机器学习模型辅助抗菌肽设计

Antimicrobial peptides (AMPs) are a class of functional small peptides with anti-bacterial activity. Research and development (R&D) of AMPs has great significance in both science and industry. AMPs are traditionally discovered and separated from nature by experiments. In contrast, prediction and design of novel AMPs based on precise mathematical models is more efficient, faster and lower-cost, and thus will become a trend of R&D of AMPs. However, due to the extremely complex and non-linear mapping relations between AMP sequence and its activity, it is hard to precisely predict the activity of a given peptide. How does the sequence determine the activity? What are the key features hidden in the sequences with high activity? Such basic scientific questions need to be answered before we design a novel peptide with desired activity. To address these issues in this project, machine learning technologies with strong non-linear mappings for regression, including back propagation (BP) neural network, support vector machine and deep learning, will be employed to build high-performance models for predicting the activity from peptide sequence. Furthermore, design of novel AMPs with high activity and low toxicity will be performed based on the accurate prediction of the well-trained models. The models, methodologies and tools developed in this work are expected to provide a new approach for addressing the issue of a type of common problems (i.e., functional relationship of sequence-activity), and also to provide a theoretical basis and technical platform for R&D of AMPs.

抗菌肽是一类具有抗菌活性的小分子多肽，其研究开发具有重要的学术意义和产业前景。与传统从自然界中发现和分离抗菌肽等实验手段相比，基于模型预测和设计的方法具有高效快捷、成本低廉等特点，是今后抗菌肽研发的趋势。但由于抗菌肽序列及其活性之间存在着异常复杂的非线性映射关系，由序列准确预测其活性仍然十分困难。抗菌肽序列是如何决定其抑菌活性的？高活性序列存在哪些关键特征？这是高活性抗菌肽设计需要解决的关键科学问题。为回答这些问题，本项目拟采用BP神经网络、支持向量机、深度学习等非线性映射能力很强的机器学习算法来构建高效的抗菌肽序列-活性定量预测模型，并基于模型的精准预测，实现高效低毒的新抗菌肽设计。本研究建立的模型、方法和工具有望为一类共性科学问题（即序列-活性之间的函数关联）提供新的解决方案，亦能为抗菌肽产品的研发提供理论依据和平台支持。

新型抗菌肽的设计、开发及应用同时具有重要的学术意义和产业前景。基于定量数学模型的准确预测来设计获得新型抗菌肽可显著提升研发效率和降低成本，是今后抗菌肽研发的趋势。据此，本项目致力于提出一种可用于新抗菌肽预测和设计的高性能机器学习模型。首先以菌丝霉素抗菌肽为例，利用易错PCR和液滴微流控分选技术构建了抗菌肽随机突变文库，筛选获得了12条活性较高的抗菌肽突变序列。然后，针对抗菌肽序列-活性之间异常复杂的非线性映射关系，采用非线性映射能力很强的支持向量机方法来构建高效的预测模型。经过核函数等参数优化设置和交叉验证训练后，其对独立测试集的预测准确率达88.16%，高于已报道的同类模型，表明其具有较好的预测性能。该模型有望应用于发现或设计新的抗菌肽序列。