Benchmarking and Computational framework for Optimal Visualization and Interpretability of high-dimensional Spatiotemporal Data

National Institute Of Environmental Health Sciences

Investigators

Linked publications, trials & patents

Abstract

Single-cell analysis have revolutionized biomedical research, enabling unprecedented insights into cellular heterogeneity, differentiation trajectories, and the mechanisms underlying complex diseases. Despite these advances, the interpretation and visualization of single-cell data remain challenging due to the high dimensionality and complexity of the datasets. Dimensionality reduction methods (DRMs) such as t-SNE, UMAP, PCA, and PHATE and their extensions have been widely adopted to address these challenges by reducing the data to more manageable dimensions for efficient visual interpretability. Each DRM focuses on optimizing specific geometric feature transformation to achieve effective visualization and interpretability (Atitey et al.). For instance, t-SNE emphasizes local structure, while UMAP is known for its ability to separate clusters effectively in high-dimensional data. Both t-SNE and UMAP seem to perform nicely in preserving local structure but struggle to preserve global structure. Regarding PCA, it finds and ranks linear combinations of data points that maximize the data’s variance in terms of the principal component (PCs). However, this projection does not necessarily translate to biological relevance. PHATE focuses on preserving the progression of cell states even if there might not be a clear biological continuity in the data. While these methods provide valuable insights individually, our recent published benchmarking framework called MIBCOVIS (Atitey et al. 2024) demonstrates that their singular focus on specific data features lead to a loss of critical information when applied in isolation. We believe integrating the non-linear outputs of different DRMs provides a better approach of significantly boosting or improving visualization and interpretability of complex processes compared to focusing on individual method improvements. Thus, we recently developed a new algorithm called GIBOOST (Atitey et al. 2025) which has been accepted for publication in Briefings in Bioinformatics. This novel computational tool is designed to enhance the visualization and interpretation of high-dimensional single-cell data. Conditional on a target performance feature of interest e.g. cluster sensitivity, GIBOOST integrates reduced data from different DRMs, each optimizing different features for effective visualization. By doing so, GIBOOST provides a more interpretable, accurate and holistic view of cellular dynamics and interactions, which are often obscured by the limitations of individual DRMs. To synthesize complementary information from different methods and enhance visualization and interpretability, GIBOOST uses the MIBCOVIS framework to identify the most important features optimized by each DRM. These features include high separability, good coverage, uniform spread, time dependency, and cluster sensitivity, each assessed by the following metrics: separability index (SI), occupation index (OI), uniformity index (UI), time order structure index (TI), and gradient boosting classifier index (GI). For any high-dimensional and complex data, GIBOOST selects and integrates the two best data reduction methods from a pool of methods with the maximum additive clustering sensitivity effect conditioned on other visualization features. The tool comprises three main steps: 1. Selection of DRMs: Select a diverse pool of DRMs with different visualization objectives. 2. Optimization Using MIBCOVIS: Select the top two methods that maximize visualization and interpretability using the MIBCOVIS Bayesian framework. The selection is based on the GI feature, which evaluates how effectively the methods maximize cluster sensitivity while considering other features such as OI, SI, UI, and TI. 3. Integration with Autoencoder: Use a GI-optimized autoencoder to integrate the most effective combination of DRM pairs by varying the number of neurons and batch size. GIBOOST selects the optimal number of neurons and batch size that maximizes cluster sensitivity after integrating the data from step 2. To demonstrate GIBOOST’s advantages, we applied it to diverse datasets representing dynamic biological processes including EMT, iPSC reprogramming, and spermatogenesis. Across these datasets, GIBOOST consistently outperformed standalone DRMs, preserving structural transitions such as epithelial-to-mesenchymal progression, intermediate pluripotent states in reprogramming, and developmental stages in spermatogenesis. This consistent performance demonstrates GIBOOST’s capacity to maintain biologically meaningful trajectories and heterogeneity. To further assess GIBOOST’s ability to preserve meaningful biological structure, we evaluated how well local neighborhood relationships were maintained following dimensionality reduction using the trustworthiness score. By varying neighborhood sizes, we found that GIBOOST consistently outperformed leading DRMs like scPHERE, scVI, SIMLR, SAUCIE, and Ensemble UMAP. Beyond local structure, we also compared GIBOOST to these methods in the context of reconstructing biologically relevant trajectories. Specifically, we analyzed the epithelial-mesenchymal transition (EMT), a well-characterized process involving transitions through epithelial (E), partial EMT (pEMT), mesenchymal (M), and partial mesenchymal-epithelial transition (pMET) states. We showed that GIBOOST inferred trajectories closely aligned with the known EMT progression, whereas other methods produced distorted trajectories that misrepresented key transitions, particularly between M and pMET states. Overall, across a comprehensive suite of metrics capturing local neighborhood fidelity, global structure preservation, and trajectory inference accuracy, GIBOOST consistently outperformed current state-of-the-art approaches, demonstrating its robust utility for high-resolution visualization and interpretation of single-cell data. We next applied GIBOOST to placental development data to further demonstrate its utility. Compared to individual DRMs, GIBOOST showed the highest performance in preserving clustering structure, as measured by the Gradient Boosting Classifier Index (GI) and cophenetic correlation. Notably, the integration of PHATE and UMAP within GIBOOST optimized cluster separability and continuity, while cophenetic correlation analysis revealed GIBOOST preserved the highest structural integrity relative to the original data . This enhanced global structural preservation enabled detailed insights into cell-cell communication during placental development. For example, GIBOOST highlighted: (1) the interconnected roles of decidual perivascular cells, decidual stromal cells, and epithelial cells in placental development; (2) the interaction between dendritic cells and T cells, promoting the formation of regulatory T cells; (3) the secretion of chemokines (CXCL12 and CCL2) by trophoblasts, which attract T cells to the placenta; and (4) the stimulation of endothelial cells by fibroblasts. Focusing on specific tissues or organ samples, GIBOOST further identified a range of crucial cell-cell interactions, including: Fibroblast-Endothelial in chorioamniotic membranes, placental villi, and basal plate; NK-Syncytiotrophoblast (SCT) in chorioamniotic membranes, placental villi, and basal plate; T cells-Extravillous Trophoblast (EVT) in placental villi and basal plate; Epithelial-Decidual perivascular cells-Decidual stromal cells in decidua, placenta, and basal plate; Macrophage-Hofbauer cells (HB) in basal plate and placental villi; T cells-NK in decidua, placental villi, and basal plate; EVT-SCT-VCT in chorioamniotic membranes, decidua, placenta, blood, placental villi, and basal plate; and T cells-NK in the blood. These findings, supported by existing experimental studies, validate GIBOOST’s ability to illuminate biologically relevant communication networks at scale. In the future, we plan to extend GIBOOST capabilities to integrate three or more DRMs, increasing flexibility. Also, given the growing variety of interpretability metrics(e.g., silhouette coefficient, trajectory pseudotime score), GIBOOST offers a generalizable framework for enhancing clustering sensitivity in both supervised and unsupervised settings. Furthermore, while deep learning models like autoencoders are powerful, they are stochastic and often lack interpretability. We plan to use methods like GIBOOST to enhance the interpretability nature of neural networks. In summary, GIBOOST offers an innovative and efficient solution for improving the visualization and interpretation of high-dimensional single-cell data. By systematically selecting and integrating complementary DRMs, it maximizes the preservation of local and global biological structure, enabling deeper insights into cellular dynamics and tissue-level communication. Its modular design and superior performance across diverse tasks make it a powerful tool for advancing single-cell analysis.