app.py

"""
Interactive Crystallization Component Predictor
===============================================
Streamlit app for Hugging Face Hub deployment
Predicts crystallization components using Simple Baseline and Advanced Baseline models
"""

import streamlit as st
import pandas as pd
import numpy as np
import joblib
import json
import os
import warnings

warnings.filterwarnings('ignore')

# Page config
st.set_page_config(
    page_title="Crystallization Predictor",
    page_icon="🔬",
    layout="wide",
    initial_sidebar_state="expanded"
)

# Get the directory of this script
BASE_DIR = os.path.dirname(os.path.abspath(__file__))

# Title and Introduction
st.title("🔬 Crystallization Component Predictor")
st.markdown("""
### Predict crystallization components using Machine Learning
This app uses trained machine learning models to predict the optimal components for protein crystallization
based on your experimental parameters.
""")
st.markdown("---")

# Sidebar
st.sidebar.header("⚙️ Model Selection")
approach = st.sidebar.radio(
    "Choose Approach:",
    ["Advanced Baseline (Recommended)", "Simple Baseline"],
    help="Advanced has concentration parsing and better accuracy"
)

st.sidebar.markdown("---")
st.sidebar.markdown("### 📊 Model Performance")

# Display performance metrics
try:
    simple_results_path = os.path.join(BASE_DIR, 'models', 'simple_baseline', 'training_results.json')
    advanced_results_path = os.path.join(BASE_DIR, 'models', 'advanced_baseline', 'training_results.json')
    
    if os.path.exists(simple_results_path):
        with open(simple_results_path, 'r') as f:
            simple_results = json.load(f)
    if os.path.exists(advanced_results_path):
        with open(advanced_results_path, 'r') as f:
            advanced_results = json.load(f)
    
    if "Simple" in approach:
        st.sidebar.metric("Name Accuracy", "61.12%")
        st.sidebar.metric("pH R²", "95.58%")
        st.sidebar.warning("⚠️ Conc: N/A")
    else:
        st.sidebar.metric("Name Accuracy", "64.18%")
        st.sidebar.metric("Conc R²", "47.33%")
        st.sidebar.metric("pH R²", "99.34%")
        st.sidebar.success("✅ All metrics working!")
        
except Exception as e:
    st.sidebar.info(f"Using default metrics")

st.sidebar.markdown("---")
st.sidebar.markdown("""
### ℹ️ About
This tool predicts three key crystallization parameters:
- **Component Name**: The chemical compound
- **Concentration**: Amount in solution (M)
- **pH**: Acidity/basicity level

**Recommended:** Advanced Baseline for complete predictions
""")

# Input Form
st.header("🎯 Input Crystallization Parameters")

col1, col2 = st.columns(2)

with col1:
    st.markdown("#### Crystallization Setup")
    cryst_method = st.selectbox(
        "Crystallization Method",
        [
            "VAPOR DIFFUSION, SITTING DROP",
            "VAPOR DIFFUSION, HANGING DROP",
            "VAPOR DIFFUSION",
            "BATCH MODE",
            "MICROBATCH"
        ],
        help="Select the crystallization technique you're using"
    )
    
    temp = st.slider(
        "Temperature (K)", 
        250.0, 320.0, 293.0, 1.0,
        help="Typical room temperature is ~293K (20°C)"
    )
    
    ph = st.slider(
        "pH", 
        0.0, 14.0, 7.0, 0.1,
        help="Initial pH of your crystallization solution"
    )

with col2:
    st.markdown("#### Crystal Properties")
    matthews = st.slider(
        "Matthews Coefficient", 
        1.0, 4.5, 2.2, 0.1,
        help="Ratio of unit cell volume to protein molecular weight (ų/Da)"
    )
    
    solvent = st.slider(
        "Percent Solvent Content (%)", 
        0.0, 100.0, 45.0, 1.0,
        help="Percentage of solvent in the crystal"
    )

st.markdown("---")

