app.py

import numpy as np
import os
import re
import torch
import gradio as gr  # 必须导入 gradio

# --- 依赖导入 ---
# 请确保这些文件存在，否则会报错
from model import CAFN
from Feature_extraction_algorithms.PSTAAP import PSTAAP_feature, load_precomputed_fr_matrix
from Feature_extraction_algorithms.Physicochemical import PC_feature

# 设置设备
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# --- 1. 模型初始化 ---
model = None
try:
    FR_MATRIX_PATH = 'Fr_train.mat' 
    if not os.path.exists(FR_MATRIX_PATH):
        # 为了演示，这里改为警告而不是阻断，实际运行时请确保文件存在
        print(f"警告：找不到矩阵文件 {FR_MATRIX_PATH}")
    else:
        load_precomputed_fr_matrix(FR_MATRIX_PATH)
    
    # 假设这里还需要加载模型权重
    # model = CAFN().to(device)
    # model.load_state_dict(torch.load('model_weights.pth'))
    # model.eval()

except Exception as e:
    print(f"初始化过程中发生错误: {e}")

# --- 3. 特征提取函数 (修复了你代码中缺失的函数定义头) ---
def extract_features_from_seq(sequence_list):
    """
    这里是你缺失的特征提取函数逻辑。
    你需要把之前的 data, data2 生成逻辑放回来。
    """
    # 模拟数据用于演示，请替换为你真实的特征提取代码
    # 假设 PSTAAP 和 PC_feature 返回 numpy 数组
    # data = PC_feature(sequence_list) ...
    # data2 = PSTAAP_feature(sequence_list) ...
    
    # 这是一个占位符，保证代码不报错
    # 实际维度需要根据你的模型输入调整 (Batch, Channel, Length)
    x1_dummy = np.random.rand(len(sequence_list), 1, 49).astype(np.float32) 
    x2_dummy = np.random.rand(len(sequence_list), 1, 49).astype(np.float32)
    return x1_dummy, x2_dummy

# --- 4. 核心预测函数 ---
def predict_single_49mer(sequence_49mer):
    """
    对单个、长度为49的序列片段进行预测。
    """
    # 如果模型未加载，返回模拟数据（方便调试UI）
    if model is None:
        print("警告：模型未加载，返回随机结果用于演示。")
        labels = ["Lysine-Acetyllysine (K-Ac)", "Lysine-Crotonyllysine (K-Cr)", "Lysine-Methyllysine (K-Me)", "Lysine-Succinyllysine (K-Succ)"]
        return {label: float(np.random.rand()) for label in labels}

    sequence_list = [sequence_49mer]
    x1_np, x2_np = extract_features_from_seq(sequence_list)
    
    tensor_x1 = torch.tensor(x1_np).to(device)
    tensor_x2 = torch.tensor(x2_np).to(device)
    
    with torch.no_grad(): # 推理时不需要梯度
        outputs = model(tensor_x1, tensor_x2)
        probabilities = torch.sigmoid(outputs).squeeze().cpu().numpy()
    
    labels = ["Lysine-Acetyllysine (K-Ac)", "Lysine-Crotonyllysine (K-Cr)", "Lysine-Methyllysine (K-Me)", "Lysine-Succinyllysine (K-Succ)"]
    
    # 处理 batch size 为 1 的情况
    if probabilities.ndim == 0:
        probabilities = [probabilities]
        
    result = {label: float(prob) for label, prob in zip(labels, probabilities)}
    return result

# --- 5. FASTA格式解析与主处理流程 ---
def parse_fasta(fasta_string):
    sequence_lines = [line for line in fasta_string.splitlines() if not line.startswith('>')]
    return "".join(sequence_lines).replace(" ", "").replace("\n", "").upper()

def process_fasta_and_predict(fasta_input):
    if not fasta_input or not isinstance(fasta_input, str):
        raise gr.Error("Please enter a valid FASTA format sequence.")

