Self-supervised deep learning for spatial transcriptomics
TRINUS learns the triadic grammar of cell-cell communication — separating intrinsic identity from niche pressure to enable in silico tissue engineering.
The Problem
Spatial transcriptomics maps coordinates — but we lack the computational logic to decipher the rules of cellular interaction.
Graph neural networks fuse a cell with its environment through isotropic smoothing. This mathematical conflation erases the very interaction-induced deviations we seek to measure — denoising data at the cost of biological signal.
Standard GNNs treat the niche as a "bag of neighbors" — missing triadic closure and anisotropic geometry. They cannot distinguish cooperative signaling dependencies from pairwise noise.
Existing tools infer forward correlations only — no framework exists for bidirectional in silico engineering: predicting tissue response and computing the minimal intervention needed.
The Solution
A self-supervised architecture that walks through three innovations — from structured attention to in silico tissue design.
Goes beyond pairwise GNNs to capture cooperative signaling dependencies. Triangle multiplicative updates enforce geometric consistency, distinguishing complex topological motifs invisible to distance-weighted aggregation.
↳ Hover the diagram to see cooperative gating activate
Isolates intrinsic lineage from niche pressure without reference labels. A vector-quantized codebook learns context-free cellular prototypes — discrete anchor points representing pure identity, stripped of environmental noise.
Forward perturbation predicts tissue response to virtual transplants. Inverse design computes the minimal spatial remodeling to rescue a target cell state. A unified differentiable framework for both directions.
↳ Click the diagram to toggle forward / inverse modes
Figure 1 — Hover the architecture components to explore the TRINUS pipeline
Downstream Applications
Decompose observed expression into intrinsic identity and niche-induced deviation.
2 tasksReveal the grammar and vector fields of tissue-wide cell communication.
3 tasksBidirectional simulation: predict tissue response and compute minimal interventions.
2 tasksImpact
Results and significance — each figure panel tells both what we found and why it matters.
TRINUS autonomously learns context-free cell prototypes without reference atlases, separating intrinsic baseline expression from interaction-induced deviations. This enables the resolution of complex spatial topologies.
Reference-free decoupling of cellular lineages and quantification of plasticity via niche pressure.
Moving beyond descriptive mapping, TRINUS models tissue responses to new cellular infiltration (forward) and computes specific molecular edits (inverse) to rescue functions in the ovarian TME.
Unified forward perturbation and inverse design — simulate transplants and compute minimal therapeutic interventions.
Integration of context-agnostic representations from large-scale foundation models (like Geneformer) with TRINUS explicitly endows them with spatial syntax.
Spatial structural context through stacked NicheLayers boosts single-cell foundation model downstream performance.
Get Started
Access the code, reproduce the results, and build upon the triadic framework for your own spatial transcriptomics analysis.
View on GitHub
