app.py

import gradio as gr
import torch
import joblib
import numpy as np
import torch.nn as nn
import matplotlib.pyplot as plt
import io
from PIL import Image
from itertools import product

# --------------- Model Definition ---------------

class VirusClassifier(nn.Module):
    def __init__(self, input_shape: int):
        super(VirusClassifier, self).__init__()
        self.network = nn.Sequential(
            nn.Linear(input_shape, 64),
            nn.GELU(),
            nn.BatchNorm1d(64),
            nn.Dropout(0.3),
            nn.Linear(64, 32),
            nn.GELU(),
            nn.BatchNorm1d(32),
            nn.Dropout(0.3),
            nn.Linear(32, 32),
            nn.GELU(),
            nn.Linear(32, 2)
        )

    def forward(self, x):
        return self.network(x)
    
    def get_gradient_importance(self, x, class_index=1):
        """
        Calculate gradient-based importance for each input feature.
        By default, we compute the gradient wrt the 'human' class (index=1).
        This method is akin to a raw gradient or 'saliency' approach.
        """
        x = x.clone().detach().requires_grad_(True)
        output = self.network(x)
        probs = torch.softmax(output, dim=1)
        
        # Probability of the specified class
        target_prob = probs[..., class_index]
        
        # Zero existing gradients if any
        if x.grad is not None:
            x.grad.zero_()
        
        # Backprop on that probability
        target_prob.backward()
        
        # Raw gradient is now in x.grad
        importance = x.grad.detach()
        
        # Optional: Multiply by input to get a more "integrated gradients"-like measure
        # importance = importance * x.detach()
        
        return importance, float(target_prob)

# --------------- Utility Functions ---------------

def parse_fasta(text: str):
    """
    Parse a FASTA string and return a list of (header, sequence) pairs.
    """
    sequences = []
    current_header = None
    current_sequence = []
    
    for line in text.split('\n'):
        line = line.strip()
        if not line:
            continue
        if line.startswith('>'):
            if current_header:
                sequences.append((current_header, ''.join(current_sequence)))
            current_header = line[1:]
            current_sequence = []
        else:
            current_sequence.append(line.upper())
    if current_header:
        sequences.append((current_header, ''.join(current_sequence)))
    return sequences

def sequence_to_kmer_vector(sequence: str, k: int = 4) -> np.ndarray:
    """
    Convert a nucleotide sequence into a k-mer frequency vector.
    Defaults to k=4.
    """
    # Generate all possible k-mers
    kmers = [''.join(p) for p in product("ACGT", repeat=k)]
    kmer_dict = {km: i for i, km in enumerate(kmers)}
    vec = np.zeros(len(kmers), dtype=np.float32)
    
    for i in range(len(sequence) - k + 1):
        kmer = sequence[i:i+k]
        if kmer in kmer_dict:
            vec[kmer_dict[kmer]] += 1

    total_kmers = len(sequence) - k + 1
    if total_kmers > 0:
        vec = vec / total_kmers

    return vec

def compute_sequence_stats(sequence: str):
    """
    Compute various statistics for a given sequence:
      - Length
      - GC content (%)
      - A/C/G/T counts
    """
    length = len(sequence)
    if length == 0:
        return {
            'length': 0,
            'gc_content': 0,
            'counts': {'A': 0, 'C': 0, 'G': 0, 'T': 0}
        }

    counts = {
        'A': sequence.count('A'),
        'C': sequence.count('C'),
        'G': sequence.count('G'),
        'T': sequence.count('T')
    }
    gc_content = (counts['G'] + counts['C']) / length * 100.0

    return {
        'length': length,
        'gc_content': gc_content,
        'counts': counts
    }

# --------------- Visualization Functions ---------------

def plot_shap_like_bars(kmers, importance_values, top_k=10):
    """
    Create a bar chart that mimics a SHAP summary plot:
      - k-mers on y-axis
      - importance magnitude on x-axis
      - color indicating positive (push towards human) vs negative (push towards non-human)
    """
    abs_importance = np.abs(importance_values)
    # Sort by absolute importance
    sorted_indices = np.argsort(abs_importance)[::-1]
    top_indices = sorted_indices[:top_k]
    
