Deep learning-based non-invasive profiling of tumor transcriptomes from cell-free DNA for precision oncology

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Circulating tumor DNA (ctDNA) profiling via liquid biopsies is a minimally invasive cancer diagnostic tool.
• Current ctDNA methods for gene expression inference lack transcriptome-wide resolution and require specialized or deep sequencing.

Purpose of the Study:

• Introduce Triton for cfDNA fragmentomic and nucleosome profiling.
• Introduce Proteus, a deep learning framework, to predict gene expression from cfDNA using standard whole genome sequencing.
• Demonstrate the clinical utility of these tools for precision oncology.

Main Methods:

• Jointly developed Triton for comprehensive cfDNA fragmentomic and nucleosome profiling.
• Utilized Proteus, a multi-modal deep learning framework, with standard depth whole genome sequencing data.
• Validated Proteus using patient-derived xenografts and four patient cohorts with matched tumor RNA-Seq.

Main Results:

• Proteus accurately reproduced gene expression profiles from ctDNA, comparable to RNA-Seq replicates.
• The model successfully predicted prognostic markers, phenotype markers, and therapeutic targets in patient cohorts.
• Proteus accurately predicted gene pathway enrichment scores, demonstrating analog utility to RNA-Seq.

Conclusions:

• Triton and Proteus enable transcriptome-wide gene expression profiling from cfDNA using standard whole genome sequencing.
• These tools offer a non-invasive solution for precision oncology, including cancer monitoring and therapeutic guidance.
• The findings highlight the potential of cfDNA-based gene expression analysis in clinical settings.