### Install TOSICA Package from GitHub

Source: https://github.com/jackiehanlab/tosica/blob/main/README.md

Installs the TOSICA Python package directly from its GitHub repository.

```bash
pip install git+https://github.com/JackieHanLab/TOSICA.git
```

--------------------------------

### Install TOSICA Package Locally

Source: https://github.com/jackiehanlab/tosica/blob/main/README.md

Installs the TOSICA Python package from the local repository root.

```bash
pip install .
```

--------------------------------

### Check PyTorch and CUDA Version

Source: https://github.com/jackiehanlab/tosica/blob/main/test/tutorial.ipynb

Verifies the installed PyTorch version and CUDA capabilities, essential for GPU acceleration.

```python
import torch
print(torch.__version__)
print(torch.cuda.get_device_capability(device=None),  torch.cuda.get_device_name(device=None))
```

--------------------------------

### Initialize Environment

Source: https://github.com/jackiehanlab/tosica/blob/main/reproducibility/Fig4.ipynb

Imports necessary libraries and configures global settings for Scanpy plotting.

```python
import scanpy as sc
import os
import math
import itertools
import warnings
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import anndata
sc.settings.set_figure_params(dpi_save=600,  frameon=False, transparent=True, fontsize=10)
warnings.filterwarnings('ignore')
```

--------------------------------

### Fine-tune with pre-trained weights

Source: https://context7.com/jackiehanlab/tosica/llms.txt

Initializes the model with existing weights to perform transfer learning on new datasets.

```python
import TOSICA
import scanpy as sc

ref_adata = sc.read('new_reference_data.h5ad')

# Fine-tune from pre-trained weights
TOSICA.train(
    adata=ref_adata,
    gmt_path='human_gobp',
    label_name='Celltype',
    project='finetuned_model',
    pre_weights='./pretrained_model/model-5.pth',  # Path to pre-trained weights
    epochs=5
)
```

--------------------------------

### TOSICA Training with Pre-weights Output

Source: https://github.com/jackiehanlab/tosica/blob/main/test/tutorial.ipynb

Shows the training progress and validation metrics for a single epoch when using pre-trained weights.

```text
[train epoch 0] loss: 0.153, acc: 0.974: 100%|██████████| 5096/5096 [01:32<00:00, 55.28it/s]
[valid epoch 0] loss: 0.040, acc: 0.992: 100%|██████████| 5096/5096 [00:31<00:00, 162.61it/s]
```

--------------------------------

### Create Conda Environment for TOSICA

Source: https://github.com/jackiehanlab/tosica/blob/main/README.md

Sets up a Conda environment with Python 3.8, scanpy, and specific PyTorch versions required for TOSICA.

```bash
conda create -n TOSICA python=3.8 scanpy
conda activate TOSICA
conda install pytorch=1.7.1 torchvision=0.8.2 torchaudio=0.7.2 cudatoolkit=10.1 -c pytorch
```

--------------------------------

### TOSICA Training Output Files

Source: https://github.com/jackiehanlab/tosica/blob/main/test/tutorial.ipynb

Lists the files generated after training, including model weights, pathway information, and label dictionaries.

```text
Training finished!
label_dictionary.csv
mask.npy
model-0.pth
model-1.pth
model-2.pth
pathway.csv
```

--------------------------------

### Import Libraries

Source: https://github.com/jackiehanlab/tosica/blob/main/test/tutorial.ipynb

Imports necessary libraries for TOSICA and general data analysis. Suppresses warnings for cleaner output.

```python
import TOSICA
import scanpy as sc
import numpy as np
import warnings 
warnings.filterwarnings ("ignore")
```

--------------------------------

### Copy Pre-trained Weights

Source: https://github.com/jackiehanlab/tosica/blob/main/test/tutorial.ipynb

Copies a pre-trained model weight file to be used for fine-tuning or transfer learning.

```bash
!cp ./hGOBP_demo/model-0.pth ./pre_weights.pth
```