    # Prepare data
    top_kmers = [kmers[i] for i in top_indices]
    top_importances = importance_values[top_indices]
    
    # Create plot
    fig, ax = plt.subplots(figsize=(8, 6))
    colors = ['green' if val > 0 else 'red' for val in top_importances]
    ax.barh(range(len(top_kmers)), np.abs(top_importances), color=colors)
    ax.set_yticks(range(len(top_kmers)))
    ax.set_yticklabels(top_kmers)
    ax.invert_yaxis()  # So that the highest value is at the top
    ax.set_xlabel("Feature Importance (Gradient Magnitude)")
    ax.set_title(f"Top-{top_k} SHAP-like Feature Importances")
    plt.tight_layout()
    return fig

def plot_kmer_distribution(kmer_freq_vector, kmers):
    """
    Plot a histogram of k-mer frequencies for the entire vector.
    (Optional if you want a quick distribution overview)
    """
    fig, ax = plt.subplots(figsize=(10, 4))
    ax.bar(range(len(kmer_freq_vector)), kmer_freq_vector, color='blue', alpha=0.6)
    ax.set_xlabel("K-mer Index")
    ax.set_ylabel("Frequency")
    ax.set_title("K-mer Frequency Distribution")
    ax.set_xticks([])
    plt.tight_layout()
    return fig

def create_step_visualization(important_kmers, human_prob):
    """
    Re-implementation of your step-wise probability plot.
    Shows how each top k-mer 'pushes' the probability from 0.5 to the final value.
    """
    fig = plt.figure(figsize=(8, 5))
    ax = fig.add_subplot(111)
    
    # Start from 0.5
    current_prob = 0.5
    steps = [('Start', current_prob, 0)]
    
    for kmer in important_kmers:
        change = kmer['impact'] * (-1 if kmer['direction'] == 'non-human' else 1)
        current_prob += change
        steps.append((kmer['kmer'], current_prob, change))

    x_vals = range(len(steps))
    y_vals = [s[1] for s in steps]

    ax.step(x_vals, y_vals, 'b-', where='post', label='Probability', linewidth=2)
    ax.plot(x_vals, y_vals, 'b.', markersize=10)
    
    # Reference line at 0.5
    ax.axhline(y=0.5, color='r', linestyle='--', label='Neutral (0.5)')
    ax.set_ylim(0, 1)
    ax.set_ylabel('Human Probability')
    ax.set_title(f'K-mer Contributions (final p={human_prob:.3f})')
    ax.grid(True, linestyle='--', alpha=0.7)

    for i, (kmer, prob, change) in enumerate(steps):
        ax.annotate(kmer, 
                    (i, prob),
                    xytext=(0, 10 if i % 2 == 0 else -20),
                    textcoords='offset points',
                    ha='center',
                    rotation=45)
        
        if i > 0:
            change_text = f'{change:+.3f}'
            color = 'green' if change > 0 else 'red'
            ax.annotate(change_text,
                        (i, prob),
                        xytext=(0, -20 if i % 2 == 0 else 10),
                        textcoords='offset points',
                        ha='center',
                        color=color)

    ax.legend()
    plt.tight_layout()
    return fig

def plot_kmer_freq_and_sigma(important_kmers):
    """
    Plot frequencies vs. sigma from mean for the top k-mers.
    This reuses logic from the original create_visualization second subplot,
    but as its own function for clarity.
    """
    fig, ax = plt.subplots(figsize=(8, 5))
    
    # Prepare data
    kmers = [k['kmer'] for k in important_kmers]
    frequencies = [k['occurrence'] for k in important_kmers]
    sigmas = [k['sigma'] for k in important_kmers]
    colors = ['green' if k['direction'] == 'human' else 'red' for k in important_kmers]
    
    x = np.arange(len(kmers))
    width = 0.35
    
    # Frequency bars
    ax.bar(x - width/2, frequencies, width, label='Frequency (%)', color=colors, alpha=0.6)
    
    # Create a twin axis for sigma
    ax2 = ax.twinx()
    # Sigma bars
    ax2.bar(x + width/2, sigmas, width, label='σ from mean', 
            color=[c if s > 0 else 'gray' for c, s in zip(colors, sigmas)], alpha=0.3)
    
    ax.set_xticks(x)
    ax.set_xticklabels(kmers, rotation=45)
    ax.set_ylabel('Frequency (%)')
    ax2.set_ylabel('Standard Deviations (σ) from Mean')
    ax.set_title("K-mer Frequencies & Statistical Significance")

