Machine learning for pan-cancer classification based on RNA sequencing data.

Area of Science:

• Oncology
• Bioinformatics
• Genomics

Background:

• Cancers of unknown primary (CUP) account for 2%-5% of malignancies, posing diagnostic and treatment challenges.
• Accurate tissue-of-origin (TOO) determination is crucial for selecting optimal cancer therapies.
• Large-scale cancer genomics initiatives generate vast datasets for developing predictive models.

Purpose of the Study:

• To review and assess machine learning (ML) methods for predicting cancer tissue-of-origin (TOO) using RNA sequencing data.
• To evaluate the reproducibility, interpretability, and robustness of ML-based TOO prediction.
• To identify potential issues in datasets and suggest improvements for clinical utility.

Main Methods:

• Systematic review of 20 recent studies employing ML for TOO prediction.
• Analysis of ML methods based on RNA sequencing data.
• Evaluation of model performance on independent datasets and feature identification.

Main Results:

• Assessed the performance, reproducibility, and interpretability of various ML approaches for TOO prediction.
• Identified strengths and weaknesses of different ML methods and their underlying datasets.
• Highlighted potential challenges in dataset quality and model generalizability.

Conclusions:

• Machine learning shows promise for predicting cancer tissue-of-origin from RNA sequencing data.
• Further research is needed to enhance model robustness, interpretability, and clinical applicability.
• Addressing dataset limitations is crucial for reliable TOO prediction in clinical settings.