Fig. 1: The establishment of iMVP.

a The schema for iMVP. The algorithms selected in our analysis are labeled in red. b Testing dimension reduction algorithms with fly embryo m⁵C dataset. Type I and Type II sites are labeled in blue and red, respectively. PCA, Principal Component Analysis; LLE, Local Linear Embedding; t-SNE, t-distributed Stochastic Neighbor Embedding; UMAP, Uniform Manifold Approximation and Projection; DensMAP, Density-Preserving UMAP. Random states were set to 42. Parallel levels were set to 4 if possible. Algorithm specific parameters were listed in Supplementary Data 7. Metrics: oi outgroup-ingroup score, higher is better, bs boundary score, lower is better. c The peak memory usage and runtime of the algorithms used in b. The bar plot displays the mean and 95% confidence interval, while the precise values of the data points are visualized as dots. n = 3 biologically independent experiments. The time consumed by JIT (Just-In-Time) compiling process in UMAP was not included. This process typically takes ~200 seconds. d Testing clustering algorithms based on the results of UMAP. Simple partitioning algorithms: K-Means and Gaussian Mixture Model (GMM); Hierarchical clustering: Agglomerative hierarchical clustering and BRICH; Graph-based: Spectral clustering, Louvain, and Leiden; Density-based: OPTICS, DBSCAN, and HDBSCAN. Random states were set to 42. Parallel levels were set to 4 if possible. Algorithm specific parameters were listed in Supplementary Data 7. e The peak memory usage and runtime of the algorithms used in d. The bar plot displays the mean and 95% confidence interval, while the precise values of the data points are visualized as dots. n = 3 biologically independent experiments. For Louvain and Leiden, the time (but not memory usage) of the construction of Nearest Neighbor Matrix with UMAP was also included. f The Precision and Recall rate of the clustering algorithms in d. Precision and Recall were computed based on our experimentally validated site information of the dataset.