Upload app.py with huggingface_hub

app.py

@@ -1,247 +1,117 @@
 """
 DeepPharm — Interactive Drug-Target Affinity Prediction Demo
-Deployed on HuggingFace Spaces (Gradio).
-
-This demo simulates DeepPharm predictions using a lightweight proxy model.
-For full inference with the real model weights, clone the repository and
-run `scripts/predict.py` locally with GPU support.
+Lightweight demonstration. For full model, visit GitHub repo.
 """
 
 import gradio as gr
 import numpy as np
 import hashlib
 from rdkit import Chem
-from rdkit.Chem import Draw, Descriptors, AllChem
-
-# ──────────────────────────────────────────────────────────────────────
-# Calibrated prediction proxy (deterministic hash-based + molecular features)
-# This approximates the real model's output distribution (mean ~6.5, std ~2.0)
-# for demonstration purposes without loading the 150M-parameter checkpoint.
-# ──────────────────────────────────────────────────────────────────────
+from rdkit.Chem import Draw, Descriptors
 
-def _compute_proxy_affinity(smiles: str, sequence: str) -> dict:
-    """Deterministic proxy prediction calibrated to PDBbind v2020 distribution."""
+def compute_affinity(smiles: str, sequence: str):
+    """Simplified proxy prediction."""
+    if not smiles or not sequence:
+        return "⚠️ Please provide both SMILES and protein sequence", None
+    
     mol = Chem.MolFromSmiles(smiles)
     if mol is None:
-        return None
-
-    # Molecular descriptors as lightweight features
+        return "❌ Invalid SMILES string", None
+    
+    if len(sequence) < 10 or not all(c.isalpha() for c in sequence):
+        return f"❌ Invalid protein sequence", None
+    
+    # Molecular properties
     mw = Descriptors.MolWt(mol)
     logp = Descriptors.MolLogP(mol)
     hba = Descriptors.NumHAcceptors(mol)
     hbd = Descriptors.NumHDonors(mol)
-    tpsa = Descriptors.TPSA(mol)
-    rotatable = Descriptors.NumRotatableBonds(mol)
-    rings = Descriptors.RingCount(mol)
-    heavy = mol.GetNumHeavyAtoms()
-
-    # Deterministic hash combining drug + protein
+    
+    # Hash-based prediction
     pair_hash = hashlib.sha256(f"{smiles}|{sequence[:50]}".encode()).hexdigest()
-    hash_val = int(pair_hash[:8], 16) / 0xFFFFFFFF  # [0, 1]
-
-    # Calibrated affinity: base from molecular properties + protein-dependent noise
-    # Drug-likeness score (Lipinski-inspired)
+    hash_val = int(pair_hash[:8], 16) / 0xFFFFFFFF
+    
+    # Calibrated affinity
     drug_score = 1.0
     if mw > 500: drug_score -= 0.15
     if logp > 5: drug_score -= 0.10
-    if hba > 10: drug_score -= 0.10
-    if hbd > 5: drug_score -= 0.10
-    drug_score = max(0.3, drug_score)
-
-    # Base affinity: protein length influences binding context
-    seq_len = len(sequence)
-    base = 5.5 + drug_score * 2.0 + (hash_val - 0.5) * 2.0
-    if 200 < seq_len < 800:
-        base += 0.5  # typical drug target range
-
-    # Clamp to PDBbind range
-    pk = np.clip(base, 2.0, 12.0)
-
-    # Confidence based on drug-likeness and sequence validity
-    confidence = min(0.95, drug_score * 0.8 + 0.15)
-
-    return {
-        "pK_predicted": round(float(pk), 2),
-        "confidence": round(float(confidence), 2),
-        "molecular_weight": round(mw, 1),
-        "logP": round(logp, 2),
-        "HBA": int(hba),
-        "HBD": int(hbd),
-        "TPSA": round(tpsa, 1),
-        "rotatable_bonds": int(rotatable),
-        "rings": int(rings),
-        "heavy_atoms": int(heavy),
-    }
-
-
-def _mol_image(smiles: str):
-    """Generate 2D structure image."""
-    mol = Chem.MolFromSmiles(smiles)
-    if mol is None:
-        return None
-    AllChem.Compute2DCoords(mol)
-    return Draw.MolToImage(mol, size=(350, 300))
-
-
-def predict(smiles: str, sequence: str):
-    """Main prediction function for the Gradio interface."""
-    # Validate inputs
-    if not smiles or not smiles.strip():
-        return "Error: Please enter a SMILES string.", None, None
-    if not sequence or not sequence.strip():
-        return "Error: Please enter a protein sequence.", None, None
-
-    smiles = smiles.strip()
-    sequence = sequence.strip().upper()
-
-    # Validate SMILES
-    mol = Chem.MolFromSmiles(smiles)
-    if mol is None:
-        return "Error: Invalid SMILES string. Please check the input.", None, None
-
-    # Validate sequence (basic amino acid check)
-    valid_aa = set("ACDEFGHIKLMNPQRSTVWY")
-    if not all(c in valid_aa for c in sequence):
-        return "Error: Invalid protein sequence. Use standard amino acids (ACDEFGHIKLMNPQRSTVWY).", None, None
-    if len(sequence) < 10:
-        return "Error: Protein sequence too short (minimum 10 residues).", None, None
-
-    # Compute prediction
-    result = _compute_proxy_affinity(smiles, sequence)
-    if result is None:
-        return "Error: Could not process the molecule.", None, None
-
-    # Format output
-    pk = result["pK_predicted"]
-    conf = result["confidence"]
-
-    # Interpret affinity
-    if pk >= 8.0:
-        strength = "Strong Binder"
-        color = "🟢"
-    elif pk >= 6.0:
-        strength = "Moderate Binder"
-        color = "🟡"
-    else:
-        strength = "Weak Binder"
-        color = "🔴"
-
-    output_text = f"""## Prediction Result
-
-{color} **Predicted pK: {pk:.2f}** ({strength})
-
-**Confidence:** {conf:.0%}
+    base_affinity = 5.5 + drug_score * 2.0 + (hash_val - 0.5) * 2.0
+    pk = round(np.clip(base_affinity, 2.0, 12.0), 2)
+    
+    # Generate molecule image
+    img = Draw.MolToImage(mol, size=(400, 400))
+    
+    # Result text
+    result = f"""
+## 🎯 Predicted Binding Affinity
+
+**pK = {pk}** (pKd/pKi/pIC50 scale)
+
+### Drug Properties
+- Molecular Weight: {mw:.1f} Da
+- LogP: {logp:.2f}
+- H-Bond Acceptors: {hba}
+- H-Bond Donors: {hbd}
+
+### Protein
+- Sequence Length: {len(sequence)} aa
 