--------------------------------

### Load Data

Source: https://github.com/jackiehanlab/tosica/blob/main/reproducibility/Fig4.ipynb

Reads the raw h5ad file into an AnnData object.

```python
adata = sc.read('./Fig4_raw.h5ad')
adata
```

--------------------------------

### Preprocess Data for Visualization

Source: https://github.com/jackiehanlab/tosica/blob/main/test/tutorial.ipynb

Applies standard preprocessing steps using scanpy, including normalization, log transformation, scaling, PCA, neighbor graph construction, and UMAP dimensionality reduction.

```python
new_adata.raw = new_adata
sc.pp.normalize_total(new_adata, target_sum=1e4)
sc.pp.log1p(new_adata)
sc.pp.scale(new_adata, max_value=10)
sc.tl.pca(new_adata, svd_solver='arpack')
sc.pp.neighbors(new_adata, n_neighbors=10, n_pcs=40)
sc.tl.umap(new_adata)
```

--------------------------------

### Complete TOSICA Training and Prediction Pipeline

Source: https://context7.com/jackiehanlab/tosica/llms.txt

A full workflow demonstrating data loading, model training, prediction, and visualization. Ensure gene alignment between reference and query datasets is critical.

```python
import TOSICA
import scanpy as sc
import warnings
warnings.filterwarnings("ignore")

# Load and inspect reference data
ref_adata = sc.read('demo_train.h5ad')
print(f"Reference: {ref_adata.n_obs} cells × {ref_adata.n_vars} genes")
print(ref_adata.obs['Celltype'].value_counts())

# Load query data and align genes to reference
query_adata = sc.read('demo_test.h5ad')
query_adata = query_adata[:, ref_adata.var_names]  # Critical: match gene order
print(f"Query: {query_adata.n_obs} cells × {query_adata.n_vars} genes")

# Train model with human GO biological processes
TOSICA.train(
    ref_adata,
    gmt_path='human_gobp',
    label_name='Celltype',
    project='pancreas_gobp',
    epochs=10
)

# Select best model (typically last epoch or based on validation accuracy)
model_path = './pancreas_gobp/model-9.pth'

# Predict cell types on query data
new_adata = TOSICA.pre(
    query_adata,
    model_weight_path=model_path,
    project='pancreas_gobp'
)

# Evaluate prediction accuracy
if 'Celltype' in new_adata.obs.columns:
    accuracy = (new_adata.obs['Prediction'] == new_adata.obs['Celltype']).mean()
    print(f"Prediction accuracy: {accuracy:.3f}")

# Generate UMAP visualization
new_adata.raw = new_adata
sc.pp.normalize_total(new_adata, target_sum=1e4)
sc.pp.log1p(new_adata)
sc.pp.scale(new_adata, max_value=10)
sc.tl.pca(new_adata)
sc.pp.neighbors(new_adata, n_neighbors=10, n_pcs=40)
sc.tl.umap(new_adata)
sc.pl.umap(new_adata, color=['Celltype', 'Prediction', 'Probability'])

# Save results
new_adata.write('annotated_results.h5ad')
```

--------------------------------

### Train TOSICA with Pre-trained Weights

Source: https://github.com/jackiehanlab/tosica/blob/main/test/tutorial.ipynb

Trains a TOSICA model using reference data and pre-trained weights, specifying the GMT pathway, label name, and number of epochs. Indicates successful loading of masks and model.

```python
TOSICA.train(ref_adata, gmt_path='human_gobp', label_name='Celltype',pre_weights='pre_weights.pth',epochs=1)
```

--------------------------------

### Train TOSICA Model

Source: https://github.com/jackiehanlab/tosica/blob/main/test/tutorial.ipynb

Trains a TOSICA model using the reference data, specifying the GMT pathway file, label name, number of epochs, and project name. Lists the contents of the output directory.

```python
TOSICA.train(ref_adata, gmt_path='human_gobp', label_name='Celltype',epochs=3,project='hGOBP_demo')
!ls ./hGOBP_demo
```

