passalis / synthetic-data-genomics

Proof of concept code from Gretel.ai and Illumina using generative neural networks to create synthetic versions of mouse genotype and phenotype data.

Synthetic Data Genomics

The code in this repository uses Gretel.ai's synthetic data APIs to create synthetic (artificial) versions of real world mouse genotype and connected phenotype datasets. We then measure the accuracy of our synthetic data by replicating the results of a Genome Wide Association Study (GWAS) on the real world genotypes and phenotypes for 1,220 mice from this paper: https://doi.org/10.1038/ng.3609.

View the full case study here: https://cdn.gretel.ai/case_studies/gretel_illumina_case_study.pdf

Installation

Requirements:

• Conda package manager
• Ubuntu 18.04 recommended
• NVidia T4 or faster GPU
• Gretel.ai API key (https://console.gretel.cloud)

Install the Conda package manager:

conda create --name genomics python=3.9
conda activate genomics
conda install jupyter

Note that

Recreate the original paper experiments

Follow the steps in EXPERIMENTS.md to download the experiment datasets and recreate the results from the paper using real world data.

Synthesize genome and phenome data, run experiments

Next, create synthetic versions of the mouse phenome and genome datasets from the original experiments.

1. synthetics/01_create_phenome_training_data.ipynb creates the genome training set and filter irrelevant fields.
2. synthetics/02_create_synthetic_mouse_phenomes.ipynb trains a synthetic model on the mouse phenome set.
3. synthetics/03_build_genome_training_set.ipynb creates a genome dataset based on abBMD SNPs
4. synthetics/04_create_synthetic_mouse_genomes.ipynb trains a synthetic model on the mouse genome set, runs GWAS analysis and compares to original results

Additional resources

• research_paper_code/notebooks/05_compare_associations.ipynb compute precision, recall and F1 scores for the final synthetic data
• synthetics/Optional_tune_synthetic_training_params optionally use Optuna to tune synthetic training parameters.
• research_paper_code/notebooks/Manhattan plot.ipynb compute Manhattan plots for both the original and synthetic genome/phenome gwas p-values