KAFSTExp: Kernel Adaptive Filtering With Nyström Approximation for Predicting Spatial Gene Expression From Histology...

Area of Science:

• Biomedical Informatics
• Computational Biology
• Genomics

Background:

• Spatial transcriptomics (ST) is crucial for tumor heterogeneity analysis but is expensive.
• Predicting gene expression from pathology images offers a cost-effective alternative.
• Current deep learning models face generalization challenges with limited ST data.

Purpose of the Study:

• To develop a novel framework, KAFSTExp, for accurate gene expression prediction from pathology images.
• To leverage kernel adaptive filtering (KAF) and foundation models for enhanced spatial transcriptomics analysis.
• To address the limitations of existing deep learning models in handling complex, nonlinear relationships in ST data.

Main Methods:

• Utilized the UNI pathology foundation model for image feature encoding.
• Implemented the kernel least mean square algorithm with Nystrom approximation for gene expression prediction.
• Developed the KAFSTExp framework integrating image analysis with gene expression prediction.

Main Results:

• KAFSTExp significantly improved prediction accuracy for normalized transcript counts.
• The method demonstrated substantial reductions in computational cost and training time.
• Achieved relative improvements in Pearson Correlation Coefficient (PCC) from 1.24% to 94.23% across datasets.

Conclusions:

• KAFSTExp offers a computationally efficient and accurate approach for spatial transcriptomics analysis.
• The framework shows strong generalization performance and clinical application value.
• This method provides a viable and cost-effective alternative to traditional ST examinations.