X-scPAE: An explainable deep learning model for embryonic lineage allocation prediction based on single-cell transcriptomics revealing key genes in embryonic cell development

Abstract

In single-cell transcriptomics research, accurately predicting cell lineage allocation and identifying differences between lineages are crucial for understanding cell differentiation processes and reducing early pregnancy miscarriages in humans. This paper introduces an explainable PCA-based deep learning attention autoencoder model, X-scPAE (eXplained Single Cell PCA - Attention Auto Encoder), which is built on the Counterfactual Gradient Attribution (CGA) algorithm. The model is designed to predict lineage allocation in human and mouse single-cell transcriptomic data, while identifying and interpreting gene expression differences across lineages to extract key genes. It first reduces dimensionality using Principal Component Analysis (PCA) and ranks the importance of principal components. An autoencoder is then employed for feature extraction, integrating an attention mechanism to capture interactions between features. Finally, the Counterfactual Gradient Attribution algorithm calculates the importance of each feature. The model achieved an accuracy of 0.945 on the test set and 0.977 on the validation set, with other metrics such as F1-score, Precision, and Recall all reaching 0.94. It significantly outperformed both baseline algorithms (XGBoost, SVM, RF, and LR) and advanced approaches like F-Score-SVM, CV2-LR, scChrBin, and TripletCell. Notably, the explainability analysis uncovered key lineage predictor genes for both humans and mice and identified crucial genes distinguishing between developmental stages and lineages. A logistic regression model built using the extracted key genes still achieved an AUROC of 0.92, surpassing the performance of other feature extraction methods, including F-Score, CV2, PCA, random feature selection, and the interpretability method Shapley. Lastly, ablation studies demonstrated the effectiveness of each model component.