The required input is the standard Seurat object in .h5seurat format. Download example data.

Overview

DeepGxP is a deep learning model that predicts protein abundances from single-cell RNA-seq gene expression profiles. shinyDeepGxP is an R Shiny app that provides a user-friendly interface to DeepGxP without programming skills. The application offers two complementary modules:

1. Interpret Model: Understand how the model predicts protein abundances
2. Predict Protein: Predict protein abundances from your uploaded scRNA-seq data

For detailed information on data formats or specific module functionality, please visit the corresponding help pages or click the question mark icon in each section.

About DeepGxP

DeepGxP is a deep learning model that predicts surface protein abundance (measured by antibody-derived tags, ADTs) from single-cell RNA expression data. Trained on CITE-seq datasets with paired RNA and surface protein profiles, the model accurately predicts the abundance of 224 surface proteins per cell. This enables multimodal single-cell analysis even when only RNA data are available.

 

ShinyDeepGxP Modules

Module 1: Interpret Model

This module helps you explore the key gene features driving DeepGxP's predictions for each protein. Using the Integrated Gradients interpretation method, the app displays:

1. Key RNA predictors of protein abundance
2. Visualization of gene importance through ranked tables and dot plots
3. Enrichment analyses on top predictors using multiple gene set collections (GO, KEGG, Reactome, etc.)
4. Functional associations and pathways relevant to protein abundance

Module 2: Predict Protein

In this core module, you can upload a Seurat object (.h5Seurat) containing single-cell RNA expression data. Your data will undergo preprocessing and imputation before protein level prediction. The outputs include:

1. Predicted protein abundance (CLR-normalized)
2. Louvain clustering based on RNA expression, predicted protein abundance, and combined modality
3. Reference-based cell type annotation using a pre-annotated PBMC CITE-seq dataset
4. UMAP visualizations for each modality

This module enhances cell type resolution and functional characterization when protein measurements are unavailable.

Additional Features

1. Fully interactive interface for navigating results
2. Integration with Seurat's WNN (weighted nearest neighbor) framework for combined RNA-protein embeddings
3. Downloadable output tables and visualizations

For more details on model architecture, training data, and evaluation metrics, please refer to the DeepGxP manuscript.

---