scPretrain

scPretrain applies multi-task self-supervised learning to improve cell type classification from single-cell RNA-seq (scRNA-seq) data by leveraging both annotated and unannotated cells.

Key Features:

• Multi-Task Self-Supervised Learning: Employs a multi-task learning framework that integrates self-supervised learning to utilize unannotated scRNA-seq cells.
• Two-Step Training: Implements a pre-training step using pseudo-labels from unannotated cells followed by a fine-tuning step with annotated cells.
• Dataset-Specific Feature Extraction Encoder: Trains a feature extraction encoder specific to each dataset during pre-training.
• Enhanced Classification and Clustering: Incorporates annotated and unannotated data to improve cell type classification accuracy and clustering performance.
• Representation Transferability: Pre-trained representations improve the performance of conventional classifiers including random forests, logistic regression, and support vector machines.

Scientific Applications:

• Cell Type Classification: Facilitates automatic classification of cells based on gene expression profiles from scRNA-seq data.
• Cross-Dataset Evaluation: Validated on 60 diverse datasets spanning different technologies, species, and organs.
• Classifier Enhancement: Produces representations that enhance the efficacy of conventional classifiers such as random forests, logistic regression, and support vector machines.

Methodology:

Pre-training uses pseudo-labels from unannotated cells in a multi-task self-supervised learning framework to train a dataset-specific feature extraction encoder, which is then fine-tuned using annotated cells.