[Paper Review] FEAST: Fully Connected Expressive Attention for Spatial Transcriptomics

Taejin Jeong, Joohyeok Kim+3 more|arXiv (Cornell University)|Mar 26, 2026

Single-cell and spatial transcriptomics0 citations

TL;DR

FEAST models tissue as a fully connected graph using negative-aware attention and off-grid pseudo-spots to predict spatial gene expression from whole slide images, achieving state-of-the-art results and offering interpretable attention maps.

ABSTRACT

Spatial Transcriptomics (ST) provides spatially-resolved gene expression, offering crucial insights into tissue architecture and complex diseases. However, its prohibitive cost limits widespread adoption, leading to significant attention on inferring spatial gene expression from readily available whole slide images. While graph neural networks have been proposed to model interactions between tissue regions, their reliance on pre-defined sparse graphs prevents them from considering potentially interacting spot pairs, resulting in a structural limitation in capturing complex biological relationships. To address this, we propose FEAST (Fully connected Expressive Attention for Spatial Transcriptomics), an attention-based framework that models the tissue as a fully connected graph, enabling the consideration of all pairwise interactions. To better reflect biological interactions, we introduce negative-aware attention, which models both excitatory and inhibitory interactions, capturing essential negative relationships that standard attention often overlooks. Furthermore, to mitigate the information loss from truncated or ignored context in standard spot image extraction, we introduce an off-grid sampling strategy that gathers additional images from intermediate regions, allowing the model to capture a richer morphological context. Experiments on public ST datasets show that FEAST surpasses state-of-the-art methods in gene expression prediction while providing biologically plausible attention maps that clarify positive and negative interactions. Our code is available at https://github.com/starforTJ/ FEAST.

Motivation & Objective

• Motivate improving spatial transcriptomics gene expression prediction by modeling all spot interactions instead of sparse graphs.
• Develop an attention-based FEAST framework with negative-aware attention to capture excitatory and inhibitory biological relationships.
• Mitigate information loss from patch-based spot extraction with off-grid sampling and a hierarchical attention design.
• Provide interpretable attention maps and validate performance against state-of-the-art baselines on public ST datasets.

Proposed method

• Model tissue as a fully connected graph via self-attention over all spot pairs.
• Introduce negative-aware attention to learn both positive and negative interactions.
• Incorporate a static positional bias to guide local vs global interactions across attention heads.
• Propose off-grid sampling to create pseudo-spots that capture intermediate morphological context.
• Use a two-stage local (k-NN) and global attention block to manage computational cost of increased spots.
• Train with mean squared error loss and evaluate with MSE, MAE, and PCC on cross-validated datasets.

Figure 1 : Conceptual overview of our proposed framework, illustrating how it addresses key challenges of prior ST methods. Existing approaches are limited by sparse graphs that miss potential interactions, standard positive-only relationships in attention that omit inhibitory interactions, and spar

Experimental results

Research questions

• RQ1 Can a fully connected attention mechanism capture all relevant inter-spot interactions in spatial transcriptomics better than sparse-graph GNNs?
• RQ2 Do negative-aware attention and off-grid pseudo-spots improve gene expression prediction and interpretability over baselines?
• RQ3 How does hierarchical attention with local and global stages balance performance and computation?
• RQ4 Does FEAST produce more biologically plausible attention maps that distinguish excitatory and inhibitory interactions?

Key findings

• FEAST achieves state-of-the-art performance, best in 7 of 9 evaluation metrics across three ST datasets.
• On Her2ST, FEAST attains MSE 0.5761, PCC 0.5524, surpassing prior best results.
• On SCC, FEAST yields PCC 0.5811, the top among evaluated methods.
• Qualitatively, FEAST produces attention maps that distinguish positive and negative interactions.
• Off-grid sampling (pseudo-spots) significantly improves target-spot prediction when surrounding context is sparse.
• Ablation confirms both negative-aware attention and off-grid sampling contribute to performance gains.

Figure 2 : The overall architecture of FEAST. The framework first extracts features and spot distance ( $\mathbf{B}_{h}$ ) from the input WSI. The features are then processed through $L$ stacked hierarchical attention layers. Each layer consists of two stages: (1) a FEAST Block applied to local $k$