Setup and Train the DOLPHIN Model

This tutorial provides a step-by-step guide on configuring the model architecture, setting hyperparameters, and visualizing cell embedding clusters using DOLPHIN.

from DOLPHIN.model import run_DOLPHIN
import numpy as np

Load Processed Dataset

Specify the graph data input and the highly variable gene (HVG)-filtered feature matrix obtained from the preprocessing step.

#load datasets
graph_data = "model_<sample_name>.pt"
feature_data = "FeatureCompHvg_<sample_name>.h5ad"
## save the output adata, default is set to the current folder
output_path = './'

Set Hyperparameters and Train the Model

The function run_DOLPHIN is used to configure hyperparameters and train the model. Below is a detailed explanation of its parameters:

---

Function Definition

```python run_DOLPHIN(data_type, graph_in, fea_in, current_out_path=’./’, params=None, device=’cuda:0’, seed_num=0)

Parameters

1. data_type Specifies the type of input single-cell RNA-seq data:

• "full-length": For full-length RNA-seq data.

• "10x": For 10x Genomics RNA-seq data.

2. graph_in The input graph dataset.

3. fea_in The input feature matrix, provided as an AnnData object (adata).

4. current_out_path Specifies the output directory where the resulting cell embeddings (X_z) will be saved. The output file will be named: DOLPHIN_Z.h5ad

5. params Model hyperparameters.

If data_type is set, you can use the default hyperparameters or provide your own in a dictionary format.

Below is a list of customizable hyperparameters:

Parameter	Description
"gat_channel"	Number of features per node after the GAT layer.
"nhead"	Number of attention heads in the graph attention layer.
"gat_dropout"	Dropout rate for the GAT layer.
"list_gra_enc_hid"	Neuron sizes for each fully connected layer of the encoder.
"gra_p_dropout"	Dropout rate for the encoder.
"z_dim"	Dimensionality of the latent Z space.
"list_fea_dec_hid"	Neuron sizes for each fully connected layer of the feature decoder.
"list_adj_dec_hid"	Neuron sizes for each fully connected layer of the adjacency decoder.
"lr"	Learning rate for optimization.
"batch"	Mini-batch size.
"epochs"	Number of training epochs.
"kl_beta"	KL divergence weight.
"fea_lambda"	Feature matrix reconstruction loss weight.
"adj_lambda"	Adjacency matrix reconstruction loss weight.

6. device Specifies the device for training. Default: "cuda:0" for GPU training (highly recommended).

7. seed_numSets the random seed for reproducibility.

run_DOLPHIN("full-length", graph_data, feature_data, output_path)

Cell Embedding Cluster Using X_z

The cell embedding matrix X_z represents the low-dimensional latent space learned by the DOLPHIN model.

This matrix can be used to visualize cell clusters and analyze their relationships in the latent space.

import scanpy as sc
from sklearn.metrics import adjusted_rand_score

adata = sc.read_h5ad("./DOLPHIN_Z.h5ad")

sc.pp.neighbors(adata, use_rep="X_z")
sc.tl.umap(adata)
sc.tl.leiden(adata)
print(len(set(adata.obs["leiden"])))
adjusted_rand_score(adata.obs["celltype"], adata.obs["leiden"])
sc.pl.umap(adata, color=['leiden', "celltype"], wspace=0.5)