利用神经网络集成预测MHC-Ⅰ类分子结合肽

目的：利用神经网络集成（NNE）预测MHC-Ⅰ类分子结合肽。 方法： 基于HLA-A*0201编码的MHC-Ⅰ类分子结合肽数据库（含有628个9聚物）及其结合能力分类，利用NNE分别对具有无、低、中和高4类亲合性的结合肽进行分类预测；同时还进一步利用T细胞真实表位集（含50个表位）评估了NNE的预测性能。 结果： 集成数为12的NNE对上述分类的平均预测命中率可达0.8,而且NNE对潜在T细胞表位的预测能力也较高，约84%的真实表位归于高和中等亲合性的潜在抗原肽一类。 结论： 可以利用神经网络集成预测MHC-Ⅰ类分子结合肽，并进而预测相应的T细胞表位。经适当修改，NNE预测工具可扩展为能涵盖任意长度的Ⅰ类分子结合肽甚至可扩展到Ⅱ类分子结合肽的预测。

Prediction of MHC class Ⅰ binding peptides using neural network ensembles

AIM: To predict MHC class Ⅰ binding peptides by using neural network ensembles. METHODS: As a combination of neural networks, neural network ensemble (NNE) was here used to improve the predictive performance. Based on a database of 628 nonamers and their classified binding capacities, the generalized NNEs were used to classify peptides respectively with non, low, moderate and high binding capacities to MHC class I molecule encoded by gene HLA-A*0201. The predictive power of NNE was further evaluated by running generalized NNE on a set of actual T-cell epitopes. RESULTS: The generalized NNEs achieved an average predictive hit rate of 0.8 for the above classifications. In addition, NNE was also efficient in the prediction of the potential T-cell epitopes, and about 84% of the actual T-cell epitopes were among the potentially antigenic peptides with high and moderate affinities. CONCLUSION: The NNEs can be applied in the prediction of MHC class Ⅰ binding peptides, and moreover, after proper modifications, they can be conveniently extended to cover peptides with any length and thus suitable for the prediction of peptides binding to other MHC class Ⅰ or even class Ⅱ molecules.