# Predict button
if st.button("🚀 Predict Components", type="primary", use_container_width=True):
    
    try:
        with st.spinner("🔄 Loading models and making predictions..."):
            
            if "Advanced" in approach:
                # Load advanced models
                model_name = joblib.load(os.path.join(BASE_DIR, 'models', 'advanced_baseline', 'model_component_name.pkl'))
                model_conc = joblib.load(os.path.join(BASE_DIR, 'models', 'advanced_baseline', 'model_component_conc.pkl'))
                model_ph = joblib.load(os.path.join(BASE_DIR, 'models', 'advanced_baseline', 'model_component_ph.pkl'))
                le = joblib.load(os.path.join(BASE_DIR, 'models', 'advanced_baseline', 'label_encoder_name.pkl'))
                scaler = joblib.load(os.path.join(BASE_DIR, 'models', 'advanced_baseline', 'scaler.pkl'))
                tfidf = joblib.load(os.path.join(BASE_DIR, 'models', 'advanced_baseline', 'tfidf.pkl'))
                
                # Feature engineering (Advanced Baseline needs 8 features)
                temp_ph_int = temp * ph
                matthews_solvent_int = matthews * solvent
                ph_diff = 0  # Unknown for new prediction
                solvent_ratio = solvent / (matthews + 1e-6)
                
                numerical = np.array([[temp, ph, matthews, solvent,
                                      temp_ph_int, matthews_solvent_int,
                                      ph_diff, solvent_ratio]])
                
            else:
                # Load simple models
                model_name = joblib.load(os.path.join(BASE_DIR, 'models', 'simple_baseline', 'model_component_name.pkl'))
                model_ph = joblib.load(os.path.join(BASE_DIR, 'models', 'simple_baseline', 'model_component_ph.pkl'))
                le = joblib.load(os.path.join(BASE_DIR, 'models', 'simple_baseline', 'label_encoder_name.pkl'))
                scaler = joblib.load(os.path.join(BASE_DIR, 'models', 'simple_baseline', 'scaler.pkl'))
                tfidf = joblib.load(os.path.join(BASE_DIR, 'models', 'simple_baseline', 'tfidf.pkl'))
                
                # Simple baseline: only 4 features
                numerical = np.array([[temp, ph, matthews, solvent]])
            
            # Scale numerical features
            numerical_scaled = scaler.transform(numerical)
            
            # TF-IDF for crystallization method
            method_tfidf = tfidf.transform([cryst_method.upper()]).toarray()
            
            # Combine features
            X_pred = np.concatenate([numerical_scaled, method_tfidf], axis=1)
            
            # Make predictions
            pred_name_idx = model_name.predict(X_pred)[0]
            pred_name = le.inverse_transform([pred_name_idx])[0]
            pred_name_proba = model_name.predict_proba(X_pred)[0]
            top_5_idx = np.argsort(pred_name_proba)[-5:][::-1]
            top_5_names = le.inverse_transform(top_5_idx)
            top_5_proba = pred_name_proba[top_5_idx]
            
            pred_ph = model_ph.predict(X_pred)[0]
            
            if "Advanced" in approach:
                pred_conc = model_conc.predict(X_pred)[0]
            
        # Display Results
        st.success("✅ Predictions Complete!")
        st.markdown("---")
        
        st.header("📊 Prediction Results")
        
        # Component Name
        st.subheader("1️⃣ Component_1_Name")
        st.markdown("**Most likely chemical component for crystallization:**")
        
        col1, col2 = st.columns([1, 2])
        
        with col1:
            st.metric("Predicted Component", pred_name)
            st.caption("Top prediction from the model")
        
        with col2:
            st.markdown("**Top 5 Predictions (with confidence):**")
            top5_df = pd.DataFrame({
                'Rank': range(1, 6),
                'Component': top_5_names,
                'Probability': [f"{p:.2%}" for p in top_5_proba]
            })
            st.dataframe(top5_df, hide_index=True, use_container_width=True)
        
        st.markdown("---")
        
        # Concentration
        st.subheader("2️⃣ Component_1_Conc")
        if "Advanced" in approach:
            col1, col2 = st.columns(2)
            with col1:
                st.metric("Predicted Concentration (log-scale)", f"{pred_conc:.4f}")
            with col2:
                actual_molarity = 10**pred_conc
                st.metric("Actual Molarity", f"{actual_molarity:.6f} M")
            
            st.info(f"💡 Use approximately **{actual_molarity:.6f} M** of {pred_name} in your crystallization trials")
        else:
            st.warning("⚠️ Not available in Simple Baseline - use Advanced Baseline for concentration predictions")
        
        st.markdown("---")
        
        # pH
        st.subheader("3️⃣ Component_1_pH")
        col1, col2 = st.columns([1, 2])
        
        with col1:
            st.metric("Predicted pH", f"{pred_ph:.2f}")
            
