Related Experiment Video
Updated: May 28, 2025
Cell-Lineage Guided Mass Spectrometry Proteomics in the Developing Frog Embryo
Published on: April 21, 2022
X-scPAE: An explainable deep learning model for embryonic lineage allocation prediction based on single-cell
Kai Liao1, Bowei Yan2, Ziyin Ding3
1The Central Laboratory of Birth Defects Prevention and Control, The Affiliated Women and Children's Hospital of Ningbo University, Ningbo, 315021, China; Ningbo Key Laboratory for the Prevention and Treatment of Embryogenic Diseases, The Affiliated Women and Children's Hospital of Ningbo University, Ningbo, 315021, China; Ningbo Key Laboratory of Genomic Medicine and Birth Defects Prevention, The Affiliated Women and Children's Hospital of Ningbo University, Ningbo, 315021, China; MOE Engineering Research Center of Gene Technology, School of Life Sciences, Fudan University, Shanghai, 200433, China.
This study introduces X-scPAE, a novel deep learning model for predicting cell lineage allocation using single-cell transcriptomics. It accurately identifies key genes, aiding in understanding cell differentiation and potentially reducing miscarriages.
Area of Science:
- Genomics
- Computational Biology
- Developmental Biology
Background:
- Accurate cell lineage prediction is vital in single-cell transcriptomics for understanding differentiation.
- Identifying lineage-specific gene expression differences is crucial for developmental biology and early pregnancy loss research.
Purpose of the Study:
- To develop an explainable deep learning model for predicting cell lineage allocation in human and mouse single-cell transcriptomic data.
- To identify and interpret key genes responsible for lineage differences and developmental stages.
Main Methods:
- Introduced X-scPAE (eXplained Single Cell PCA - Attention Auto Encoder), a PCA-based deep learning attention autoencoder model.
- Integrated the Counterfactual Gradient Attribution (CGA) algorithm for feature importance calculation.
- Employed an autoencoder with an attention mechanism for feature extraction and PCA for dimensionality reduction.
Main Results:
- X-scPAE achieved high accuracy (0.945 test, 0.977 validation) and outperformed existing baseline and advanced methods.
- Explainability analysis identified key lineage predictor genes in humans and mice.
- A logistic regression model using extracted genes achieved an AUROC of 0.92, surpassing other feature extraction techniques.
Conclusions:
- X-scPAE offers a powerful and explainable approach for cell lineage prediction and key gene identification in single-cell transcriptomics.
- The model's ability to extract biologically relevant genes has implications for understanding cell differentiation and reducing pregnancy loss.
- Ablation studies confirmed the contribution of each component to the model's overall effectiveness.
