tGPT¶
⚠️ Status: partial | Version: v1.0
Overview¶
Autoregressive next-token prediction of gene expression values (GPT-style, not masked)
When to choose tGPT
User wants autoregressive/generative modeling, next-token prediction of gene expression, or GPT-style generation
Specifications¶
| Property | Value |
|---|---|
| Model | tGPT |
| Version | v1.0 |
| Tasks | embed, integrate |
| Modalities | RNA |
| Species | human |
| Gene IDs | symbol |
| Embedding Dim | 512 |
| GPU Required | Yes |
| Min VRAM | 16 GB |
| Recommended VRAM | 32 GB |
| CPU Fallback | No |
| Adapter Status | ⚠️ partial |
Quick Start¶
import omicverse as ov
# 1. Check model spec
info = ov.fm.describe_model("tgpt")
# 2. Profile your data
profile = ov.fm.profile_data("your_data.h5ad")
# 3. Validate compatibility
check = ov.fm.preprocess_validate("your_data.h5ad", "tgpt", "embed")
# 4. Run inference
result = ov.fm.run(
task="embed",
model_name="tgpt",
adata_path="your_data.h5ad",
output_path="output_tgpt.h5ad",
device="auto",
)
# 5. Interpret results
metrics = ov.fm.interpret_results("output_tgpt.h5ad", task="embed")
Input Requirements¶
| Requirement | Detail |
|---|---|
| Gene ID scheme | symbol |
| Preprocessing | Standard preprocessing. Gene expression values are tokenized autoregressively. |
| Data format | AnnData (.h5ad) |
| Batch key | .obs column for batch integration (optional) |
Output Keys¶
After running ov.fm.run(), results are stored in the AnnData object:
| Key | Location | Description |
|---|---|---|
X_tgpt |
adata.obsm |
Cell embeddings (512-dim) |
import scanpy as sc
adata = sc.read_h5ad("output_tgpt.h5ad")
embeddings = adata.obsm["X_tgpt"] # shape: (n_cells, 512)
# Downstream analysis
sc.pp.neighbors(adata, use_rep="X_tgpt")
sc.tl.umap(adata)
sc.tl.leiden(adata, resolution=0.5)
sc.pl.umap(adata, color=["leiden"])
Resources¶
- Repository / Checkpoint: https://github.com/deeplearningplus/tGPT
- License: Check upstream LICENSE
Hands-On Tutorial¶
For a step-by-step walkthrough with code, see the tGPT Tutorial Notebook.