Update to 118K tissue-curated model (kNN demo mode)

app.py

@@ -7,17 +7,11 @@ Upload a gene expression matrix -> get UBERON tissue classification + embeddings
 v3 (May 2026): Trained on 118,263 tissue-curated samples from TCGA, GTEx, ARCHS4
 (cell lines excluded, classes balanced). 42 UBERON tissues, 94.9% balanced accuracy.
 
-Deployment note:
-    The full model weights (vae_tissue.final_model.pth, ~8 GB) exceed the
-    Hugging Face free-tier storage limit and are NOT shipped with this Space.
-    When the .pth is absent, the app runs in demo mode: uploaded samples are
-    projected with TruncatedSVD and classified by kNN against the pre-computed
-    118K reference embeddings. When the .pth IS present (e.g. running locally),
-    the app uses the trained VAE encoder for exact inference. Both modes use the
-    same 118K reference embeddings; only the encoding of new uploads differs.
-
-Run locally (full VAE inference, needs the .pth in model/):
-    streamlit run app.py
+When the full model weights (vae_tissue.final_model.pth, ~8 GB) are absent
+(Hugging Face free tier), the app runs in demo mode: uploaded samples are
+projected with TruncatedSVD and classified by kNN against the pre-computed
+118K reference embeddings. When the .pth IS present (local), the trained VAE
+encoder is used for exact inference.
 
 Author: Amit Pande, MDC Berlin/BIMSB
 """
@@ -30,8 +24,7 @@ from pathlib import Path
 from sklearn.neighbors import KNeighborsClassifier
 from collections import Counter
 
-# -- Config --
-MODEL_DIR = Path("model")
+MODEL_DIR = Path(".")
 K = 5
 LATENT_DIM = 121
 
@@ -43,7 +36,6 @@ st.markdown(
     "across **42 UBERON tissue categories** (94.9% balanced accuracy, 121-dim latent space)."
 )
 
-# -- Load model (optional) + artifacts + reference embeddings --
 @st.cache_resource
 def load_all():
     art        = joblib.load(MODEL_DIR / "vae_tissue.artifacts.joblib")
@@ -68,7 +60,6 @@ def load_all():
     knn = KNeighborsClassifier(n_neighbors=K, metric='cosine', n_jobs=-1)
     knn.fit(ref_emb.values, ref_clin['uberon_tissue'].values)
 
-    # Load the full VAE only if the weights are present (local / paid Space).
     model = None
     pth = MODEL_DIR / "vae_tissue.final_model.pth"
     if pth.exists():
@@ -77,7 +68,7 @@ def load_all():
             model = torch.load(pth, map_location='cpu', weights_only=False)
             model.eval()
         except Exception as ex:
-            st.sidebar.warning(f"Model weights present but not loaded ({ex}); using kNN demo mode.")
+            st.sidebar.warning(f"Weights present but not loaded ({ex}); using kNN demo mode.")
             model = None
 
     return model, gene_list, scaler, knn, ref_emb, ref_clin
@@ -96,15 +87,13 @@ try:
             "ℹ️ Running in **demo mode**. The 118K reference embeddings are used as-is; "
             "uploaded samples are projected with TruncatedSVD and classified by kNN. "
             "Full VAE encoding of new samples requires the model weights "
-            "(vae_tissue.final_model.pth), which are deposited at Zenodo and are too large "
-            "for the free Hugging Face tier."
+            "(vae_tissue.final_model.pth), deposited at Zenodo and too large for the free tier."
         )
 except Exception as e:
     st.error(f"Load error: {e}")
-    st.info("Place reference files in `model/` (embeddings_*.csv, *_clin.csv, vae_tissue.artifacts.joblib).")
+    st.info("Place reference files in repo root (embeddings_*.csv, *_clin.csv, vae_tissue.artifacts.joblib).")
     st.stop()
 
-# -- Helpers --
 def orient_matrix(df):
     genes_in_cols = len(set(df.columns) & set(gene_list))
     genes_in_rows = len(set(df.index)   & set(gene_list))
@@ -112,11 +101,10 @@ def orient_matrix(df):
         df = df.T
     return df
 
-def embed_and_classify(df_aligned_scaled):
-    """Return (embeddings, pred_labels, max_probs). Uses VAE if available, else SVD+kNN."""
+def embed_and_classify(X_scaled):
     if model is not None:
         import torch
-        X_t = torch.tensor(df_aligned_scaled, dtype=torch.float32)
+        X_t = torch.tensor(X_scaled, dtype=torch.float32)
         with torch.no_grad():
             h      = model.encoders[0](X_t)
             mu     = model.FC_mean(h[0])
@@ -133,10 +121,10 @@ def embed_and_classify(df_aligned_scaled):
         max_probs   = probs.numpy().max(axis=1)
     else:
         from sklearn.decomposition import TruncatedSVD
-        n_comp      = min(LATENT_DIM, df_aligned_scaled.shape[0] - 1, df_aligned_scaled.shape[1])
+        n_comp      = min(LATENT_DIM, X_scaled.shape[0] - 1, X_scaled.shape[1])
         n_comp      = max(n_comp, 1)
         svd         = TruncatedSVD(n_components=n_comp, random_state=42)
-        emb_reduced = svd.fit_transform(df_aligned_scaled)
+        emb_reduced = svd.fit_transform(X_scaled)
         if emb_reduced.shape[1] < LATENT_DIM:
             pad = np.zeros((emb_reduced.shape[0], LATENT_DIM - emb_reduced.shape[1]))
             emb = np.concatenate([emb_reduced, pad], axis=1)
@@ -146,14 +134,13 @@ def embed_and_classify(df_aligned_scaled):
         max_probs   = np.full(len(pred_labels), np.nan)
     return emb, pred_labels, max_probs
 
-# -- File upload --
 st.markdown("---")
 col1, col2 = st.columns([2, 1])
 with col1:
     uploaded = st.file_uploader(
         "Upload CSV/TSV (genes × samples or samples × genes)",
         type=["csv", "tsv", "txt"],
-        help="HGNC gene symbols. Log2-transformed expression values (TPM, RPKM, or counts).")
+        help="HGNC gene symbols. Log2-transformed expression values.")
 with col2:
     st.markdown(f"""
     **Expected input:**
@@ -227,4 +214,4 @@ st.caption(
     "Flexynesis Tissue VAE (v3) · Akalin Lab, MDC Berlin/BIMSB · "
     "118,263 samples · 42 UBERON tissues · 94.9% balanced accuracy · "
     "github.com/BIMSBbioinfo/flexynesis"
-)
+)

embeddings_test.csv

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:65c933c42a2e665bcbf8f37ee08e812446b29a6c084f4f36cde3e3f8af2402e2
-size 14762761
+oid sha256:b8353267526ba6fe1b189650a11eb9918282fa9537515f9893a74195263ef688
+size 37219853

embeddings_train.csv

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8effbf1939159417c5c3b217cb6321449149c1d1809b68270c3dcee4ab43a088
-size 83634269
+oid sha256:917d5c6677790a8725edaf78a589fa0c8cc675019548b2add49778f6edaeb2cc
+size 155547606

vae_tissue.artifacts.joblib

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9a2ca19cf98acb56bfa9d5b75c41cc912a293586954d178914da1286b86bd0b4
-size 523817
+oid sha256:2bafdb85f7da1e7976055e42329dcaf8d228cb15217692b6be9a36de5de0c29c
+size 658142