LysePred: A Multiscale Convolutional Neural Network for Predicting Hemolytic Activity of Antimicrobial Peptides.

Area of Science:

• Biochemistry and Cheminformatics
• Computational Biology and Drug Discovery

Background:

• Antimicrobial peptides (AMPs) show promise as antibiotic alternatives but often exhibit hemolytic toxicity, hindering clinical use.
• Existing computational methods for predicting hemolytic toxicity face challenges like high complexity and limited pattern recognition.

Purpose of the Study:

• To develop LysePred, an efficient and accurate computational tool for predicting antimicrobial peptide hemolytic toxicity.
• To address limitations in existing prediction methods by capturing multiscale sequence patterns.

Main Methods:

• Developed LysePred, a multiscale convolutional neural network with parallel branches and varied kernel sizes.
• Employed exponentially spaced kernel sizes to capture both local amino acid motifs and longer-range amphipathic patterns.
• Validated performance on six benchmark datasets and the HemoPI2 dataset.

Main Results:

• LysePred achieved top-tier performance, outperforming existing methods in MCC and ACC.
• Demonstrated exceptional stability and computational efficiency with a parsimonious design (~0.55 M parameters).
• Ablation studies confirmed the multiscale architecture's importance, and interpretability analyses showed biologically meaningful pattern learning.

Conclusions:

• LysePred provides a practical, efficient, and interpretable solution for predicting hemolytic toxicity in antimicrobial peptide development.
• The tool facilitates rapid screening, accelerating the discovery of safer antimicrobial peptides.
• LysePred's multiscale approach effectively integrates local and global sequence information for accurate predictions.