Increasing prediction accuracy of pathogenic staging by sample augmentation with a GAN

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in oncology

Background:

• Accurate cancer stage prediction is crucial for effective patient treatment.
• Machine learning, particularly deep learning, shows promise for cancer stage prediction using high-throughput genomic data.
• A significant challenge is the need for large datasets, often unavailable due to the high cost and time of clinical sample acquisition, especially when dealing with over 10,000 genes.

Purpose of the Study:

• To address the limitation of small clinical sample sizes in cancer stage prediction.
• To develop and evaluate a novel data augmentation method using Generative Adversarial Networks (GANs) for genomic data.
• To improve the accuracy of cancer stage prediction models by increasing the effective dataset size.

Main Methods:

• Selected significant genes from DNA mutation data using random forest feature ranking.
• Augmented RNA expression data for selected genes using Generative Adversarial Networks (GANs).
• Compared classification performance using original and augmented datasets with Random Forest, Deep Neural Networks (DNNs), and 1-Dimensional Convolutional Neural Networks (1DCNN).

Main Results:

• Data augmentation with GANs significantly improved classification accuracy, with a 39% increase in F1 score using 1DCNN with a 5-fold data increase (GAN5).
• Results using only 30% of the augmented data outperformed models trained on the entire original dataset.
• The study represents the first use of GANs for augmenting numeric DNA and RNA data for cancer stage prediction.

Conclusions:

• Generative Adversarial Networks (GANs) combined with 1-Dimensional Convolutional Neural Networks (1DCNN) enable accurate cancer stage prediction even with limited clinical samples.
• The proposed sample augmentation method effectively increases classification accuracy and has the potential to reduce the cost and time associated with clinical sample collection.
• This data augmentation technique is applicable to other cancer-related tasks, including prognostic prediction and classification.