sCellST predicts single-cell gene expression from H& E images.

Area of Science:

• Computational biology
• Histopathology
• Genomics

Background:

• Understanding spatial cell organization in tissues is crucial for biology and medicine.
• Hematoxylin and eosin (H&E) slides offer morphological context, while spatial gene expression profiling provides molecular data but is costly and inaccessible.
• Current methods for predicting gene expression from images use small patches, limiting resolution and fine-grained morphological analysis.

Purpose of the Study:

• To develop a deep learning approach for predicting single-cell gene expression directly from histological images.
• To overcome the limitations of existing patch-based methods in capturing fine-grained morphological variations.
• To enable molecular-level interpretation of standard histological slides at scale.

Main Methods:

• A novel deep learning model was developed to predict single-cell gene expression based on cell morphology from H&E-stained images.
• The model's performance was evaluated against patch-based methods on spot-level prediction tasks.
• The approach was applied to two cancer datasets to assess its ability to recover biological expression patterns.

Main Results:

• The deep learning model achieved performance comparable to patch-based methods on spot-level prediction.
• The model successfully recovered biologically meaningful gene expression patterns within the analyzed cancer datasets.
• The approach demonstrated the capability to distinguish between fine cell populations based on morphology.

Conclusions:

• This deep learning approach enables accurate prediction of single-cell gene expression from standard histological images.
• It offers a scalable solution for molecular-level interpretation of histological data, complementing existing spatial profiling techniques.
• The method provides new avenues for studying tissue organization and cellular diversity in both health and disease contexts.