Multi-task spatial distillation reveals cell-type-resolved programmed cell death landscapes in the human kidney.

Area of Science:

• Spatial transcriptomics
• Computational biology
• Renal pathology

Background:

• Kidney injury and chronic kidney disease involve spatially varied programmed cell death (PCD).
• Current methods struggle to simultaneously analyze cell composition, PCD activity, and spatial distribution in spatial transcriptomics (ST) data.

Purpose of the Study:

• To develop CoDeST, a novel multi-task spatial inference framework.
• To jointly predict kidney cell-type composition and PCD program activity (apoptosis, pyroptosis, necroptosis, ferroptosis) within ST data.

Main Methods:

• CoDeST employs a three-stage pseudo-to-real training strategy.
• This includes self-supervised pretraining, supervised deconvolution on pseudo-spots, and real-ST domain adaptation using teacher-student distillation and spatial regularization.

Main Results:

• CoDeST demonstrated competitive cell-type proportion recovery in pseudo-spot benchmarks.
• Analysis on real kidney ST data confirmed its ability to transfer deconvolution signals.
• The framework achieves stable deconvolution and PCD mapping, balancing spatial coherence and boundary contrast.

Conclusions:

• CoDeST offers a specialized framework for kidney research.
• It enables joint spatial mapping of cell composition and PCD programs.
• This facilitates deeper insights into kidney injury and disease progression.