Spaces:
Configuration error
Configuration error
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,97 +1,82 @@
|
|
| 1 |
-
|
| 2 |
-
# BulkRNABert Inference - Working Version
|
| 3 |
-
# ============================================================
|
| 4 |
-
|
| 5 |
-
# STEP 1: Install the compatible transformers version
|
| 6 |
-
# Community-verified fix for the 'attention_maps_to_save' error
|
| 7 |
-
!pip install -q "transformers<=4.33.3" torch
|
| 8 |
-
|
| 9 |
-
from huggingface_hub import hf_hub_download
|
| 10 |
import numpy as np
|
| 11 |
import pandas as pd
|
|
|
|
|
|
|
| 12 |
from transformers import AutoConfig, AutoModel, AutoTokenizer
|
| 13 |
|
| 14 |
-
|
|
|
|
| 15 |
config = AutoConfig.from_pretrained(
|
| 16 |
"InstaDeepAI/BulkRNABert",
|
| 17 |
trust_remote_code=True,
|
| 18 |
)
|
| 19 |
-
config.embeddings_layers_to_save = (4,)
|
| 20 |
|
| 21 |
tokenizer = AutoTokenizer.from_pretrained(
|
| 22 |
"InstaDeepAI/BulkRNABert",
|
| 23 |
-
trust_remote_code=True
|
| 24 |
)
|
|
|
|
| 25 |
model = AutoModel.from_pretrained(
|
| 26 |
"InstaDeepAI/BulkRNABert",
|
| 27 |
config=config,
|
| 28 |
-
trust_remote_code=True
|
| 29 |
)
|
|
|
|
| 30 |
print(f"✓ Model loaded. Required genes: {config.n_genes}")
|
| 31 |
|
| 32 |
-
# STEP 3: Download reference data (defines the required gene set)
|
| 33 |
-
# The model expects exactly 19,062 genes in a specific order [citation:6]
|
| 34 |
csv_path = hf_hub_download(
|
| 35 |
repo_id="InstaDeepAI/BulkRNABert",
|
| 36 |
filename="data/tcga_sample.csv",
|
| 37 |
repo_type="model",
|
| 38 |
)
|
| 39 |
-
|
| 40 |
-
# STEP 4: Load YOUR data (adjust paths as needed)
|
| 41 |
-
tcga = pd.read_csv(
|
| 42 |
-
"/content/drive/MyDrive/TCGA_UCEC_INPUT/TCGA_UCEC_582_SAMPLE_LOG2CPM.tsv",
|
| 43 |
-
sep="\t", index_col=0
|
| 44 |
-
).T
|
| 45 |
-
cptac = pd.read_csv(
|
| 46 |
-
"/content/drive/MyDrive/UCEC_paper/CPTAC_expression_matrix.tsv",
|
| 47 |
-
sep="\t", index_col=0
|
| 48 |
-
).T
|
| 49 |
-
|
| 50 |
-
print(f"TCGA: {tcga.shape}")
|
| 51 |
-
print(f"CPTAC: {cptac.shape}")
|
| 52 |
-
|
| 53 |
-
# STEP 5: Align genes to the required 19,062-gene set
|
| 54 |
ref_df = pd.read_csv(csv_path)
|
| 55 |
ref_genes = [c for c in ref_df.columns if c != "identifier"]
|
| 56 |
-
print(f"
|
|
|
|
| 57 |
|
| 58 |
def align_genes(df, required_genes):
|
| 59 |
-
"""Align dataframe to required gene set, fill missing with 0."""
|
| 60 |
aligned = pd.DataFrame(0, index=df.index, columns=required_genes)
|
| 61 |
for g in required_genes:
|
| 62 |
if g in df.columns:
|
| 63 |
aligned[g] = df[g]
|
| 64 |
return aligned
|
| 65 |
|
| 66 |
-
tcga_aligned = align_genes(tcga, ref_genes)
|
| 67 |
-
cptac_aligned = align_genes(cptac, ref_genes)
|
| 68 |
|
| 69 |
-
# STEP 6: Generate embeddings (official pipeline)
|
| 70 |
def get_embeddings(df, model, tokenizer, config):
|
| 71 |
-
"""Generate BulkRNABert embeddings following the official example."""
