Deep Time Point Selector and Profiler (DTPSP)

DTPSP is a machine learning model for selecting optimal time points in single-cell analysis. The goal is to maximize the information learned from single-cell experiments by predicting the most informative time points for data collection. The model uses features derived from gene expression data to inform decision-making.

Fig. 1: DTPSP method overview

Overview

a, Reference time point selection: A subset of time points, S, is selected from temporal bulk gene expression data spanning T time points and sent to a deep learning model to predict gene expression at non-selected time points. Gene expression data from S are augmented with neighboring genes (NG) with similar expression patterns, time embeddings (E), and a target label (L) that specifies the time point to be predicted. The deep regressor model undergoes pretraining with an autoencoder-based reconstruction task before being fine-tuned for regression to predict gene expression. The subset achieving the best prediction performance, SB, is identified as the most informative set of time points for subsequent analyses. b, Single-cell profiling and processing: The selected optimal time points, SB, are subjected to single-cell sequencing experiments to obtain real single-cell data. Standard single-cell processing pipelines are then applied to these data. c, Single-cell inference for unselected time points (temporal single-cell semiprofiling): For non-selected time points, a VAE-GAN model is employed to infer single-cell data matrices. The VAE-GAN is first trained on a reconstruction task to learn the data distribution of the selected time points with real single-cell data. Subsequently, bulk gene expression data are incorporated into the model to transition the process from reconstruction to inference, generating single-cell data for the target time points. This results in a semi-profiled single-cell temporal trajectory where single-cell data are available for all time points. d, Single-cell analysis across all time points: The semi-profiled single-cell temporal trajectory enables comprehensive single-cell analyses, including visualization, deconvolution, enrichment analysis, and more.

Publication

Our research manuscript detailing DTPSP is now available on bioRxiv: DTPSP: A Deep Learning Framework for Optimized Time Point Selection in Time-Series Single-Cell Studies

Key Contributions and Impact

Efficient Time-Point Selection

DTPSP identifies the most informative time points in time-series studies. By prioritizing key time points, DTPSP enables efficient allocation of resources to in-depth profiling methods like multi-omics and single-cell studies, reducing redundancy.

Enhanced Study of Dynamic Systems

The model supports the exploration of dynamic biological systems, such as development, disease progression, and responses to treatments, ensuring that vital temporal dynamics are captured.

Scalable and Versatile

DTPSP provides a scalable framework adaptable to various experimental designs, significantly improving efficiency in systems biology research.

Tech Features

• Simplified Workflow: Trains on public datasets to ensure accessibility and reproducibility.
• Single-Cell Application: Input format is single-cell RNA-seq data.
• Prediction Accuracy: Employs a hybrid neural network model (autoencoder and regressor).
• Visualization: Ability to visualize average MAE and R² values for all shown time points, as well as UMAP single-cell visualizations.
• Example Tutorial: An example notebook covering all steps, from model training and time point selection to single-cell visualization.

Data Preparation

Input Requirements

• Format: Data should be in AnnData's .h5ad format
• Gene Expression Matrices:
  • Genes as variables (adata.var_names)
  • Cells as observations (adata.obs_names)

Example Data

The example_data folder provides data that can be used in the Example_Code.ipynb notebook to train a model, output average MAE and R² values, and visualize single-cell reconstruction. Preprocessing steps include:

• Normalizing
• Thresholding
• Normalizing (again)
• Avergaing across cells
• Log1p transformation

Example Use

To run the example code, please follow these steps:

1. Download the Required Files:

   • Download the Example_Code.ipynb file and the example_data folder.

2. Set Up the Environment:

   • Ensure that Python and all required dependencies are installed. The necessary packages can be installed directly from within the notebook if not already available.

3. Execute the Code:

   • Place the example_data folder in the same directory as the notebook.
   • Open Example_Code.ipynb using Jupyter Notebook, Jupyter Lab, or Google Colab.
   • Run the cells in the notebook sequentially to train the model and generate the outputs.

4. Outputs:

   • The notebook will produce the trained model as well as average MAE and R² values.

These steps ensure reproducibility of the findings described in this work.

API Documentation

An exntensive API for DTPSP is under development.

License

This project is licensed under the GNU General Public License. See the LICENSE file for more details.

Happy Profiling!