--------------------------------

### TOSICA Model Training Output

Source: https://github.com/jackiehanlab/tosica/blob/main/test/tutorial.ipynb

Displays the training progress and validation metrics for each epoch, indicating loss and accuracy.

```text
[train epoch 0] loss: 1.876, acc: 0.296: 100%|██████████| 5096/5096 [01:39<00:00, 51.10it/s]
[valid epoch 0] loss: 0.305, acc: 0.967: 100%|██████████| 5096/5096 [00:34<00:00, 149.26it/s]
[train epoch 1] loss: 0.168, acc: 0.975: 100%|██████████| 5096/5096 [01:40<00:00, 50.88it/s]
[valid epoch 1] loss: 0.042, acc: 0.993: 100%|██████████| 5096/5096 [00:33<00:00, 152.25it/s]
[train epoch 2] loss: 0.059, acc: 0.992: 100%|██████████| 5096/5096 [01:39<00:00, 51.08it/s]
[valid epoch 2] loss: 0.024, acc: 0.996: 100%|██████████| 5096/5096 [00:33<00:00, 150.93it/s]
```

--------------------------------

### Visualize Pathway Signaling with Scatter Plots

Source: https://github.com/jackiehanlab/tosica/blob/main/reproducibility/Fig4.ipynb

Generates a scatter plot to compare two specific Reactome pathways across cancer types.

```python
sc.pl.scatter(sub, x='REACTOME_CYTOKINE_SIGNALING_IN_IMMUNE_SYSTEM', y='REACTOME_SIGNALING_BY_INSULIN_RECEPTOR', 
              color='cancer', size=10, title='Macro_LYVE1',use_raw=True)
```

--------------------------------

### Perform PCA, UMAP, and Louvain Clustering

Source: https://github.com/jackiehanlab/tosica/blob/main/reproducibility/Fig4.ipynb

This snippet preprocesses a subset of the AnnData object by scaling, performing PCA, computing nearest neighbors, and applying UMAP and Louvain clustering. It then visualizes the Louvain clusters on the UMAP plot. Ensure 'adata' is an AnnData object and 'Prebigtype' is a valid observation key.

```python
sub = adata[adata.obs.Prebigtype=='Mono',:]
#sc.pp.scale(sub, max_value=10)
sc.tl.pca(sub, svd_solver='arpack')
sc.pp.neighbors(sub, n_neighbors=10, n_pcs=50)
sc.tl.umap(sub)
sc.tl.louvain(sub,resolution=0.3, key_added='louvain_03')
sc.pl.umap(sub, color='louvain_03', legend_loc='on data', frameon=False, legend_fontsize=10,title='louvain_03')
```

--------------------------------

### Train TOSICA Model

Source: https://github.com/jackiehanlab/tosica/blob/main/README.md

Trains the TOSICA model using a reference AnnData object and a GMT file. The model and associated files are saved in a specified project directory.

```python
TOSICA.train(ref_adata, gmt_path,project=<my_project>,label_name=<label_key>)
```

--------------------------------

### Inspect Metadata

Source: https://github.com/jackiehanlab/tosica/blob/main/reproducibility/Fig4.ipynb

Checks the distribution of categorical metadata fields.

```python
adata.obs.cancer.value_counts()
```

```python
adata.obs.bigtype.value_counts()
```

```python
adata.obs.Celltype.value_counts()
```

--------------------------------

### TOSICA Training Finished Message

Source: https://github.com/jackiehanlab/tosica/blob/main/test/tutorial.ipynb

Confirms that the training process, potentially with pre-trained weights, has been completed.

```text
Training finished!
```