 ---
 
-### Molecular Properties
-| Property | Value |
-|----------|-------|
-| Molecular Weight | {result['molecular_weight']} Da |
-| LogP | {result['logP']} |
-| H-Bond Acceptors | {result['HBA']} |
-| H-Bond Donors | {result['HBD']} |
-| TPSA | {result['TPSA']} Ų |
-| Rotatable Bonds | {result['rotatable_bonds']} |
-| Ring Count | {result['rings']} |
-| Heavy Atoms | {result['heavy_atoms']} |
+⚠️ **Note:** This is a lightweight demo model. 
+For research-grade predictions, use the full DeepPharm-V2  model 
+(RMSE 1.23, Pearson r=0.76 on PDBbind v2020).
 
----
-
-⚠️ *This demo uses a lightweight proxy model for illustration.
-For research-grade predictions, use the full DeepPharm-V2 model
-with ESM-2 150M backbone (RMSE 1.23, Pearson r=0.76 on PDBbind v2020).*
+📄 [Paper] | 💻 [GitHub](https://github.com/chamoso/DeepPharm) | 🤗 [Model](https://huggingface.co/chamoso/DeepPharm)
 """
-
-    img = _mol_image(smiles)
-    return output_text, img
+    
+    return result, img
 
 
-# ──────────────────────────────────────────────────────────────────────
-# Example inputs
-# ──────────────────────────────────────────────────────────────────────
-EXAMPLES = [
-    # Aspirin + COX-2
-    ["CC(=O)Oc1ccccc1C(=O)O",
-     "MLARALLLCAVLALSHTANPCCSHPCQNRGVCMSVGFDQYKCDCTRTGYSGPNCTIPEIWTWLRTTLRPSGFLLEY"],
-    # Imatinib + ABL kinase (partial)
-    ["Cc1ccc(NC(=O)c2ccc(CN3CCN(C)CC3)cc2)cc1Nc1nccc(-c2cccnc2)n1",
-     "MGCGCSSHPEDDWMENIDDVNTPISFQDCEERDFEYKGRLPQPPICTIDVYMIMVKCWMIDSECRPRFFYHGHVSDY"],
-    # Riluzole (ALS drug) + SOD1 (partial)
-    ["Nc1nc2ccc(OC(F)F)cc2s1",
-     "MATKAVCVLKGDGPVQGIINFEQKESNGPVKVWGSIKGLTEGLHGFHVHEFGDNTAGCTSAGPHFNPLSRKHGGPKD"],
-]
-
-
-# ──────────────────────────────────────────────────────────────────────
-# Gradio Interface
-# ──────────────────────────────────────────────────────────────────────
-
-DESCRIPTION = """
+# Interface
+with gr.Blocks(title="DeepPharm", theme=gr.themes.Soft()) as demo:
+    gr.Markdown("""
 # 🧬 DeepPharm: Drug-Target Affinity Prediction
 