|
| 72 |
expr = df.to_numpy()
|
| 73 |
-
expr = np.log10(1 + expr)
|
| 74 |
-
|
| 75 |
assert expr.shape[1] == config.n_genes, \
|
| 76 |
f"Expected {config.n_genes} genes, got {expr.shape[1]}"
|
| 77 |
-
|
| 78 |
ids = tokenizer.batch_encode_plus(expr, return_tensors="pt")["input_ids"]
|
| 79 |
-
|
| 80 |
-
import torch
|
| 81 |
with torch.no_grad():
|
| 82 |
out = model(ids)
|
| 83 |
-
|
| 84 |
-
# Official: extract layer-4 embeddings, average across genes [citation:1]
|
| 85 |
return out["embeddings_4"].mean(axis=1).numpy()
|
| 86 |
|
| 87 |
-
tcga_emb = get_embeddings(tcga_aligned, model, tokenizer, config)
|
| 88 |
-
cptac_emb = get_embeddings(cptac_aligned, model, tokenizer, config)
|
| 89 |
|
| 90 |
-
|
| 91 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 92 |
|
| 93 |
-
|
| 94 |
-
save_dir = "/content/drive/MyDrive/UCEC_paper"
|
| 95 |
-
np.save(f"{save_dir}/tcga_bulkrnabert_emb.npy", tcga_emb)
|
| 96 |
-
np.save(f"{save_dir}/cptac_bulkrnabert_emb.npy", cptac_emb)
|
| 97 |
-
print(f"✓ Saved to {save_dir}/")
|
|
|
|
| 1 |
+
import gradio as gr
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
import numpy as np
|
| 3 |
import pandas as pd
|
| 4 |
+
import torch
|
| 5 |
+
from huggingface_hub import hf_hub_download
|
| 6 |
from transformers import AutoConfig, AutoModel, AutoTokenizer
|
| 7 |
|
| 8 |
+
print("Loading BulkRNABert...")
|
| 9 |
+
|
| 10 |
config = AutoConfig.from_pretrained(
|
| 11 |
"InstaDeepAI/BulkRNABert",
|
| 12 |
trust_remote_code=True,
|
| 13 |
)
|
| 14 |
+
config.embeddings_layers_to_save = (4,)
|
| 15 |
|
| 16 |
tokenizer = AutoTokenizer.from_pretrained(
|
| 17 |
"InstaDeepAI/BulkRNABert",
|
| 18 |
+
trust_remote_code=True,
|
| 19 |
)
|
| 20 |
+
|
| 21 |
model = AutoModel.from_pretrained(
|
| 22 |
"InstaDeepAI/BulkRNABert",
|
| 23 |
config=config,
|
| 24 |
+
trust_remote_code=True,
|
| 25 |
)
|
| 26 |
+
model.eval()
|
| 27 |
print(f"✓ Model loaded. Required genes: {config.n_genes}")
|
| 28 |
|
|
|
|
|
|
|
| 29 |
csv_path = hf_hub_download(
|
| 30 |
repo_id="InstaDeepAI/BulkRNABert",
|
| 31 |
filename="data/tcga_sample.csv",
|
| 32 |
repo_type="model",
|
| 33 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 34 |
ref_df = pd.read_csv(csv_path)
|
| 35 |
ref_genes = [c for c in ref_df.columns if c != "identifier"]
|
| 36 |
+
print(f"✓ Reference genes: {len(ref_genes)}")
|
| 37 |
+
|
| 38 |
|
| 39 |
def align_genes(df, required_genes):
|
|
|
|
| 40 |
aligned = pd.DataFrame(0, index=df.index, columns=required_genes)
|
| 41 |
for g in required_genes:
|
| 42 |
if g in df.columns:
|
| 43 |
aligned[g] = df[g]
|
| 44 |
return aligned
|
| 45 |
|
|
|
|
|
|
|
| 46 |
|
|
|
|
| 47 |
def get_embeddings(df, model, tokenizer, config):
|
|
|
|
| 48 |
expr = df.to_numpy()
|
| 49 |
+
expr = np.log10(1 + expr)
|
|
|
|
| 50 |
assert expr.shape[1] == config.n_genes, \
|
| 51 |
f"Expected {config.n_genes} genes, got {expr.shape[1]}"
|
|
|
|
| 52 |
ids = tokenizer.batch_encode_plus(expr, return_tensors="pt")["input_ids"]
|
|
|
|
|
|
|
| 53 |
with torch.no_grad():
|
| 54 |
out = model(ids)
|
|
|
|
|
|
|
| 55 |
return out["embeddings_4"].mean(axis=1).numpy()
|
| 56 |
|
|
|
|
|
|
|
| 57 |
|
| 58 |
+
def process_file(file):
|
| 59 |
+
if file is None:
|
| 60 |
+
return {"error": "No file uploaded"}
|
| 61 |
+
try:
|
| 62 |
+
df = pd.read_csv(file.name, sep="\t", index_col=0).T
|
| 63 |
+
aligned = align_genes(df, ref_genes)
|
| 64 |
+
emb = get_embeddings(aligned, model, tokenizer, config)
|
| 65 |
+
return {
|
| 66 |
+
"status": "success",
|
| 67 |
+
"n_samples": int(emb.shape[0]),
|
| 68 |
+
"embedding_dim": int(emb.shape[1]),
|
| 69 |
+
"first_sample_preview": emb[0][:10].tolist(),
|
| 70 |
+
}
|
| 71 |
+
except Exception as e:
|
| 72 |
+
return {"error": str(e)}
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
iface = gr.Interface(
|
| 76 |
+
fn=process_file,
|
| 77 |
+
inputs=gr.File(label="Upload TSV expression file (genes × samples)"),
|
| 78 |
+
outputs=gr.JSON(label="Embeddings Result"),
|
| 79 |
+
title="BulkRNABert Embeddings",
|
| 80 |
+
)
|
| 81 |
|
| 82 |
+
iface.launch(server_name="0.0.0.0", server_port=7860)
|
|
|
|
|
|
|
|
|
|
|
|