Generative pretraining from large-scale transcriptomes for single-cell deciphering.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Exponential growth of single-cell transcriptome data presents assimilation challenges.
• Need for advanced computational methods to extract meaningful biological insights.

Purpose of the Study:

• Introduce generative pretraining from transcriptomes (tGPT) for effective transcriptome feature representation learning.
• Evaluate tGPT's performance on single-cell analysis tasks and its application to bulk tissues.

Main Methods:

• Developed tGPT, an autoregressive model predicting gene ranking based on neighboring genes.
• Trained tGPT on 22.3 million single-cell transcriptomes.
• Validated tGPT using four diverse single-cell datasets and bulk tissue samples.

Main Results:

• tGPT-derived single-cell clusters and lineage trajectories show high alignment with known cell states.
• Learned feature patterns from tumor bulk tissues correlate with genomic alterations, prognosis, and immunotherapy outcomes.
• Demonstrated tGPT's utility in integrating and deciphering large-scale transcriptome datasets.

Conclusions:

• tGPT offers a novel analytical paradigm for massive transcriptome data integration and interpretation.
• Facilitates enhanced understanding and clinical translation of single-cell transcriptome data.
• Enables deeper insights into cellular heterogeneity and disease mechanisms.