中文摘要：

目的探讨高维生物学数据的多阶段组合降维策略。方法以微阵列数据的判别分析为例，采用实际数据和模拟数据相结合的方法，提出“初步选维→进→步降维”的两阶段组合降维策略，并与后续的“判别一验证”相结合，形成了“选维→降维→判别→验证”的判别分析思路。以后续判别分析的预测效果、预测结果的稳定性与敏感性等为指标，对2种单→降维（PCA，PLS）方法和4种组合降维方法（PCA＋SIR、PCA＋SAVE、PLS＋SIR和PLS＋SAVE）进行了考察。结果从判别模型的预测效果、预测结果的稳定性及敏感性来看，PLS优于PCA，PLS＋SIR／SAVE的组合降维效果更佳。结论用t计分法选维，以“PLS＋SIR／SAVE”法进行降维的两阶段组合降维策略，对于微阵列数据判别分析，是实用的、可行的。

英文摘要：

Objective To explore multi-stage combinational dimension reduction strategy for analyzing high-dimensional data in biology field. Methods Two-stage combinational strategy incorporated in a four- step procedure, i. e. ＂variable pre-selection→further dimensionality reduction→discrimination→validation＂, was put forward and applied to publicly available microarray data as well as simulated ones. In this process, the rela- tive performances of six dimension reduction methods, including PCA, PLS ,PCA ＋ SIR,PCA ＋ SAVE ,PLS ＋ SIR and PLS ＋ SAVE, were evalua- ted. Results Considering the prediction quality, the stability of the pre- diction results as well as the sensitivity to the number of genes： （ 1 ） PLS performed was superior to PCA; （2） PLS ＋ SIR or PLS ＋ SAVE performed much better than other methods. Conclusion The results indicate that two stage combinational strategy proposed, i. e. variable pre-selection based on t-scores followed by PLS ＋ SIR or PLS ＋ SAVE,is feasible and practical in the discriminate analysis for microarray data.