An Augmented Sample Selection Framework for Prediction of Anticancer Peptides.

Area of Science:

• Biotechnology
• Computational Biology
• Oncology

Background:

• Anticancer peptides (ACPs) show promise for cancer therapy.
• Traditional ACP identification is inefficient and costly.
• Deep learning models offer potential for ACP prediction but require large datasets.

Purpose of the Study:

• To address the limitations of small datasets and noisy augmented data in ACP prediction.
• To propose a novel augmented sample selection framework (ACPs-ASSF) for enhancing ACP prediction models.

Main Methods:

• Train an initial prediction model on raw data.
• Generate augmented samples and use the model to estimate prediction uncertainty and pseudo-labels.
• Select high-confidence, low-uncertainty augmented samples consistent with original labels for retraining.

Main Results:

• The proposed ACPs-ASSF framework significantly improved prediction accuracy.
• Accuracy gains of up to 5.41% on the ACP240 dataset and 5.68% on the ACP740 dataset were observed compared to traditional data augmentation.
• The method effectively filters noisy augmented samples, enhancing model generalization.

Conclusions:

• ACPs-ASSF offers an effective strategy to improve ACP prediction accuracy by intelligently selecting augmented data.
• This framework overcomes the challenge of limited training data and noisy samples in deep learning for ACP identification.
• The improved prediction accuracy holds potential for accelerating the discovery of novel anticancer peptides for clinical applications.