            # pH classification
            if pred_ph < 6:
                ph_class = "Acidic"
                ph_emoji = "🔴"
            elif pred_ph < 8:
                ph_class = "Neutral"
                ph_emoji = "🟢"
            else:
                ph_class = "Basic"
                ph_emoji = "🔵"
            
            st.caption(f"{ph_emoji} {ph_class} solution")
        
        with col2:
            # pH visualization
            ph_percent = (pred_ph / 14) * 100
            ph_color = "red" if pred_ph < 6 else ("green" if pred_ph < 8 else "blue")
            st.markdown(f"""
            <div style='background: linear-gradient(to right, red, yellow, green, cyan, blue); 
                        height: 40px; border-radius: 10px; margin: 10px 0; border: 2px solid #333;'></div>
            <div style='display: flex; justify-content: space-between; font-size: 14px;'>
                <span><b>0</b> (Acidic)</span>
                <span><b>7</b> (Neutral)</span>
                <span><b>14</b> (Basic)</span>
            </div>
            <div style='text-align: center; margin-top: 15px;'>
                <b style='font-size: 24px; color: {ph_color};'>pH = {pred_ph:.2f}</b>
            </div>
            """, unsafe_allow_html=True)
        
        st.info(f"💡 Adjust your buffer to maintain pH ≈ **{pred_ph:.2f}** for optimal crystallization")
        
        # Input Summary
        st.markdown("---")
        st.subheader("📥 Input Summary")
        input_df = pd.DataFrame({
            'Parameter': [
                'Crystallization Method', 
                'Temperature', 
                'Input pH', 
                'Matthews Coefficient', 
                'Solvent Content'
            ],
            'Value': [
                cryst_method, 
                f"{temp:.1f} K ({temp-273.15:.1f}°C)", 
                f"{ph:.1f}", 
                f"{matthews:.2f} Ų/Da", 
                f"{solvent:.1f}%"
            ]
        })
        st.table(input_df)
        
        # Download Results
        st.markdown("---")
        st.subheader("💾 Download Results")
        
        results_dict = {
            'Crystallization Method': cryst_method,
            'Temperature (K)': temp,
            'Temperature (°C)': temp - 273.15,
            'Input pH': ph,
            'Matthews Coefficient': matthews,
            'Solvent Content (%)': solvent,
            'Predicted Component': pred_name,
            'Component Probability': f"{top_5_proba[0]:.4f}",
            'Predicted pH': f"{pred_ph:.2f}",
        }
        
        if "Advanced" in approach:
            results_dict['Predicted Concentration (log)'] = f"{pred_conc:.4f}"
            results_dict['Predicted Concentration (M)'] = f"{10**pred_conc:.6f}"
        
        results_df = pd.DataFrame([results_dict])
        csv = results_df.to_csv(index=False)
        
        st.download_button(
            label="📥 Download Predictions as CSV",
            data=csv,
            file_name="crystallization_predictions.csv",
            mime="text/csv",
        )
        
    except FileNotFoundError as e:
        st.error(f"""
        ❌ **Model files not found!**
        
        Error: {e}
        
        Please ensure model files are in the correct directory:
        - `models/simple_baseline/`
        - `models/advanced_baseline/`
        """)
    except Exception as e:
        st.error(f"❌ **Prediction Error:** {e}")
        with st.expander("🔍 Show full error details"):
            import traceback
            st.code(traceback.format_exc())

# Model Comparison Section
st.markdown("---")
st.header("📈 Model Comparison")

comparison_df = pd.DataFrame({
    'Model': ['Simple Baseline', 'Advanced Baseline', 'Transformer'],
    'Name Accuracy': ['61.12%', '64.18% ⭐', '53.85%'],
    'Conc R²': ['N/A', '47.33%', '18.72%'],
    'pH R²': ['95.58%', '99.34% ⭐', '99.27%'],
    'Speed': ['⚡ Fast', '⚡ Fast', '🐌 Slow'],
    'Recommendation': ['Basic use', '✅ Best overall', 'Research only']
})

st.dataframe(
    comparison_df, 
    hide_index=True, 
    use_container_width=True,
    column_config={
        "Model": st.column_config.TextColumn("Model", width="medium"),
        "Name Accuracy": st.column_config.TextColumn("Name Accuracy", width="medium"),
        "Conc R²": st.column_config.TextColumn("Concentration R²", width="medium"),
        "pH R²": st.column_config.TextColumn("pH R²", width="medium"),
    }
)

st.markdown("""
**Model Selection Guide:**
- **Simple Baseline**: Fast predictions, no concentration. Good for quick pH and component estimates.
- **Advanced Baseline**: ⭐ Recommended for most users. Includes all three predictions with high accuracy.
- **Transformer**: Deep learning approach, requires more data for better performance.
""")