--------------------------------

### Load Query AnnData

Source: https://github.com/jackiehanlab/tosica/blob/main/test/tutorial.ipynb

Loads a query dataset using scanpy and aligns its genes with the reference AnnData. Displays the AnnData object and cell type value counts.

```python
query_adata = sc.read('demo_test.h5ad')
query_adata = query_adata[:,ref_adata.var_names]
print(query_adata)
print(query_adata.obs.Celltype.value_counts())
```

--------------------------------

### Configure Stage Categories and Colors

Source: https://github.com/jackiehanlab/tosica/blob/main/reproducibility/Fig4.ipynb

Converts the Stage observation to a categorical type and assigns custom hex color codes for visualization.

```python
sub.obs.Stage = sub.obs.Stage.astype('category')
sub.uns['Stage_colors'] = ["#BEBEBE", "#CE8E8E", "#DE5F5F" ,"#EE2F2F", "#FF0000"]
```

--------------------------------

### PAGA Analysis

Source: https://github.com/jackiehanlab/tosica/blob/main/reproducibility/Fig4.ipynb

Performs PAGA analysis to visualize cluster connectivity.

```python
sc.tl.paga(sub, groups='Precelltype')
sc.pl.paga(sub, threshold=0.3, show=False,color='dpt_pseudotime')
```

--------------------------------

### Analyze Gene-to-Token Weights

Source: https://context7.com/jackiehanlab/tosica/llms.txt

Load and analyze the gene-to-token weight matrix to identify key marker genes for specific pathways or pathways associated with a gene. Ensure the `gene2token_weights.csv` file is generated during prediction.

```python
import pandas as pd
import numpy as np

# Load gene-to-token weights (generated during prediction)
gene2token = pd.read_csv('./my_annotation_model/gene2token_weights.csv', index_col=0)

# Find top genes for a specific pathway
pathway_name = 'GOBP_IMMUNE_RESPONSE'
if pathway_name in gene2token.index:
    pathway_weights = gene2token.loc[pathway_name].sort_values(ascending=False)
    top_genes = pathway_weights.head(20)
    print(f"Top genes for {pathway_name}:")
    print(top_genes)

# Find pathways where a specific gene has highest weight
gene_name = 'CD3D'
if gene_name in gene2token.columns:
    gene_weights = gene2token[gene_name].sort_values(ascending=False)
    top_pathways = gene_weights.head(10)
    print(f"\nTop pathways involving {gene_name}:")
    print(top_pathways)
```

--------------------------------

### Load Reference AnnData

Source: https://github.com/jackiehanlab/tosica/blob/main/test/tutorial.ipynb

Loads a reference dataset using scanpy and ensures it contains the same genes as the reference. Prints the AnnData object and cell type value counts.

```python
ref_adata = sc.read('demo_train.h5ad')
ref_adata = ref_adata[:,ref_adata.var_names]
print(ref_adata)
print(ref_adata.obs.Celltype.value_counts())
```

--------------------------------

### Train with custom gene sets

Source: https://context7.com/jackiehanlab/tosica/llms.txt

Uses a custom GMT file for pathway masks to allow domain-specific annotations during model training.

```python
import TOSICA
import scanpy as sc

ref_adata = sc.read('reference_data.h5ad')

# Train with custom GMT file containing domain-specific pathways
TOSICA.train(
    adata=ref_adata,
    gmt_path='/path/to/custom_pathways.gmt',
    label_name='Celltype',
    project='custom_pathway_model',
    max_g=300,       # Maximum genes per pathway
    max_gs=300,      # Maximum number of pathways to include
    n_unannotated=1, # Number of fully connected tokens for unmapped genes
    epochs=10
)
```

--------------------------------

### Dimensionality Reduction and Visualization

Source: https://github.com/jackiehanlab/tosica/blob/main/reproducibility/Fig4.ipynb

Scales data, computes PCA, builds a neighbor graph, and generates UMAP plots.

```python
sc.pp.scale(adata)
sc.tl.pca(adata, svd_solver='arpack')
sc.pp.neighbors(adata, n_neighbors=10, n_pcs=40)
sc.tl.umap(adata)
sc.pl.umap(adata, color='Celltype',edgecolor="none",show=False)
sc.pl.umap(adata, color='Precelltype',edgecolor="none",show=False)
```

