Learning embedding features based on multisense-scaled attention architecture to improve the predictive performance...

Area of Science:

• Biochemistry and Molecular Biology
• Bioinformatics and Computational Biology
• Cancer Research

Background:

• Anticancer peptides (ACPs) show promise as targeted cancer therapeutics.
• Current machine learning predictors for ACPs rely on experience-based feature engineering, limiting adaptability and performance.
• Existing methods lack robustness and struggle with diverse datasets.

Purpose of the Study:

• To introduce ACPred-LAF, a novel deep learning predictor for identifying anticancer peptides.
• To overcome limitations of traditional feature engineering in ACP prediction.
• To enhance the accuracy, adaptivity, and robustness of ACP identification models.

Main Methods:

• Developed a novel deep-learning-based predictor, ACPred-LAF.
• Proposed a multisense and multiscaled embedding algorithm for automatic feature extraction.
• Constructed a new benchmark dataset, ACP-Mixed, by integrating existing ACP datasets to mitigate evaluation bias.

Main Results:

• Learned, self-adaptive embedding features outperform hand-crafted features in capturing discriminative information for ACP prediction.
• ACPred-LAF demonstrates superior performance compared to state-of-the-art methods on benchmark and newly constructed datasets.
• Model robustness was validated through data interference experiments.

Conclusions:

• ACPred-LAF offers a more adaptive and robust approach to anticancer peptide prediction.
• The novel embedding algorithm effectively captures sequential characteristics of ACPs.
• The ACP-Mixed dataset and ACPred-LAF web server facilitate future research and application in cancer therapy.