    lines1, labels1 = ax.get_legend_handles_labels()
    lines2, labels2 = ax2.get_legend_handles_labels()
    ax.legend(lines1 + lines2, labels1 + labels2, loc='best')
    
    plt.tight_layout()
    return fig

# --------------- Main Prediction Logic ---------------

def predict_fasta(
    file_obj, 
    k_size=4, 
    top_k=10,
    advanced_analysis=False
):
    """
    Main function to predict classes for each sequence in an uploaded FASTA.
    Returns:
      - Combined textual report for all sequences
      - A list of generated PIL Image plots
    """
    # 1. Read raw text from file or string
    if file_obj is None:
        return "Please upload a FASTA file", []
    
    try:
        if isinstance(file_obj, str):
            text = file_obj
        else:
            text = file_obj.decode('utf-8', errors='replace')
    except Exception as e:
        return f"Error reading file: {str(e)}", []
    
    # 2. Parse the FASTA
    sequences = parse_fasta(text)
    if not sequences:
        return "No valid FASTA sequences found!", []
    
    # 3. Load model & scaler
    try:
        device = 'cuda' if torch.cuda.is_available() else 'cpu'
        model = VirusClassifier(input_shape=(4 ** k_size)).to(device)
        state_dict = torch.load('model.pt', map_location=device)
        model.load_state_dict(state_dict)
        model.eval()
        
        scaler = joblib.load('scaler.pkl')
    except Exception as e:
        return f"Error loading model/scaler: {str(e)}", []
    
    # 4. Prepare k-mer dictionary for reference
    all_kmers = [''.join(p) for p in product("ACGT", repeat=k_size)]
    kmer_dict = {km: i for i, km in enumerate(all_kmers)}

    # 5. Iterate over sequences and build output
    final_text_report = []
    plots = []
    
    for idx, (header, seq) in enumerate(sequences, start=1):
        seq_stats = compute_sequence_stats(seq)
        
        # Convert sequence -> raw freq -> scaled freq
        raw_kmer_freq = sequence_to_kmer_vector(seq, k=k_size)
        scaled_kmer_freq = scaler.transform(raw_kmer_freq.reshape(1, -1))
        X_tensor = torch.FloatTensor(scaled_kmer_freq).to(device)
        
        # Predict
        with torch.no_grad():
            output = model(X_tensor)
            probs = torch.softmax(output, dim=1)
        
        # Determine class
        pred_class = torch.argmax(probs, dim=1).item()
        pred_label = 'human' if pred_class == 1 else 'non-human'
        human_prob = float(probs[0][1])
        non_human_prob = float(probs[0][0])
        confidence = float(torch.max(probs[0]).item())
        
        # Compute gradient-based importance
        importance, target_prob = model.get_gradient_importance(X_tensor, class_index=1)
        importance = importance[0].cpu().numpy()  # shape: (num_features,)

        # Identify top-k features (by absolute gradient)
        abs_importance = np.abs(importance)
        sorted_indices = np.argsort(abs_importance)[::-1]
        top_indices = sorted_indices[:top_k]
        
        # Build a list of top k-mers
        top_kmers_info = []
        for i in top_indices:
            kmer_name = all_kmers[i]
            imp_val = float(importance[i])
            direction = 'human' if imp_val > 0 else 'non-human'
            freq_perc = float(raw_kmer_freq[i] * 100.0)  # in percent
            sigma = float(scaled_kmer_freq[0][i])  # This is the scaled value (stdev from mean if the scaler is StandardScaler)
            
            top_kmers_info.append({
                'kmer': kmer_name,
                'impact': abs(imp_val),
                'direction': direction,
                'occurrence': freq_perc,
                'sigma': sigma
            })
        
        # Text summary for this sequence
        seq_report = []
        seq_report.append(f"=== Sequence {idx} ===")
        seq_report.append(f"Header: {header}")
        seq_report.append(f"Length: {seq_stats['length']}")
        seq_report.append(f"GC Content: {seq_stats['gc_content']:.2f}%")
        seq_report.append(f"A: {seq_stats['counts']['A']}, C: {seq_stats['counts']['C']}, G: {seq_stats['counts']['G']}, T: {seq_stats['counts']['T']}")
        seq_report.append(f"Prediction: {pred_label} (Confidence: {confidence:.4f})")
        seq_report.append(f"  Human Probability: {human_prob:.4f}")
        seq_report.append(f"  Non-human Probability: {non_human_prob:.4f}")
        seq_report.append(f"\nTop-{top_k} Influential k-mers (by gradient magnitude):")
        for tkm in top_kmers_info:
            seq_report.append(
                f"  {tkm['kmer']}: pushes towards {tkm['direction']} "
                f"(impact={tkm['impact']:.4f}), occurrence={tkm['occurrence']:.2f}%, "
                f"sigma={tkm['sigma']:.2f}"
            )
        
        final_text_report.append("\n".join(seq_report))
        
