Tracing unknown tumor origins with a biological-pathway-based transformer model

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Cancer of unknown primary (CUP) presents a diagnostic challenge, with the primary tumor site remaining elusive after standard evaluations.
• Accurate identification of the primary tumor site is crucial for effective patient treatment and management.

Purpose of the Study:

• To develop and validate BPformer, a novel deep learning approach for determining the tissue of origin in metastatic cancers, particularly CUP.
• To integrate transformer models with biological pathway knowledge for enhanced tumor origin prediction.

Main Methods:

• BPformer was developed using transcriptomic data from 10,410 primary tumors across 32 distinct cancer types.
• The method leverages a transformer architecture enhanced with prior biological pathway information.
• Model performance was assessed on primary tumors and both primary and metastatic sites of metastatic tumors.

Main Results:

• BPformer achieved high accuracy rates: 94% for primary tumors, 92% for primary sites of metastatic tumors, and 89% for metastatic sites.
• The method outperformed existing state-of-the-art techniques in tumor origin identification.
• Validation in a retrospective study confirmed consistency with histopathology and immunohistochemistry diagnoses.
• BPformer successfully ranked biological pathways by their contribution to origin identification, distinguishing conserved from variable oncogenic pathways.

Conclusions:

• BPformer represents a significant advancement in diagnosing cancer of unknown primary, offering high accuracy and reliability.
• The method's ability to analyze biological pathways provides insights into cancer biology and potential therapeutic targets.
• BPformer's performance suggests its utility as a valuable tool in clinical oncology for improving diagnostic accuracy and patient outcomes.