File size: 1,554 Bytes
656047d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
---
library_name: perturblab
tags:
- biology
- genomics
- scfoundation
- foundation-model
license: apache-2.0
base_model: biomap-research/scFoundation
---

# scfoundation-rde

## Model Description
This is the **RDE (Reaction-Diffusion Embedding)** model from scFoundation. It provides specialized embeddings optimized for specific downstream tasks.

Model weights were originally from the [biomap-research/scFoundation](https://github.com/biomap-research/scFoundation) repository and have been re-uploaded here for ease of use with the `perturblab` library.

## Model Details
- **Model Type**: RDE embedding model
- **Architecture**: xTrimoGene with MAE (Masked Autoencoder), Performer/Transformer modules
- **Parameters**: 100M parameters
- **Training Data**: 50M+ human single-cell transcriptomics data
- **Input**: Single-cell or bulk RNA-seq expression data (19,264 fixed genes)
- **Output**: RDE-optimized embeddings

## Source
- **Original Repository**: [biomap-research/scFoundation](https://github.com/biomap-research/scFoundation)
- **Paper**: [Large Scale Foundation Model on Single-cell Transcriptomics](https://www.nature.com/articles/s41592-024-02305-7) (_Nature Methods_, 2024)

## Usage

```python
from perturblab.model.scfoundation import scFoundationModel

# Load model
model = scFoundationModel.from_pretrained('scfoundation-rde', device='cuda')

# Generate embeddings
embeddings = model.predict_embedding(
    adata, 
    output_type='cell',
    pool_type='max'
)
```

## Note
Intended for internal use with the PerturbLab framework.