Impact of Data Quality on Deep Learning Prediction of Spatial Transcriptomics from Histology Images.

Area of Science:

• Computational Biology
• Genomics
• Histopathology

Background:

• Spatial transcriptomics provides gene expression data at tissue locations.
• Deep learning models predict spatial gene expression from histology images due to high costs.
• The impact of data quality on these predictive models is understudied.

Purpose of the Study:

• To investigate how data quality from different spatial transcriptomics technologies (Xenium, Visium) affects deep learning predictions.
• To identify specific aspects of molecular and image data quality that influence predictive performance.
• To compare the impact of data quality versus model architecture on predictive accuracy.

Main Methods:

• Comparative analysis of data quality from Xenium and Visium spatial transcriptomics.
• In silico ablation experiments to assess the impact of molecular data sparsity and noise.
• In silico imputation experiments to evaluate data rescue strategies.
• Assessment of image resolution effects on predictive performance and interpretability.

Main Results:

• Increased sparsity and noise in molecular data significantly degraded predictive performance.
• In silico imputation offered limited improvements and did not generalize well.
• Reduced image resolution negatively impacted predictive performance and model interpretability.
• Data quality improvements proved to be an effective strategy for enhancing predictive models.

Conclusions:

• Data quality is a critical factor in deep learning-based spatial gene expression prediction.
• Improving data quality is an orthogonal strategy to model architecture optimization.
• Consideration of technological limitations affecting data quality is essential for developing robust predictive methodologies.