    sequence = parse_fasta(fasta_input)
    
    if len(sequence) < 49:
        raise gr.Error(f"Sequence too short! Needs at least 49 AA. Current: {len(sequence)}.")

    predictions_map = {}
    k_indices = [m.start() for m in re.finditer('K', sequence)]
    
    for k_index in k_indices:
        start, end = k_index - 24, k_index + 25
        if start >= 0 and end <= len(sequence):
            fragment = sequence[start:end]
            prediction_result = predict_single_49mer(fragment)
            if prediction_result:
                predictions_map[k_index] = prediction_result

    if not predictions_map:
        return [(sequence, None)], {}, "No valid K sites (with enough context) found."

    highlight_data = []
    last_pos = 0
    sorted_predictable_indices = sorted(predictions_map.keys())

    for k_index in sorted_predictable_indices:
        highlight_data.append((sequence[last_pos:k_index], None))
        highlight_data.append(("K", str(k_index)))
        last_pos = k_index + 1
    
    highlight_data.append((sequence[last_pos:], None))
    
    initial_info = "Processing complete! Click on the highlighted 'K' site below."
    return highlight_data, predictions_map, initial_info

# --- 6. Gradio事件处理函数 ---
def show_results_for_site(evt: gr.SelectData, state_data):
    if evt.value:
        k_index_str = evt.value[1] # 获取索引
        if k_index_str is None: # 如果点击了非高亮部分
             return None, "Please click on a highlighted 'K'."
             
        k_index = int(k_index_str)
        result_dict = state_data.get(k_index)
        
        if result_dict:
            site_info = f"Prediction results for 'K' at position {k_index + 1}:"
            return result_dict, site_info
            
    return None, "Please click on the highlighted 'K' site."

# --- 7. 创建并启动 Gradio 界面 ---

# 修复点 1: 补全字符串
fasta_example = """>sp|P05141|ADT2_HUMAN ADP/ATP translocase 2 OS=Homo sapiens OX=9606 GN=SLC25A5 PE=1 SV=7
MTDAAVSFAKDFLAGGVAAAISKTAVAPIERVKLLLQVQHASKQITADKQYKGIIDCVVR
IPKEQGVLSFWRGNLANVIRYFPTQALNFAFKDKYKQIFLGGVDKRTQFWLYFAGNLASG"""

# 修复点 2: 使用 with gr.Blocks() 包裹界面代码，并修复缩进
with gr.Blocks(css=".predictable-k { color: red; font-weight: bold; }") as demo:
    gr.Markdown(
        """
        # DeepKMulti Model: Multi-label Classifier for Lysine Modifications
        **Supports FASTA format input. Click on highlighted 'K' sites to view predictions.**
        """
    )
    
    with gr.Row():
        with gr.Column(scale=2):
            fasta_input = gr.Textbox(
                lines=10, 
                label="Input FASTA Sequence",
                value=fasta_example, # 设置默认值方便测试
                placeholder="Paste FASTA sequence here..."
            )
            submit_btn = gr.Button("Submit Prediction", variant="primary")
            
        with gr.Column(scale=3):
            gr.Markdown("### Prediction Results")
            info_text = gr.Textbox(label="Status", interactive=False, value="Waiting for input...")
            predictions_state = gr.State({}) 
            results_output = gr.Label(num_top_classes=4, label="Probabilities")
            
    gr.Markdown("---")
    gr.Markdown("### Visualized Sequence (Click 'K' to view details)")
    
    highlighted_output = gr.HighlightedText(
        label="Sequence Analysis",
        combine_adjacent=False,
        show_legend=False,
        color_map={"K": "red"} 
    )
    
    # 绑定事件
    submit_btn.click(
        fn=process_fasta_and_predict,
        inputs=[fasta_input],
        outputs=[highlighted_output, predictions_state, info_text]
    )
    
    highlighted_output.select(
        fn=show_results_for_site,
        inputs=[predictions_state],
        outputs=[results_output, info_text]
    )

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