semanticst 0.1.0

SemanticST: Spatially informed semantic graph learning for effective clustering and integration of spatial transcriptomics

SemanticST

---

🔷 SemanticST Overview

SemanticST is a graph neural network-based unsupervised deep learning approach for spatial transcriptomics data analysis. SemanticST employs a sophisticated learning strategy to integrate gene expression and spatial information, enabling the model to learn a latent representation of spatial transcriptomics (ST) data. SemanticST uses a learnable weighted graph representation, termed Semantic Graphs, to better capture the complexity and diversity of biological processes. For each semantic graph, a unique embedding is learned using an autoencoder with Graph Convolutional Network (GCN) layers, representing distinct semantic features in latent space. To combine these representations, SemanticST introduces a learnable weight, referred to as the semantic score, for each semantic graph. The final graph representation is then dynamically fused by weighting and combining the individual embeddings based on their semantic scores, resulting in a more accurate and adaptive graph representation. SemanticST incorporates the MinCut loss function together with the Deep MinCut (DMC) learning algorithm. This approach not only captures the global structure of the graph but also reduces redundant training time. More importantly, it ensures that the learned embeddings are both meaningful and interpretable, providing a more robust and insightful representation of the graph. Notably, we incorporated a mini-batch training option in SemanticST by training the model on spatial graphs in smaller batches. This feature allows the learned semantic graph to maintain both local and global perspectives across batches, making SemanticST memory-efficient and scalable, enabling its application to any spatial transcriptomics dataset, regardless of the number of samples.

🔷 Requirements

To run SemanticST, install the following dependencies:

*python==3.9.20
*numpy==1.23.4
*anndata==0.10.9
*h5py==3.12.1
*leidenalg==0.10.2
*louvain==0.8.2
*matplotlib==3.9.2
*numba==0.60.0
*scanpy==1.10.3
*scikit-image==0.24.0
*scikit-learn==1.5.2
*scikit-misc==0.3.1
*torch==2.5.1
*torch-geometric==2.6.1
*torch_scatter==2.1.2+pt25cu124
*torch_sparse==0.6.18+pt25cu124
*torchaudio==2.5.1
*torchvision==0.20.1
*torchviz==0.0.2

Installation

You can install SemanticST via PyPI:

pip install semanticst

---

Or install from GitHub:

pip install git+https://github.com/roxana9/SemanticST.git

---

To install all dependencies, run:

pip install -r requirements.txt

🔷 Step-by-Step Tutorial

For a full tutorial on how to use SemanticST, visit our documentation:

🔗 SemanticST Tutorial Webpage

(The webpage is still under development, but a temporary version is available.)