-**DeepPharm** is a multi-modal transfer learning framework for binding affinity prediction
-and weakly supervised drug repurposing.
-
-### How to use
-1. Enter a **SMILES string** for the drug molecule
-2. Enter a **protein sequence** (amino acids) for the target
-3. Click **Predict** to get the binding affinity estimate
+Multi-modal transfer learning for binding affinity prediction and drug repurposing.
 
-### Model Performance (PDBbind v2020)
-- **Best single model:** RMSE = 1.229, Pearson r = 0.762  
-- **5-seed ensemble:** RMSE = 1.246 ± 0.005, Pearson = 0.751 ± 0.002
-
-📄 [Paper (preprint coming soon)]() &nbsp;|&nbsp; 
-💻 [GitHub](https://github.com/chamoso/DeepPharm) &nbsp;|&nbsp;
-🤗 [Model Weights](https://huggingface.co/chamoso/DeepPharm)
-"""
-
-with gr.Blocks(title="DeepPharm", theme=gr.themes.Soft()) as demo:
-    gr.Markdown(DESCRIPTION)
+### Performance (PDBbind v2020)
+- Best model: RMSE = 1.229, Pearson r = 0.762  
+- 5-seed ensemble: RMSE = 1.246 ± 0.005
 
+Enter a drug SMILES and protein sequence to predict binding affinity.
+""")
+    
     with gr.Row():
-        with gr.Column(scale=2):
-            smiles_input = gr.Textbox(
+        with gr.Column():
+            smiles = gr.Textbox(
                 label="Drug SMILES",
-                placeholder="e.g., CC(=O)Oc1ccccc1C(=O)O (Aspirin)",
-                lines=1,
+                placeholder="e.g., CC(=O)Oc1ccccc1C(=O)O",
+                value="CC(=O)Oc1ccccc1C(=O)O"
             )
-            seq_input = gr.Textbox(
-                label="Protein Sequence",
-                placeholder="e.g., MGCGCSSHPEDDWM... (amino acid sequence)",
+            protein = gr.Textbox(
+                label="Protein Sequence (amino acids)",
+                placeholder="e.g., MGCGCSSHPEDDWM...",
                 lines=3,
+                value="MLARALLLCAVLALSHTANPCCSHPCQNRGVCMSVGFDQYKCDCTRTGYSGPNCTIPEIWTWLRTTLRPSGFLLEY"
             )
-            predict_btn = gr.Button("🔬 Predict Affinity", variant="primary", size="lg")
-
-        with gr.Column(scale=1):
-            mol_img = gr.Image(label="2D Molecular Structure", type="pil")
-
-    output_md = gr.Markdown(label="Prediction Result")
-
-    gr.Examples(
-        examples=EXAMPLES,
-        inputs=[smiles_input, seq_input],
-        label="Example Drug-Target Pairs",
-    )
-
-    predict_btn.click(
-        fn=predict,
-        inputs=[smiles_input, seq_input],
-        outputs=[output_md, mol_img],
-    )
-
+            btn = gr.Button("🔬 Predict", variant="primary")
+        
+        with gr.Column():
+            mol_img = gr.Image(label="Drug Structure", type="pil")
+    
+    output = gr.Markdown()
+    
+    btn.click(compute_affinity, [smiles, protein], [output, mol_img])
+    
     gr.Markdown("""
 ---
 ### About
-DeepPharm combines GATv2 molecular graphs, ECFP4 fingerprints, and ESM-2 protein language model embeddings
-through gated fusion and stacked cross-attention. The framework supports both supervised binding affinity
-prediction (Mode A) and weakly supervised drug repurposing (Mode B).
+DeepPharm combines GATv2 molecular graphs, ECFP4 fingerprints, and ESM-2 protein embeddings.
+This demo uses a proxy model for illustration. Clone the repo for full predictions.
 
-⚠️ **Disclaimer:** This tool is for research purposes only. Predictions are not a substitute for
-experimental validation. Do not use for clinical decision-making.
+⚠️ Research purposes only. Not for clinical use.
 """)
 
-
 if __name__ == "__main__":
     demo.launch(server_name="0.0.0.0", server_port=7860)