Manatee: a Variational Autoencoder Model for Predicting Transcription Factor Perturbation-Induced Transcriptomes

Overview

Manatee variational autoencoder (VAE) model predicts transcription factor (TF) perturbation-induced transcriptomes. The workflow of Manatee is shown in the following schematics. Specifically:

During training, the Manatee latent space is constrained to approximate TF expression with an additional TF loss term.

The Manatee in silico perturbation prediction start with adjusting the latent space. For TFs to be up- and down-regulated, corresponding latent values are sampled from the top and bottom Q quantile of the given reference, respectively. Adjusted profiles are subsequently flowed through the trained decoder for final results.

To assess if a perturbation yields the desired transcriptomic phenotype, the relative strength among r(Perturb, End), r(Start, End) and r(Perturb, Start) Pearson correlations, quantified by t-statistics, is calculated. The assessment plot is used to distinguish ideal, compromised (under) and non-significant (N.S.) groups. Start, End and Perturb represent the original, target and perturbed groups, respectively

Data Processing

We use the GSE72857 dataset as an example to show the Manatee functions.

Assuming we are at the root directory of the Manatee repository. The rawdata is in ./GSE72857/processed/. Further data processing for sections In Silico Perturbation and In Silico Screening can be done by running processed.R. The yielded data files are provided in ./GSE72857/perturb/ and ./GSE72857/screen/ respectively for reproducing our results.

Model Training:

vae=./src/train_vae.py
job=GSE72857
data_path=./GSE72857/processed/data_x.csv.gz
gene_path=./GSE72857/processed/genes.txt
tf_path=./GSE72857/processed/tfs.txt
out_dir=./GSE72857/model/

python3 $vae --job=$job --data_path=$data_path --gene_path=$gene_path --tf_path=$tf_path --out_dir=$out_dir --cv=5 --lr=1e-4 --depth=3 --alpha=0.8

A pre-trained model can be found here. You will need to download it into ./GSE72857/model/ to run the following code chunks.

Model Benchmarking:

We examine whether Manatee is able to capture biological information, by benchmarking its two modes:

The “predict” mode encodes original transcriptomes (X), reparametrizes the latent space as decoder inputs Z*, and reconstructs X’.

The “generate” mode, on the other hand, directly decodes X’ from Z*.

vae=./src/train_vae.py
gene_path=./GSE72857/processed/genes.txt
tf_path=./GSE72857/processed/tfs.txt
model_path=./GSE72857/model/GSE72857best_fold.pt
x_path=./GSE72857/processed/data_x.csv.gz
z_path=./GSE72857/processed/data_z.csv.gz
out_dir=./GSE72857/benchmark/

python3 $vae --job=basic --mode=predict --data_path=$x_path --gene_path=$gene_path --tf_path=$tf_path --model_path=$model_path --out_dir=$out_dir --depth=3
python3 $vae --job=basic --mode=generate --data_path=$z_path --gene_path=$gene_path --tf_path=$tf_path --model_path=$model_path --out_dir=$out_dir --depth=3

The result shows that both modes could recapitulate original transcriptomes (X'=X):

In Silico Perturbation

We use Manatee to model the hematopoietic CMP to GMP VS MEP development, which is driven by the Gata1-Spi1 TF module:

vae=./src/train_vae.py
gene_path=./GSE72857/processed/genes.txt
tf_path=./GSE72857/processed/tfs.txt
model_path=./GSE72857/model/GSE72857best_fold.pt
z_path=./GSE72857/perturb/z_perturb.csv.gz
out_dir=./GSE72857/perturb/

python3 $vae --job=perturb --mode=generate --data_path=$z_path --gene_path=$gene_path --tf_path=$tf_path --model_path=$model_path --out_dir=$out_dir --depth=3

The result shows the recapitulation of the lineage bifurcation:

In Silico Screening

We leverage Manatee for the in silico screening of perturbations that could yield the target transcriptomic phenotype. As a proof-of-concept, we screened alternative TF duos driving the above hematopoiesis process. Without losing generality, we screened all 25 combinations of the top 5 highly expressed TFs within MEP and GMP:

vae=./src/train_vae.py
gene_path=./GSE72857/processed/genes.txt
tf_path=./GSE72857/processed/tfs.txt
model_path=./GSE72857/model/GSE72857best_fold.pt
z_path=./GSE72857/screen/z_screen.csv.gz
out_dir=./GSE72857/screen/

python3 $vae --job=screen --mode=generate --data_path=$z_path --gene_path=$gene_path --tf_path=$tf_path --model_path=$model_path --out_dir=$out_dir --depth=3

The result shows the Gata1-Cebpa duo to be an alternative lineage controlling module. Noticeably, Cebpa has been experimentally validated as a master regulator controlling the identity of GMP:

A Comprehensive Model for Mouse Development

We speculate the great value of “comprehensive models”, which represent all major biological processes related to a certain species. We thus provide a Manatee model trained from the TOME mouse embryogenesis single cell dataset, which contains 468 cell populations and 113 lineages. Such a model can be used for mouse in silico perturbation and in silico screening analyses, as well as for further algorithmic developments.