# Visualizations Section
st.markdown("---")
st.header("📊 Performance Visualizations")

viz_path = os.path.join(BASE_DIR, 'visualizations')

if os.path.exists(viz_path):
    try:
        tab1, tab2, tab3, tab4 = st.tabs([
            "📊 Name Accuracy", 
            "📈 Concentration R²", 
            "🧪 pH R²", 
            "🎯 Complete Comparison"
        ])
        
        with tab1:
            img_path = os.path.join(viz_path, '01_component_name_comparison.png')
            if os.path.exists(img_path):
                st.image(img_path, use_column_width=True)
                st.caption("Comparison of component name prediction accuracy across all models")
                
        with tab2:
            img_path = os.path.join(viz_path, '02_component_conc_comparison.png')
            if os.path.exists(img_path):
                st.image(img_path, use_column_width=True)
                st.caption("Concentration prediction performance (R² scores)")
                
        with tab3:
            img_path = os.path.join(viz_path, '03_component_ph_comparison.png')
            if os.path.exists(img_path):
                st.image(img_path, use_column_width=True)
                st.caption("pH prediction performance (R² scores)")
                
        with tab4:
            img_path = os.path.join(viz_path, '05_complete_comparison.png')
            if os.path.exists(img_path):
                st.image(img_path, use_column_width=True)
                st.caption("Comprehensive comparison of all approaches and metrics")
    except Exception as e:
        st.info(f"Visualizations are being loaded... {e}")
else:
    st.info("📊 Visualization files not found in this deployment")

# Information Section
st.markdown("---")
st.header("ℹ️ How It Works")

with st.expander("🔬 About Protein Crystallization"):
    st.markdown("""
    **Protein crystallization** is a crucial step in structural biology for determining 3D protein structures using X-ray crystallography.
    
    **Key Parameters:**
    - **Crystallization Method**: The technique used (e.g., vapor diffusion, batch mode)
    - **Temperature**: Affects protein stability and crystal growth
    - **pH**: Critical for protein solubility and crystal formation
    - **Matthews Coefficient**: Indicates crystal packing density
    - **Solvent Content**: Amount of solvent in the crystal lattice
    
    This tool helps predict optimal conditions based on historical crystallization data.
    """)

with st.expander("🤖 About the Models"):
    st.markdown("""
    **Simple Baseline:**
    - Random Forest classifier for component name
    - XGBoost regressor for pH
    - Uses 4 numerical features + TF-IDF of method
    
    **Advanced Baseline:**
    - Ensemble of Random Forest, XGBoost, LightGBM, and CatBoost
    - Includes concentration prediction with log-transformation
    - Uses 8 engineered features including interactions
    - Best overall performance: 64% name accuracy, 99% pH R²
    
    **Training Data:**
    - Based on protein crystallization experiments from PDB
    - Includes various crystallization methods and conditions
    - Models trained on structured crystallization data
    """)

with st.expander("📖 How to Use"):
    st.markdown("""
    1. **Select a model** in the sidebar (Advanced Baseline recommended)
    2. **Input your parameters**:
       - Choose crystallization method
       - Set temperature, pH, Matthews coefficient, and solvent content
    3. **Click "Predict Components"** to get predictions
    4. **Review results**:
       - Component name with confidence scores
       - Concentration (if using Advanced Baseline)
       - Optimal pH for crystallization
    5. **Download** results as CSV for your records
    
    💡 **Tip:** Start with the recommended default values and adjust based on your specific protein and experimental setup.
    """)

# Footer
st.markdown("---")
st.markdown("""
<div style='text-align: center; color: gray; padding: 20px;'>
    <p><b>🔬 Crystallization Component Prediction System</b></p>
    <p><i>Advanced Baseline achieves: 64% Name Accuracy | 47% Conc R² | 99% pH R²</i></p>
    <p>Built with Scikit-learn, XGBoost, LightGBM, CatBoost & Streamlit</p>
    <p style='font-size: 12px; margin-top: 10px;'>
        For research and educational purposes. Validate predictions experimentally.
    </p>
</div>
""", unsafe_allow_html=True)