--------------------------------

### Save and Display Predictions

Source: https://github.com/jackiehanlab/tosica/blob/main/test/tutorial.ipynb

Saves the AnnData object with predictions to a file and displays the resulting AnnData object, showing the new 'Prediction' and 'Probability' observation columns.

```python
new_adata.write('demo_attn.h5ad')
new_adata
```

--------------------------------

### Analyze Attention Matrices with Scanpy

Source: https://context7.com/jackiehanlab/tosica/llms.txt

Visualize pathway contributions and create UMAP embeddings based on attention features. Requires prior prediction to generate the `new_adata` object.

```python
import TOSICA
import scanpy as sc
import numpy as np

# After prediction
new_adata = TOSICA.pre(query_adata, model_weight_path='./model/model-5.pth', project='model')

# Preserve raw attention scores
new_adata.raw = new_adata

# Preprocess attention matrix for visualization
sc.pp.normalize_total(new_adata, target_sum=1e4)
sc.pp.log1p(new_adata)
sc.pp.scale(new_adata, max_value=10)

# Compute UMAP embedding based on pathway attention
sc.tl.pca(new_adata, svd_solver='arpack')
sc.pp.neighbors(new_adata, n_neighbors=10, n_pcs=40)
sc.tl.umap(new_adata)

# Visualize predictions vs ground truth
sc.pl.umap(new_adata, color=['Celltype', 'Prediction'], legend_loc='on data')

# Visualize specific pathway attention scores
sc.pl.umap(new_adata, color=['GOBP_IMMUNE_RESPONSE', 'GOBP_CELL_CYCLE'])

# Heatmap of top pathways per cell type
sc.tl.rank_genes_groups(new_adata, 'Prediction', method='wilcoxon')
sc.pl.rank_genes_groups_heatmap(new_adata, n_genes=10)
```

--------------------------------

### Train a TOSICA model

Source: https://context7.com/jackiehanlab/tosica/llms.txt

Trains a transformer model on a reference dataset using predefined gene set masks. Requires an AnnData object with cell-type labels.

```python
import TOSICA
import scanpy as sc

# Load reference dataset with cell type labels
ref_adata = sc.read('reference_data.h5ad')

# Train with pre-configured human Gene Ontology biological process mask
TOSICA.train(
    adata=ref_adata,
    gmt_path='human_gobp',  # Options: human_gobp, human_immune, human_reactome, human_tf, mouse_gobp, mouse_reactome, mouse_tf
    label_name='Celltype',  # Column name in adata.obs containing cell type labels
    project='my_annotation_model',  # Output folder name
    epochs=10,
    batch_size=8,
    embed_dim=48,
    depth=2,
    num_heads=4,
    lr=0.001
)

# Output files created in ./my_annotation_model/:
# - mask.npy: Pathway-gene mask matrix
# - pathway.csv: List of pathways/gene sets used
# - label_dictionary.csv: Cell type label encoding
# - model-{epoch}.pth: Model weights for each epoch
```

--------------------------------

### Generate UMAP Visualization

Source: https://github.com/jackiehanlab/tosica/blob/main/test/tutorial.ipynb

Generates a UMAP plot colored by 'Celltype' and 'Prediction' from the AnnData object, with the legend placed on the plot.

```python
sc.pl.umap(new_adata, color=['Celltype', 'Prediction'],legend_loc='on data')
```

--------------------------------

### Normalize and Log-transform

Source: https://github.com/jackiehanlab/tosica/blob/main/reproducibility/Fig4.ipynb

Performs total count normalization and log transformation on the AnnData object.

```python
sc.pp.normalize_total(adata, target_sum=1e4)
sc.pp.log1p(adata)
adata.raw = adata
```

