors(factors=range(3), df=True).head()
# Factor1 Factor2 Factor3# AATCCTGCACATCGCC-1 0.012582 -0.093512 -0.011228# AAGACGTGTGATGCCC-1 0.001091 -0.027217 -0.011331# AAGGAGCGTCGGCATG-1 -0.015097 0.093493 -0.010593# AATCCGTCACGAGACG-1 -0.046222 0.225920 0.010083# ACACCGAGGAGGTTGA-1 0.011766 -0.055964 -0.011298

Variance explained by each factor per view and per group is calculated during the tranining and stored in the model file and can be accessed with get_r2():

model.get_r2().head()
# Factor	View Group	R2# 0	Factor1	drugs group1	13.589131# 1	Factor1	methylation group1	17.330235# 2	Factor1	rna group1	7.032133# 3	Factor1	mutations group1	22.725224# 4	Factor2	drugs group1	26.374409

MEFISTO models support

MEFISTO models can feature a few additional concepts such as covariates and interpolated factors. Covariates can be accessed via model.covariates_names and model.covariates. If interpolated factors were learnt for new values during training, they are exposed at model.interpolated_factors and also can be obtained in a long DataFrame with model.get_interpolated_factors(df_long=True).

Utility functions

A few utility functions such as calculate_factor_r2 to calculate the variance explained by a factor are provided as well.

Plotting functions

A few basic plots can be constructed with plotting functions provided such as plot_factors and plot_weights. They rely on and limited by plotting functionality of Seaborn.

Please check the notebooks for detailed examples. Some of the implemented plots are demonstrated below.

Contributions

In case you work with MOFA+ models in Python, you might find mofax useful. Please consider contributing to this module by suggesting the missing functionality to be implemented in the form of issues and in the form of pull requests.