過度使用抗生素已導致多重藥物耐藥（MDR）病原體的出現，對公共衛生造成了重大 威脅。從結構和化學上多樣的化合物中衍生出的抗微生物肽（AMPs）為解決這個問題提 供了一個有前途的解決方案。然而，由於流體膜中結構-功能關係的動態性質，鑒定有效 的 AMPs 一直具有挑戰性。機器學習（ML）算法已被應用來應對這一挑戰，我們使用 ML 基於螺旋狀AMP 序列訓練了一個預測模型，以針對特定的細菌進行靶向。已開發了多種 ML 模型來預測 AMPs。然而，這些方法大多主要依賴於氨基酸序列，對於預測針對特定 菌株的抗微生物活性，僅有限地納入了 3D 結構特徵和 ML 模型內的特定相互作用。在本 研究中，我們提出了一個線性 AMP 的預測模型，該模型包括以下重要特徵：（i）透過圖 論將AMP 的 3D 結構進行比較；（ii）根據分子動力學模擬推導出的特定相互作用的實現； （iii）考慮到細菌的負電性特性的 AMP 膜插入模型。此外，我們的預測模型已經開發 出來區分對特定菌株具有活性的肽與其他肽，從而增強了 AMP 開發中的特異性。此外， 我們還使用了低序列識別的肽來評估我們模型的準確性。我們的預測模型可以快速設計 出有效的 AMPs。設計了 7 個潛在的 AMPs 並進行了體外細胞實驗，顯示成功率高達 71 ％ ， 以 提 高 AMPs 的 抗 菌 效 果 。 AMP-META 可 免 費 訪 問 http://ai- meta.chem.ncu.edu.tw/amp-meta;Excessive use of antibiotics has led to the emergence of multidrug-resistant (MDR) pathogens, which pose a significant threat to public health. Antimicrobial peptides (AMPs) derived from structurally and chemically diverse compounds offer a promising solution to combat this problem. However, identifying effective AMPs has been challenging due to the dynamic nature of structure-function relationships in fluid membranes. Machine learning (ML) algorithms have been employed to address this challenge, and we trained a predictive model based on helical AMP sequences using ML to target specific bacteria. Various ML models have been developed to predict AMPs. However, most of these methods have primarily relied on amino acid sequences, with limited incorporation of 3D structural features and specific interactions within ML models to predict antimicrobial potency against specific strains. In this study, we present a predictive model for linear AMPs, which incorporates the following significant features: (i) Graph theory considerations, taking into account the native contacts of 3D AMPs; (ii) Implementation of specific interactions derived from molecular dynamics simulations; and (iii) Incorporation of a membrane-insertion model of AMPs that considers the negatively-charged nature of bacteria. Furthermore, our predictive model has been developed to differentiate peptides that are active against particular strains from others, thereby enhancing the specificity in AMP development. Additionally, low-sequence-identified peptides were employed to evaluate the accuracy of our model. Our predictive models enable rapid engineering of potent AMPs. Designed 7 potential AMPs and conducted in vitro cell experiments, showing a success rate of up to 71% to improve the antibacterial effect of AMPs. AMP-META is freely accessible at http://ai-meta.chem.ncu.edu.tw/amp-meta