--------------------------------

### Display Top 20 Correlations

Source: https://github.com/jackiehanlab/tosica/blob/main/reproducibility/Fig4.ipynb

Displays the top 20 gene pathways with the highest Spearman correlation coefficients with age, along with their corresponding p-values.

```python
rcc[0:20]
```

--------------------------------

### Predict Cell Types with TOSICA

Source: https://github.com/jackiehanlab/tosica/blob/main/test/tutorial.ipynb

Uses a trained TOSICA model to predict cell types for the query data. Specifies the model weight path and project name. Shows the device and number of cells processed.

```python
model_weight_path = './hGOBP_demo/model-0.pth'
new_adata = TOSICA.pre(query_adata, model_weight_path = model_weight_path,project='hGOBP_demo')
```

--------------------------------

### TOSICA.train

Source: https://github.com/jackiehanlab/tosica/blob/main/README.md

Trains the TOSICA model using a reference AnnData object and a gene set mask.

```APIDOC
## TOSICA.train

### Description
Trains the Transformer model for cell-type annotation based on a reference dataset and a specified gene set mask.

### Parameters
#### Request Body
- **ref_adata** (AnnData) - Required - An AnnData object of the reference dataset.
- **gmt_path** (string) - Required - Path to .gmt files or a pre-prepared mask name.
- **project** (string) - Optional - Name of the project folder to save model outputs.
- **label_name** (string) - Required - The name of the label column in ref_adata.obs.

### Response
- **mask.npy** (file) - Mask matrix.
- **pathway.csv** (file) - Gene set list.
- **label_dictionary.csv** (file) - Label list.
- **model-n.pth** (file) - Model weights.
```

--------------------------------

### Calculate Spearman Correlation for Pathways

Source: https://github.com/jackiehanlab/tosica/blob/main/reproducibility/Fig4.ipynb

Computes Spearman correlation coefficients between cell stages and pathway expression indices for a specific cell type and cancer subset.

```python
from scipy.stats import spearmanr, pearsonr
sub = adata[adata.obs.Celltype=='Macro_LYVE1',:]
sub = sub[sub.obs.cancer=='ESCA',:]
rcc = pd.DataFrame(sub.var['pathway_index'])
rcc['rcc'] = 0
rcc['p'] = 0
for i in range(len(rcc)):
    rr = spearmanr(sub.obs.Stage,sub.X[:,i])
    rcc.loc[rcc.index[i],'rcc'] = rr[0]
    rcc.loc[rcc.index[i],'p'] = rr[1]
rcc.sort_values(by='rcc', ascending=False)
```

--------------------------------

### Trajectory Inference

Source: https://github.com/jackiehanlab/tosica/blob/main/reproducibility/Fig4.ipynb

Subsets the data and performs trajectory analysis using Diffusion Maps and DPT.

```python
sub = adata[adata.obs.Prebigtype=='cDC',:]
sub = sub[sub.obs.bigtype=='cDC',:]
sub = sub[sub.obs.tissue=='T',:]
sc.pp.scale(sub, max_value=10)
sc.tl.pca(sub, svd_solver='arpack')
sc.pp.neighbors(sub, n_neighbors=10, n_pcs=40)
sc.tl.umap(sub)
sc.pl.umap(sub, color='Precelltype')
sc.tl.diffmap(sub)
sub.uns['iroot'] = np.flatnonzero(sub.obs['Precelltype']  == 'cDC3_LAMP3')[5]
sc.tl.dpt(sub)
sc.pl.diffmap(sub, color='Precelltype')
```

