A platform-independent AI tumor lineage and site (ATLAS) classifier

Area of Science:

• Computational biology
• Oncology
• Artificial intelligence in medicine

Background:

• Accurate cancer diagnosis is crucial for effective treatment.
• Next-generation sequencing offers molecular insights but lacks independent site-of-origin and lineage assessment.
• Metastatic disease presents classification challenges for existing methods.

Purpose of the Study:

• To develop and validate an AI-driven approach for simultaneously determining cancer site-of-origin and lineage.
• To assess the performance of AI models in both localized and metastatic tumor samples.
• To explore emergent properties of AI lineage scores, including zero-shot learning capabilities.

Main Methods:

• Utilized gradient-boosted machine learning to create separate AI models for tumor lineage and site-of-origin (ATLAS).
• Trained models on RNA expression data from 8249 tumor samples.
• Independently validated performance on 10,376 tumor samples, including 1490 metastatic cases.

Main Results:

• Achieved 91.4% accuracy for cancer site-of-origin and 97.1% for cancer lineage.
• High-confidence predictions demonstrated 98-99% accuracy in both localized and metastatic samples.
• Demonstrated zero-shot learning capabilities, differentiating mesothelioma subtypes and identifying poor-outcome neuroendocrine tumors.

Conclusions:

• The ATLAS AI platform provides accurate and independent assessment of cancer site-of-origin and lineage.
• This approach is platform-independent, easily translatable to existing RNA-seq workflows.
• ATLAS can enhance histopathologic assessment, particularly for tumors of unknown primary and challenging metastatic cases.