An image-based approach to prediction of in situ gene expressions at specific loci
10.7644/j.issn.1674-9960.2025.02.002
- VernacularTitle:基于图像的位点原位基因表达预测的方法
- Author:
Liangchen YUE
1
;
Zhen RONG
Author Information
1. 军事科学院军事医学研究院,北京 100850
- Keywords:
deep learning;
B2G;
digital histopathology slide images;
gene expression
- From:
Military Medical Sciences
2025;49(2):90-100
- CountryChina
- Language:Chinese
-
Abstract:
Objective To develop a deep learning algorithm(Bio-section to Gene,B2G)for predicting spatially resolved gene expression profiles directly from histopathological images.Methods Digital histopathological images were processed through an integrated framework comprising convolutional neural networks(CNNs)and the Transformer architecture.Local histological features were extracted by the CNN module while global feature correlations were captured by the Transformer module.Cellular characteristics in digital tissue sections were systematically identified,followed by regression-based prediction of spatially resolved gene expression profiles.Results The B2G algorithm demonstrated significantly higher prediction accuracy than existing methods(weighted median PCC 0.1776).This framework exhibited robust performance across multiple cancer types and histological preparation protocols.Conclusion This computational approach may provide a morphology-driven strategy for spatial transcriptomic analysis.The framework could facilitate cost-effective biomarker discovery in clinical specimens while reducing reliance on specialized molecular techniques.Additionally,it might enable further exploration of tumor microenvironment heterogeneity.
