STEAM: Spatial Transcriptomics Evaluation Algorithm and Metric for clustering performance.

Area of Science:

• Spatial biology
• Genomics
• Computational biology

Background:

• Spatial transcriptomics enables gene expression analysis within tissue context.
• Validating clustering in spatial omics is challenging due to the lack of ground truth labels.
• A computational framework is needed for unbiased assessment of clustering performance.

Purpose of the Study:

• To introduce the Spatial Transcriptomics Evaluation Algorithm and Metric (STEAM) pipeline.
• To provide a user-friendly computational tool for evaluating clustering consistency and reliability in spatial omics data.
• To offer actionable guidance for refining spatial omics data clusters.

Main Methods:

• STEAM utilizes machine learning classification and prediction to maintain spatial proximity and gene expression patterns.
• The pipeline enables iterative correction of misclassified cells.
• STEAM was benchmarked on diverse public datasets (multicell to single-cell resolution, normal/diseased tissues, spatial transcriptomics/proteomics).

Main Results:

• STEAM demonstrated robustness and generalizability across various spatial omics datasets.
• Performance was evaluated using metrics like Kappa score, F1 score, accuracy, adjusted rand index, and normalized mutual information.
• STEAM supports multisample training for cross-replicate consistency assessment and compares multiple clustering approaches.

Conclusions:

• STEAM is a valuable tool for evaluating clustering robustness in spatial omics data.
• It aids in benchmarking different clustering methods, including spatial-aware and spatial-ignorant strategies.
• STEAM facilitates reproducible discoveries in the field of spatial biology.