ReconST API Reference

ReconST provides a lightweight, autoencoder-based framework for gene panel selection in spatial transcriptomics.
This page summarizes the key classes and functions, organized by workflow: data preparation, model definition, training, evaluation, and gene selection.

---

1. Model

FeatureScreeningAutoencoder

Autoencoder with a learnable feature-importance layer for gene selection.

Architecture - Learnable gene-importance layer
- Encoder: input_size → 512 → 256 → embedding_size
- Decoder: embedding_size → 256 → 512 → input_size

Parameters - input_size (int): Number of genes
- embedding_size (int): Latent dimension
- dp (float): Dropout probability
- lk (float): LeakyReLU negative slope

Attributes - feature_importance: Learnable gene weights
- encoder, decoder: Sequential modules

Example

import torch
from reconst import FeatureScreeningAutoencoder

model = FeatureScreeningAutoencoder(input_size=2000, embedding_size=64)
x = torch.randn(32, 2000)
screened, latent, recon = model(x)

---

2. Data Utilities

create_data_loader(adata, batch_size=256, train_split=0.8, shuffle=True)

Create training and test DataLoaders from an AnnData object.

Example

import scanpy as sc
from reconst import create_data_loader

adata = sc.read_h5ad("data.h5ad")
train_loader, test_loader = create_data_loader(adata, batch_size=128)

---

prepare_common_genes(adata1, adata2)

Find and align common genes between two datasets.

---

3. Training

train_model(model, train_loader, test_loader, num_epochs=20, lr=1e-3, weight_decay=1e-5, l_lambda=1e-4, device='cpu')

Train the autoencoder with MSE + L1 sparsity penalty.

Example

from reconst import train_model, FeatureScreeningAutoencoder

model = FeatureScreeningAutoencoder(2000, 64).to('cuda')
train_losses, test_losses = train_model(
    model, train_loader, test_loader,
    num_epochs=50, lr=1e-3, l_lambda=1e-4, device='cuda'
)

---

4. Evaluation

evaluate_model(model, test_loader, gene_mask=None, device='cpu')

Compute reconstruction MSE with all genes or a selected subset.

Example

loss = evaluate_model(model, test_loader)

---

5. Gene Selection

select_genes(model, threshold=0.001)

Extract selected genes based on learned feature importance.

Example

gene_mask, importances = select_genes(model, threshold=0.01)
selected_names = adata.var_names[gene_mask]

---

Summary

ReconST exposes a streamlined, end-to-end workflow:

1. Load and align AnnData
2. Build DataLoaders
3. Initialize the autoencoder
4. Train with sparsity
5. Evaluate
6. Select informative genes