| """Gradio web app for the HMoE dopaminergic-subtype classifier. |
| |
| Drag-and-drop an .h5ad (or Seurat .rds) of raw sc/snRNA-seq counts (dopaminergic |
| neurons) and download a CSV assigning each cell to a subtype. Deployable as a |
| Hugging Face Space (Gradio SDK); also runs locally with `python app.py`. |
| """ |
| from __future__ import annotations |
|
|
| import tempfile |
| from pathlib import Path |
|
|
| import gradio as gr |
|
|
| from hmoe_da import load_model |
| from hmoe_da.io import read_adata |
|
|
| MODEL = load_model() |
|
|
| DESCRIPTION = """ |
| # Dopaminergic-neuron subtype classifier (HMoE) |
| |
| Assign each cell in your **raw** sc/snRNA-seq dataset to one of 18 dopaminergic |
| subtypes across three families (**Sox6**, **Calb1**, **Gad2**). |
| |
| **Input:** an `.h5ad` (raw integer counts in `.X`) or a Seurat `.rds`, with gene |
| symbols in `var_names` (human UPPER or mouse Title case both work). |
| **Output:** a CSV of per-cell subtype, family, confidence, and margin. |
| |
| _Awatramani Lab, Northwestern University (Schonfeld et al., manuscript in preparation)._ |
| """ |
|
|
|
|
| def _counts_markdown(labels) -> str: |
| vc = labels.value_counts() |
| total = int(vc.sum()) |
| lines = ["| subtype | cells | % |", "|---|---:|---:|"] |
| for name, n in vc.items(): |
| lines.append(f"| {name} | {int(n)} | {100*n/total:.1f}% |") |
| lines.append(f"| **total** | **{total}** | |") |
| return "\n".join(lines) |
|
|
|
|
| def run(data_file, exclude_gad2_family, count_source): |
| if data_file is None: |
| raise gr.Error("Please upload an .h5ad or .rds file.") |
| try: |
| ad = read_adata(data_file.name) |
| except Exception as e: |
| raise gr.Error(f"Could not read file: {e}") |
|
|
| is_h5ad = str(data_file.name).lower().endswith(".h5ad") |
| if is_h5ad and count_source == "adata.raw": |
| if ad.raw is None: |
| raise gr.Error("adata.raw is None; choose 'adata.X' or upload raw counts.") |
| ad = ad.raw.to_adata() |
| elif is_h5ad and count_source.startswith("layer:"): |
| layer = count_source.split("layer:", 1)[1] |
| if layer not in ad.layers: |
| raise gr.Error(f"Layer {layer!r} not found. Available: {list(ad.layers)}") |
| ad.X = ad.layers[layer] |
|
|
| try: |
| ad = MODEL.predict(ad, include_gad2=not exclude_gad2_family, verbose=False) |
| except ValueError as e: |
| raise gr.Error(str(e)) |
|
|
| cols = [c for c in ad.obs.columns if c.startswith("hmoe")] |
| df = ad.obs[cols].copy() |
| df.index.name = "cell" |
| out = Path(tempfile.mkdtemp()) / "hmoe_predictions.csv" |
| df.to_csv(out) |
|
|
| cov = ad.uns["hmoe_gene_coverage"] |
| header = ( |
| f"**Predicted {ad.n_obs} cells.** " |
| f"Gene coverage {cov['n_matched']}/{cov['n_features']} " |
| f"({cov['pct_matched']:.1f}%). " |
| f"Gad2 family {'excluded' if exclude_gad2_family else 'included'}.\n\n" |
| ) |
| return header + _counts_markdown(ad.obs["hmoe_subtype"]), str(out) |
|
|
|
|
| def build(): |
| with gr.Blocks(title="HMoE DA subtype classifier") as demo: |
| gr.Markdown(DESCRIPTION) |
| with gr.Row(): |
| with gr.Column(): |
| f = gr.File(label="Upload .h5ad or .rds (raw counts)", |
| file_types=[".h5ad", ".rds"]) |
| gr.Markdown( |
| "<sub>Seurat `.rds` is read where R + SeuratObject are " |
| "installed (local runs). On the hosted app, please upload " |
| "`.h5ad` — convert in R with `sceasy` or `SeuratDisk`.</sub>" |
| ) |
| src = gr.Dropdown( |
| ["adata.X", "adata.raw", "layer:counts"], |
| value="adata.X", |
| label="Where are the raw counts? (.h5ad only)", |
| ) |
| excl = gr.Checkbox( |
| label="Exclude Gad2 family (reassign Gad2 calls to best non-Gad2 subtype)", |
| value=False, |
| ) |
| btn = gr.Button("Predict subtypes", variant="primary") |
| with gr.Column(): |
| msg = gr.Markdown() |
| dl = gr.File(label="Download per-cell CSV") |
| btn.click(run, inputs=[f, excl, src], outputs=[msg, dl]) |
| return demo |
|
|
|
|
| if __name__ == "__main__": |
| build().launch() |
|
|