--------------------------------

### TOSICA.pre

Source: https://github.com/jackiehanlab/tosica/blob/main/README.md

Predicts cell types for a query dataset using a pre-trained TOSICA model.

```APIDOC
## TOSICA.pre

### Description
Performs cell-type prediction on a query dataset using weights generated during the training phase.

### Parameters
#### Request Body
- **query_adata** (AnnData) - Required - An AnnData object of the query dataset.
- **model_weight_path** (string) - Optional - Path to the model weights file (.pth).
- **project** (string) - Required - Name of the project folder created during training.

### Response
- **new_adata.X** (matrix) - Attention matrix.
- **new_adata.obs['Prediction']** (list) - Predicted labels.
- **new_adata.obs['Probability']** (list) - Probability of the prediction.
- **new_adata.var['pathway_index']** (list) - Gene set of each column.
```

--------------------------------

### Predict Cell Types with TOSICA Model

Source: https://github.com/jackiehanlab/tosica/blob/main/README.md

Predicts cell types for a query AnnData object using a trained TOSICA model. Optionally, specific model weights can be provided.

```python
new_adata = TOSICA.pre(query_adata, model_weight_path = <path to optional weight>,project=<my_project>)
```

--------------------------------

### Plot Violin Plot of Gene Expression

Source: https://github.com/jackiehanlab/tosica/blob/main/reproducibility/Fig4.ipynb

Generates a violin plot to visualize the expression of specified genes across different stages. Requires the 'scanpy' library and an AnnData object.

```python
sc.pl.violin(sub, ['REACTOME_SIGNALING_BY_FGFR','REACTOME_INTERFERON_SIGNALING'], groupby='Stage',use_raw=True)
```

--------------------------------

### Ensure Consistent Gene Order

Source: https://github.com/jackiehanlab/tosica/blob/main/README.md

Ensures that the gene (variable) names in the query AnnData object are consistent with and in the same order as the reference AnnData object before prediction.

```python
query_adata = query_adata[:,ref_adata.var_names]
```

--------------------------------

### Scatter Plot of Gene Expression

Source: https://github.com/jackiehanlab/tosica/blob/main/reproducibility/Fig4.ipynb

Generates a scatter plot to visualize the relationship between two genes, colored by cell type. Requires the 'scanpy' library and a pre-processed AnnData object.

```python
sc.pl.scatter(sub,x='REACTOME_TOLL_RECEPTOR_CASCADES',y='REACTOME_NOD1_2_SIGNALING_PATHWAY',color='Precelltype',use_raw=True)
```

--------------------------------

### Save AnnData Object

Source: https://github.com/jackiehanlab/tosica/blob/main/reproducibility/Fig4.ipynb

Saves the current state of the AnnData object to a .h5ad file. This is useful for preserving processed data for later analysis or sharing.

```python
adata.write_h5ad('./Fig4.h5ad')
```

--------------------------------

### Rank Genes Groups

Source: https://github.com/jackiehanlab/tosica/blob/main/reproducibility/Fig4.ipynb

Identifies marker genes for clusters using the Wilcoxon rank-sum test.

```python
sc.tl.rank_genes_groups(sub, 'Precelltype', method='wilcoxon')
dc_pathway = pd.DataFrame(sub.uns['rank_genes_groups']['names'])
dc_pathway
```

--------------------------------

### Assign Colors and Categories for Plotting

Source: https://github.com/jackiehanlab/tosica/blob/main/test/tutorial.ipynb

Assigns specific colors and reorders categories for 'Prediction' and 'Celltype' observation columns in the AnnData object to ensure consistent visualization.

```python
col = np.array([
"#98DF8A","#E41A1C" ,"#377EB8", "#4DAF4A" ,"#984EA3" ,"#FF7F00" ,"#FFFF33" ,"#A65628" ,"#F781BF" ,"#999999","#1F77B4","#FF7F0E","#279E68","#FF9896"
]).astype('<U7')

celltype = ("alpha","beta","ductal","acinar","delta","PP","PSC","endothelial","epsilon","mast","macrophage","schwann",'t_cell')
new_adata.obs['Prediction'] = new_adata.obs['Prediction'].astype('category')
new_adata.obs['Prediction'] = new_adata.obs['Prediction'].cat.reorder_categories(list(celltype))
new_adata.uns['Prediction_colors'] = col[1:]

celltype = ("MHC class II","alpha","beta","ductal","acinar","delta","PP","PSC","endothelial","epsilon","mast")
new_adata.obs['Celltype'] = new_adata.obs['Celltype'].astype('category')
new_adata.obs['Celltype'] = new_adata.obs['Celltype'].cat.reorder_categories(list(celltype))
new_adata.uns['Celltype_colors'] = col[:11]
```

