MADS-box gene classification in angiosperms by clustering and machine learning approaches

Abstract

The MADS-box gene family is an important transcription factor family involved in floral organogenesis. The previously proposed ABCDE model suggests that different floral organ identities are controlled by various combinations of classes of MADS-box genes. The five-class ABCDE model cannot cover all the species of angiosperms, especially the orchid. Thus, we developed a two-stage approach for MADS-box gene classification to advance the study of floral organogenesis of angiosperms. First, eight classes of reference datasets (A, AGL6, B12, B34, BPI, C, D, and E) were curated and clustered by phylogenetic analysis and unsupervised learning, and they were confirmed by the literature. Second, feature selection and multiple prediction models were curated according to sequence similarity and the characteristics of the MADS-box gene domain using support vector machines. Compared with the BindN and COILS features, the local BLAST model yielded the best accuracy. For performance evaluation, the accuracy of Phalaenopsis aphrodite MADS-box gene classification was 93.3%, which is higher than 86.7% of our previous classification prediction tool, iMADS. Phylogenetic tree construction – the most common method for gene classification yields classification errors and is time-consuming for analysis of massive, multi-species, or incomplete sequences. In this regard, our new system can also confirm the classification errors of all the random selection that were incorrectly classified by phylogenetic tree analysis. Our model constitutes a reliable and efficient MADS-box gene classification system for angiosperms.