Correlation-based feature selection of single cell transcriptomics data from multiple sources

• Published in: Journal of big data 2025-01, Vol.12 (1), p.4-23, Article 4
• Main Authors: Mitić, Nenad S., Malkov, Saša N., Maljković Ružičić, Mirjana M., Veljković, Aleksandar N., Čukić, Ivan Lj, Lin, Xin, Lyu, Minjie, Brusić, Vladimir
• Format: Article
• Language: English
• Subjects: Algorithms; Big Data and Artificial Intelligence in Emerging Scientific Fields; Bioinformatics; Classification; Clustering; Communications Engineering; Computational Science and Engineering; Computer Science; Data mining; Data Mining and Knowledge Discovery; Database Management; Feature selection; High-dimensional data; Information Storage and Retrieval; Machine learning; Mathematical Applications in Computer Science; Multidimensional data; Multidimensional methods; Networks; Prediction models; Transcriptomics data
• ISSN: 2196-1115; 2196-1115
• DOI: 10.1186/s40537-024-01051-z

When applying data mining or machine learning techniques to large and diverse datasets, it is often necessary to construct descriptive and predictive models. Descriptive models are used to discover relationships between the attributes of the data while predictive models identify the characteristics of the data that will be collected in the future. Bioinformatics data is high-dimensional, making it practically impossible to apply the majority of “classical” algorithms for classification and clustering. Even if the algorithms are useful, training with large multidimensional data significantly increases processing time. The algorithms specialized for working with high-dimensional data often cannot process data containing large data sets with several thousand dimensions (features). Dimension reduction methods (such as PCA) do not provide satisfactory results, and also obscure the meaning of the original attributes in the data. For the constructed models to be usable, they must fulfill the requirement of scalability, as the amount of bioinformatics data is increasing rapidly. Furthermore, the significance of individual data features can differ from source to source. This paper describes an attribute selection method for efficient classification of high-dimensional (30,698) transcriptomics data collected from different sources. The proposed method was tested with 22 classification algorithms. The classification results for the selected attribute sets are comparable to the results for the complete attribute set.