        # 6. Generate Plots (for each sequence)
        if advanced_analysis:
            # 6A. SHAP-like bar chart
            fig_shap = plot_shap_like_bars(
                kmers=all_kmers, 
                importance_values=importance, 
                top_k=top_k
            )
            buf_shap = io.BytesIO()
            fig_shap.savefig(buf_shap, format='png', bbox_inches='tight', dpi=150)
            buf_shap.seek(0)
            plots.append(Image.open(buf_shap))
            plt.close(fig_shap)

            # 6B. k-mer distribution histogram
            fig_kmer_dist = plot_kmer_distribution(raw_kmer_freq, all_kmers)
            buf_dist = io.BytesIO()
            fig_kmer_dist.savefig(buf_dist, format='png', bbox_inches='tight', dpi=150)
            buf_dist.seek(0)
            plots.append(Image.open(buf_dist))
            plt.close(fig_kmer_dist)
        
        # 6C. Original step visualization for top k k-mers
        # Sort by actual 'impact' to preserve that step logic
        # (largest absolute impact first)
        top_kmers_info_step = sorted(top_kmers_info, key=lambda x: x['impact'], reverse=True)
        fig_step = create_step_visualization(top_kmers_info_step, human_prob)
        buf_step = io.BytesIO()
        fig_step.savefig(buf_step, format='png', bbox_inches='tight', dpi=150)
        buf_step.seek(0)
        plots.append(Image.open(buf_step))
        plt.close(fig_step)
        
        # 6D. Frequency vs. sigma bar chart
        fig_freq_sigma = plot_kmer_freq_and_sigma(top_kmers_info_step)
        buf_freq_sigma = io.BytesIO()
        fig_freq_sigma.savefig(buf_freq_sigma, format='png', bbox_inches='tight', dpi=150)
        buf_freq_sigma.seek(0)
        plots.append(Image.open(buf_freq_sigma))
        plt.close(fig_freq_sigma)
    
    # Combine all text results
    combined_text = "\n\n".join(final_text_report)
    return combined_text, plots

# --------------- Gradio Interface ---------------

def run_prediction(
    file_obj, 
    k_size, 
    top_k,
    advanced_analysis
):
    """
    Wrapper for Gradio to handle the outputs in (text, List[Image]) form.
    """
    text_output, pil_images = predict_fasta(
        file_obj=file_obj, 
        k_size=k_size,
        top_k=top_k,
        advanced_analysis=advanced_analysis
    )


    return text_output, pil_images


with gr.Blocks() as demo:
    gr.Markdown("# Virus Host Classifier (Improved!)")
    gr.Markdown(
        "Upload a FASTA file and configure k-mer size, number of top features, "
        "and whether to run advanced analysis (plots of SHAP-like bars & k-mer distribution)."
    )
    
    with gr.Row():
        with gr.Column():
            fasta_file = gr.File(label="Upload FASTA file", type="binary")
            kmer_slider = gr.Slider(minimum=2, maximum=6, value=4, step=1, label="K-mer Size")
            topk_slider = gr.Slider(minimum=5, maximum=20, value=10, step=1, label="Top-k Features")
            advanced_check = gr.Checkbox(value=False, label="Advanced Analysis")
            predict_button = gr.Button("Predict")
            
        with gr.Column():
            results_text = gr.Textbox(
                label="Results", 
                lines=20, 
                placeholder="Prediction results will appear here..."
            )
    
    # We can display multiple images in a Gallery or as separate outputs.
    plots_gallery = gr.Gallery(label="Analysis Plots").style(grid=[2], height="auto")
    
    predict_button.click(
        fn=run_prediction,
        inputs=[fasta_file, kmer_slider, topk_slider, advanced_check],
        outputs=[results_text, plots_gallery]
    )

if __name__ == "__main__":
    demo.launch(share=True)