The Overlooked Role of Specimen Preparation in Bolstering Deep Learning-Enhanced Spatial Transcriptomics Workflows

Area of Science:

• Integrative genomics and computational pathology
• Deep learning applications in spatial transcriptomics

Background:

• Deep learning (DL) applied to spatial transcriptomics (ST) shows promise for linking gene expression to tissue architecture in disease.
• Current ST methods face challenges with tissue preparation variability, high costs, and limited scalability, hindering broader adoption.
• Enhanced specimen processing is needed to improve assay reliability, resolution, and scalability for DL-based ST.

Approach:

• Investigated an enhanced specimen processing workflow for DL-based ST assessment using the Visium CytAssist assay.
• The workflow included automated H&E staining, whole-slide imaging at 40x resolution, and multiplexing of patient tissue sections.
• Compared DL models (Inceptionv3) trained on enhanced versus traditional ST workflows using colorectal cancer (CRC) patient cohorts.

Key Points:

• The enhanced workflow significantly outperformed the traditional ST workflow in DL model performance.
• Gene expression profiles predicted from enhanced tissue slides showed greater topological consistency with ground truth.
• This led to improved statistical precision in identifying biomarkers associated with distinct spatial structures in CRC.

Conclusions:

• The enhanced specimen processing workflow improves DL-based ST accuracy and reliability for biomarker discovery.
• This approach can enhance diagnostic and prognostic biomarker detection, potentially linking to metastasis and recurrence.
• Collaboration between histotechnicians, pathologists, and genomics specialists is crucial for advancing spatial transcriptomics in cancer research.