Transfer learning with convolutional neural networks for cancer survival prediction using gene-expression data

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Precision medicine in oncology relies on integrating diverse data for accurate patient prognosis and personalized treatment.
• Next-generation sequencing (NGS) provides vast gene-expression data (RNA-seq), but analysis is challenged by high dimensionality (many genes) and low sample numbers.
• Standard machine learning methods struggle with gene-expression data for predicting clinical outcomes like patient survival due to overfitting.

Purpose of the Study:

• To develop and evaluate a novel methodology for analyzing gene-expression data using convolutional neural networks (CNNs) for improved cancer survival prediction.
• To investigate the efficacy of transforming RNA-seq data into images for CNN analysis, addressing data's unstructured nature.
• To determine if pre-training CNNs on diverse cancer types enhances lung cancer progression prediction.

Main Methods:

• Pre-trained CNN architectures on gene-expression data from 31 tumor types in the Pan-Cancer dataset.
• Developed a method to convert RNA-seq samples into gene-expression images for CNN processing.
• Fine-tuned the pre-trained CNNs on lung cancer samples to predict progression-free interval.

Main Results:

• The proposed image-based CNN approach shows promise for analyzing gene-expression data.
• Pre-training on a large, multi-cancer dataset potentially improves feature extraction for lung cancer prediction.
• Investigated the benefit of cross-tumor data for enhancing lung cancer progression prediction compared to standard ML.

Conclusions:

• Transforming RNA-seq data into images is a viable strategy for applying CNNs in cancer genomics.
• Leveraging cross-tumor data through pre-training can potentially improve the predictive power of machine learning models for specific cancer types.
• This approach offers a new avenue for advancing personalized medicine in oncology through enhanced data analysis.