Transcriptome Transformer: improving patient survival prediction via multitask learning of transcriptomic and...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Accurate patient survival prediction is crucial for guiding treatment strategies and improving outcomes.
• Clinical features offer prognostic information but often miss the molecular complexity of diseases.
• Transcriptomic data provides a complementary view of disease by reflecting gene expression patterns.

Purpose of the Study:

• To introduce Transcriptome Transformer (TxT), a novel multitask learning framework for enhanced patient survival prediction.
• To leverage a transcriptome-centric approach using Transformer architecture to model complex gene-gene interactions.
• To improve survival prediction by jointly analyzing transcriptomic data and clinical features for a comprehensive biological representation.

Main Methods:

• Developed TxT, a multitask learning framework utilizing a Transformer-based architecture with multihead attention mechanisms.
• Employed a transcriptome-centric approach to capture complex gene dependencies and dynamic gene-gene interactions.
• Integrated transcriptomic data with clinical features for joint analysis across multiple prediction tasks.

Main Results:

• TxT outperformed existing methods in survival prediction and related clinical tasks on both single-task and multitask datasets.
• The framework provided biological insights via attention-derived gene interaction networks, highlighting immune pathways in Luminal A patients.
• Differential attention analysis confirmed that integrating clinical features improves the prioritization of biologically relevant genes influencing tumor progression.

Conclusions:

• TxT offers a more complete representation of patient biology, leading to superior survival prediction accuracy.
• The model provides valuable biological insights into disease mechanisms and patient stratification.
• Integrating transcriptomic data with clinical features via advanced AI frameworks like TxT represents a promising direction for precision medicine.