--------------------------------

### Rank Genes by Group

Source: https://github.com/jackiehanlab/tosica/blob/main/reproducibility/Fig4.ipynb

Ranks genes for a specific cell type ('Macro_LYVE1') against a grouping variable ('cancer') using the Wilcoxon rank-sum test. The top 10 genes are then displayed. Requires 'scanpy' and 'pandas'.

```python
sub = adata[adata.obs.Celltype=='Macro_LYVE1',:]
sc.tl.rank_genes_groups(sub, 'cancer', method='wilcoxon')
txt = pd.DataFrame(sub.uns['rank_genes_groups']['names']).head(10)
txt
```

--------------------------------

### Predict cell types

Source: https://context7.com/jackiehanlab/tosica/llms.txt

Applies a trained model to query data to generate annotations and attention matrices. Ensure query data gene order matches the reference dataset.

```python
import TOSICA
import scanpy as sc

# Load query dataset (must have same genes as reference in same order)
query_adata = sc.read('query_data.h5ad')
ref_adata = sc.read('reference_data.h5ad')

# Ensure gene order matches reference dataset
query_adata = query_adata[:, ref_adata.var_names]

# Run prediction
new_adata = TOSICA.pre(
    adata=query_adata,
    model_weight_path='./my_annotation_model/model-9.pth',
    project='my_annotation_model',
    cutoff=0.1,      # Probability threshold; below this marks as 'Unknown'
    batch_size=50,
    n_step=10000     # Cells loaded into memory at once
)

# Access results
print(new_adata.obs['Prediction'])    # Predicted cell type labels
print(new_adata.obs['Probability'])   # Prediction confidence scores
print(new_adata.X)                     # Attention matrix (cells × pathways)
print(new_adata.var['pathway_index']) # Pathway names for each column

# Save annotated results
new_adata.write('annotated_query.h5ad')
```

--------------------------------

### Plot Gene Expression Scatter Plot

Source: https://github.com/jackiehanlab/tosica/blob/main/reproducibility/Fig4.ipynb

Use this function to create a scatter plot comparing the expression levels of two genes, colored by cell age. Ensure 'sub' is a pre-filtered AnnData object and relevant genes are present in its observation DataFrame.

```python
sc.pl.scatter(sub, x='REACTOME_INTERFERON_SIGNALING', y='REACTOME_INNATE_IMMUNE_SYSTEM',
              color='Age', size=10, title='Mono_CD14',use_raw=False)
```

--------------------------------

### Calculate Spearman Correlation with Age

Source: https://github.com/jackiehanlab/tosica/blob/main/reproducibility/Fig4.ipynb

Calculates the Spearman correlation coefficient between 'Age' and the expression of each gene pathway for cells of type 'Mono_CD14'. Results are stored in a DataFrame and sorted by correlation.

```python
from scipy.stats import spearmanr, pearsonr
sub = adata[adata.obs.Precelltype=='Mono_CD14',:]
rcc = pd.DataFrame(sub.var['pathway_index'])
rcc['rcc'] = 0
rcc['p'] = 0
for i in range(len(rcc)):
    rr = spearmanr(sub.obs.Age,sub.X[:,i])
    rcc.loc[rcc.index[i],'rcc'] = rr[0]
    rcc.loc[rcc.index[i],'p'] = rr[1]
rcc = rcc.sort_values(by='rcc', ascending=False)
rcc
```

=== COMPLETE CONTENT === This response contains all available snippets from this library. No additional content exists. Do not make further requests.