ponyo 0.6.1

Install functions to simulate gene expression compendia

ponyo

Alexandra J. Lee and Casey S. Greene 2020

University of Pennsylvania

This repository is named after the the character Ponyo, from Hayao Miyazaki's animated film Ponyo, as she uses her magic to simulate a human appearance after getting a sample of human blood. The method simulates new gene expression data by training a generative neural network on existing gene expression data to learn a representation of gene expression patterns.

Installation

This package can be installed using pip:

pip install ponyo

Types of simulations

There are 3 types of simulations that ponyo implements:

Name	Description
Simulation by random sampling	This approach simulates gene expression data by randomly sampling from the latent space distribution. The function to run this approach is divided into 2 components: simulate_by_random_sampling is a wrapper which loads VAE trained models from directory <root>/<analysis name>/"models"/<NN_architecture> and run_sample_simulation which runs the simulation. Note: simulate_by_random_sampling assumes the files are organized as described above. If this directory organization doesn't apply to you, then you can directly use run_sample_simulation by passing in the trained VAE models. An example of how to use this can be found here.
Simulation by latent transformation	This approach simulates gene expression data by encoding experiments into the latent space and then shifting samples from that experiment in the latent space. Unlike the "Simulation by random sampling" approach, this method accounts for experiment level information by shifting samples from the same experiment together. The function to run this approach is divided into 2 components: simulate_by_latent_transformation is a wrapper which loads VAE trained models from directory <root>/<analysis name>/"models"/<NN_architecture> and run_latent_transformation_simulation which runs the simulation. Note: simulate_by_latent_transformation assumes the files are organized as described above. If this directory organization doesn't apply to you, then you can directly use run_latent_transformation_simulation by passing in the VAE models trained using run_tybalt_training in vae.py. There are 2 flavors of this approach: simulate_by_latent_transform inputs a dataset with multiple experiments (these are your template experiments) and then it outputs the same number of new simulated experiments that are created by shifting each of those input template experiments. An example of how to use this can be found here. The second flavor is shift_template_experiment which inputs a single template experiment and can output multiple simulated experiments based on that one template by shifting that template experiment to different locations in the latent space. An example for how to use this can be found here.

How to use

Example notebooks using ponyo on test data can be found in human_tests

Additionally, this method has been used in simulate-expression-compendia and generic-expression-patterns repositories.

Setting random seeds

To keep the VAE training deterministic, you will need to set multiple random seeds:

1. numpy random
2. python random
3. tensorflow random

For an example of this, see human_tests

Configuration file

The tables lists the core parameters required to generate simulated data using modules from ponyo. Those marked with * indicate those parameters that will vary depending on the type of approach .

Name	Description
local_dir	str: Parent directory on local machine to store intermediate results
dataset_name	str: Name for analysis directory containing notebooks using ponyo
raw_data_filename	str: File storing raw gene expression data
normalized_data_filename	str: File storing normalized gene expression data. This file is generated by normalize_expression_data().
metadata_filename*	str: File containing metadata associated with data
experiment_ids_filename*	str: File containing list of experiment ids that have gene expression data available
scaler_transform_filename	str: Python pickle file to store mapping from normalized to raw gene expression range. This file is generated by normalize_expression_data().
simulation_type	str: Name of simulation approach directory to store results locally
NN_architecture	str: Name of neural network architecture to use. Format NN_<intermediate layer>_<latent layer>
learning_rate	float: Step size used for gradient descent. In other words, it's how quickly the methods is learning
batch_size	str: Training is performed in batches. So this determines the number of samples to consider at a given time
epochs	int: Number of times to train over the entire input dataset
kappa	float: How fast to linearly ramp up KL loss
intermediate_dim	int: Size of the hidden layer
latent_dim	int: Size of the bottleneck layer
epsilon_std	float: Standard deviation of Normal distribution to sample latent space
validation_frac	float: Fraction of input samples to use to validate for VAE training
num_simulated_samples*	int: If using random sampling approach, simulate a compendia with these many samples
num_simulated_experiments*	int: If using latent-transformation approach, simulate a compendia with these many experiments
num_simulated*	int: If using template-based approach, simulate these many experiments
metadata_delimiter*	str: Delimiter to parse metadata file
metadata_experiment_colname*	str: Column header that contains experiment id that maps expression data and metadata
metadata_sample_colname*	str: Column header that contains sample id that maps expression data and metadata
project_id*	int: If using template-based approach, experiment id to use as template experiment

For guidance on setting VAE training prameters, see configurations used in simulate-expression-compendia and generic-expression-patterns repositories

Acknowledgements

We would like to thank Marvin Thielk for adding coverage to tests and Ben Heil for contributing code to add more flexibility.