Deep learning-based clustering method for single-cell RNA data

This study proposes innovative ideas in feature engineering for the clustering problem of single-cell RNA data and provides new technical paths for existing clustering methods. Through the proposed feature engineering techniques, we effectively address the curse of dimensionality in the clustering analysis of single-cell RNA data. We introduce an Auto Encoder network that facilitates the reduction and reconstruction of data dimensionality. Meanwhile, this project proposes a novel hidden layer selection method and further utilizes the ISOMAP technique to achieve deeper dimensionality reduction. This study compares two clustering technology paths: one based on cell data and the other on cell feature vectors. It achieves innovation in clustering technology paths by introducing the OPTICS algorithm as a supplement to the k-means algorithm, introducing UMAP and t-SNE for dimensionality reduction, and by exploring the MDS algorithm for low-dimensional data embedding. By introducing internal and external clustering performance evaluation metrics, this thesis verifies that the technical combination of the proposed MDS dimensionality reduction algorithm combined with the k-means method constitutes an innovative clustering technology path that outperforms the traditional methods, and promotes the innovation and improvement of clustering analysis of single-cell RNA data.