Platelet-Based Liquid Biopsies through the Lens of Machine Learning.

Area of Science:

• Oncology
• Bioinformatics
• Genomics

Background:

• Liquid biopsies provide a minimally invasive method for cancer diagnosis and monitoring.
• Analyzing sequencing data from liquid biopsies with machine learning is complex and requires rigorous validation.
• Key validation challenges include large patient cohorts, sample collection bias, and model interpretability.

Purpose of the Study:

• To evaluate the efficacy of using RNA sequencing data from tumor-educated platelets (TEPs) for binary cancer classification (cancer vs. no-cancer).
• To develop and validate robust machine learning models for cancer detection using TEPs.
• To identify key features and splice variants contributing to accurate cancer classification.

Main Methods:

• Compiled a large-scale dataset of RNA sequencing data from over a thousand donors.
• Employed convolutional neural networks (CNNs) and boosting methods for binary classification.
• Utilized Kyoto Encyclopedia of Genes and Genomes (KEGG) for splice variant analysis and boosting algorithms for feature selection.
• Validated model robustness on independent test data from new hospitals.

Main Results:

• Achieved an area under the curve (AUC) of 0.96 for cancer classification.
• Identified specific splice variant clusters and key predictive features using machine learning.
• Demonstrated consistent model performance on novel, independent hospital test data, indicating robustness.

Conclusions:

• RNA sequencing data from tumor-educated platelets (TEPs) show significant potential for non-invasive cancer patient classification.
• The developed machine learning models are robust and accurate, paving the way for advanced cancer diagnostics.
• This approach offers a promising avenue for profound cancer diagnostics through liquid biopsy analysis.