Structure–Function Mapping of Therapeutic Membrane-Active Peptides Using an Interpretable Fourier Transform-Based Sequence Encoding and Feature Selection

Introduction: : Traditional antibiotics and small molecule-based therapies often encounter resistance and toxicity issues, prompting the search for novel alternatives such as membrane-active peptides (MAPs). MAPs, including cell-penetrating peptides (CPPs) and membrane-disrupting peptides (MDPs), have shown promise due to their ability to selectively target microbial or cancer cell membranes. However, challenges remain in ensuring their therapeutic specificity and minimizing side effects such as hemolysis. Computational prediction methods can accelerate peptide discovery by identifying functional and safe candidates more efficiently than traditional experimental approaches. While prior models have utilized features like amino acid composition and dipeptide frequency, they often fail to capture crucial sequence-order information that influences peptide functionality. To address this gap, we developed a novel Fourier transform-based encoding method to analyze the oscillations of amino acid properties in peptide sequences. Our machine learning models, particularly nonlinear support vector machines (SVMs), leverage these Fourier-derived features to predict the functional activity and safety profiles of diverse MAP classes with improved accuracy and interpretability compared to traditional or deep learning models.

Materials and

Methods: : Datasets for various MAP classes—including antimicrobial (ABP-GN, ABP-GP), hemolytic (HemoPI-2, HemoPI-3), antiviral (AVP), antifungal (AFP), anticancer (ACP), anti-parasitic (APP), cell-penetrating (CPP), and mammalian-targeting peptides (TMP)—were curated from established repositories and prior literature. Fourier transforms were applied to peptide sequences to extract oscillation features from amino acid physicochemical properties (BLOSUM100 components), capturing sequence-order information. To improve both model performance and interpretability, we integrated MISTIC (Model Informed Selection Through Importance and Contribution), a custom feature selection framework designed to identify and prioritize the most informative features. Nonlinear SVM models were then trained using the MISTIC-selected feature sets to optimize predictive accuracy while minimizing model complexity. Model evaluations included cross-validation and independent testing. Comparisons were made with traditional models using amino acid composition features and deep learning approaches based on evolutionary scale modeling (ESM2) embeddings. Clustering analyses (hierarchical and circular clustering) were performed to explore the biological relationships between peptide classes based on feature rankings. Dendrograms and heatmaps visualized these relationships, providing deeper biological insights into sequence-function correlations critical for therapeutic peptide design.

Results, Conclusions, and Discussions:: The Fourier-based machine learning models demonstrated strong predictive performance for classifying membrane-active peptides (MAPs). As shown in Figure 1, area under the ROC curve (AUC) scores ranged from 0.8466 to 0.9564, using only 59 to 144 Fourier-derived features, underscoring the efficiency of our feature selection strategy compared to larger, less interpretable models. As shown in Figure 2, ESM2-based models achieved similar accuracy (p = 0.9419, paired t-test) but required over 8 million parameters, whereas our models used fewer than 200 features—achieving comparable performance with substantially lower complexity and improved interpretability. Importantly, the models retained high transparency, enabling clear identification of sequence-order features associated with peptide activity. Clustering analysis of the top 20 selected Fourier-based features revealed biologically meaningful groupings, as shown in Figure 3. For example, antibacterial peptides (ABP-GN and ABP-GP) clustered closely due to shared features involved in bacterial membrane disruption. Hemolytic peptides (HemoPI-2 and HemoPI-3) also clustered together, likely driven by hydrophobic and aromatic residues facilitating red blood cell lysis. Furthermore, anti-cancer (ACP) and anti-parasitic peptides (APP) grouped together, reflecting their shared targeting of complex, acidic, or densely packed membranes in higher organisms. In contrast, mammalian-targeting peptides (TMPs) formed a distinct cluster, highlighting their specialized features for interacting with healthy, non-pathogenic mammalian cells. Our feature selection and ranking framework substantially improved model efficiency without sacrificing biological relevance. These findings highlight the utility of Fourier-based sequence encoding for designing specific, safe therapeutic peptides. The resulting structure–function fingerprints offer a powerful foundation for rational peptide engineering, enabling the development of targeted therapies while minimizing off-target effects such as cytotoxicity and hemotoxicity. Overall, this work advances predictive modeling and peptide design by providing an interpretable and biologically grounded computational framework.

Acknowledgements and/or References (Optional)::