app.py

#!/usr/bin/env python3
"""
Tranception Design App - Hugging Face Spaces Version (Zero GPU Fixed)
"""
import os
import sys

# Set up caching to avoid re-downloading models
os.environ['HF_HOME'] = '/tmp/huggingface'
os.environ['TRANSFORMERS_CACHE'] = '/tmp/huggingface/transformers'
os.environ['HF_DATASETS_CACHE'] = '/tmp/huggingface/datasets'
# Ensure proper Hugging Face endpoint
os.environ['HF_ENDPOINT'] = 'https://huggingface.co'
# Disable offline mode to allow downloads
os.environ['TRANSFORMERS_OFFLINE'] = '0'

# Patch for transformers 4.17.0 URL issue in HF Spaces
import urllib.parse

def patch_transformers_url():
    """Fix URL scheme issue in transformers 4.17.0"""
    try:
        import transformers.file_utils
        original_get_from_cache = transformers.file_utils.get_from_cache
        
        def patched_get_from_cache(url, *args, **kwargs):
            # Fix URLs that start with /api/ by prepending https://huggingface.co
            if isinstance(url, str) and url.startswith('/api/'):
                url = 'https://huggingface.co' + url
            return original_get_from_cache(url, *args, **kwargs)
        
        transformers.file_utils.get_from_cache = patched_get_from_cache
        print("Applied URL patch for transformers")
    except Exception as e:
        print(f"Warning: Could not patch transformers URL handling: {e}")

import torch
import transformers
patch_transformers_url()
from transformers import PreTrainedTokenizerFast
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
import gradio as gr
from huggingface_hub import hf_hub_download
import shutil
import uuid
import gc
import time
import datetime
import threading

# Simplified Zero GPU handling
try:
    import spaces
    SPACES_AVAILABLE = True
    print("Zero GPU support detected")
except ImportError:
    SPACES_AVAILABLE = False
    print("Running without Zero GPU support")
except Exception as e:
    # Catch any other initialization errors
    SPACES_AVAILABLE = False
    print(f"Zero GPU initialization warning: {e}")
    print("Running without Zero GPU support")

# Runtime mode tracking
RUNTIME_MODE = "GPU" if SPACES_AVAILABLE else "CPU"

# Keep-alive state
last_activity = datetime.datetime.now()
activity_lock = threading.Lock()

def update_activity():
    """Update last activity timestamp"""
    global last_activity
    with activity_lock:
        last_activity = datetime.datetime.now()

# Add current directory to path
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))

# Check if we need to download and extract the tranception module
if not os.path.exists("tranception"):
    print("Downloading Tranception repository...")
    try:
        # Clone the repository structure
        result = os.system("git clone https://github.com/OATML-Markslab/Tranception.git temp_tranception")
        if result != 0:
            raise Exception("Failed to clone Tranception repository")
        # Move the tranception module to current directory
        shutil.move("temp_tranception/tranception", "tranception")
        # Clean up
        shutil.rmtree("temp_tranception")
    except Exception as e:
        print(f"Error setting up Tranception: {e}")
        if os.path.exists("temp_tranception"):
            shutil.rmtree("temp_tranception")
        raise

import tranception
from tranception import config, model_pytorch

# Model loading configuration
MODEL_CACHE = {}

def get_model_path(model_name):
    """Get model path - always use HF Hub for Zero GPU spaces"""
    # In HF Spaces, models are cached automatically by the transformers library
    # Always return the HF Hub path to leverage this caching
    return f"PascalNotin/{model_name}"

def load_model_cached(model_type):
    """Load model with caching to avoid re-downloading"""
    global MODEL_CACHE
    
    # Check if model is already in cache
    if model_type in MODEL_CACHE:
        print(f"Using cached {model_type} model")
        return MODEL_CACHE[model_type]
    
    print(f"Loading {model_type} model...")
    model_name = f"Tranception_{model_type}"
    model_path = get_model_path(model_name)
    
    try:
        # Create cache directory if it doesn't exist
        cache_dir = "/tmp/huggingface/transformers"
        os.makedirs(cache_dir, exist_ok=True)
        
        # Try loading with minimal parameters first
        model = tranception.model_pytorch.TranceptionLMHeadModel.from_pretrained(
            model_path,
            cache_dir=cache_dir
        )
        MODEL_CACHE[model_type] = model
        print(f"{model_type} model loaded and cached")
        return model
    except Exception as e:
        print(f"Error loading {model_type} model: {e}")
        print(f"Attempting alternative loading method...")
        
        # Try alternative loading approach with full URL
        try:
            # Use full URL to bypass any path resolution issues
            full_url = f"https://huggingface.co/PascalNotin/Tranception_{model_type}"
            model = tranception.model_pytorch.TranceptionLMHeadModel.from_pretrained(
                full_url,
                cache_dir=cache_dir
            )
            MODEL_CACHE[model_type] = model
            print(f"{model_type} model loaded successfully with full URL")
            return model
        except Exception as e2:
            print(f"Alternative loading also failed: {e2}")
            
            # Final attempt: manually download config first
            try:
                import json
                import requests
                
                # Download config.json manually
                config_url = f"https://huggingface.co/PascalNotin/Tranception_{model_type}/raw/main/config.json"
                print(f"Manually downloading config from: {config_url}")
                
                response = requests.get(config_url)
                if response.status_code == 200:
                    # Save config locally
                    local_model_dir = f"/tmp/Tranception_{model_type}"
                    os.makedirs(local_model_dir, exist_ok=True)
                    
                    with open(f"{local_model_dir}/config.json", "w") as f:
                        json.dump(response.json(), f)
                    
                    # Now try loading from the HF model ID again
                    model = tranception.model_pytorch.TranceptionLMHeadModel.from_pretrained(
                        f"PascalNotin/Tranception_{model_type}",
                        cache_dir=cache_dir,
                        local_files_only=False
                    )
                    MODEL_CACHE[model_type] = model
                    print(f"{model_type} model loaded successfully after manual config download")
                    return model
                else:
                    print(f"Failed to download config: {response.status_code}")
            except Exception as e3:
                print(f"Manual download also failed: {e3}")
            
            # Fallback to Medium if requested model fails
            if model_type != "Medium":
                print("Falling back to Medium model...")
                return load_model_cached("Medium")
            raise

AA_vocab = "ACDEFGHIKLMNPQRSTVWY"
tokenizer = PreTrainedTokenizerFast(tokenizer_file="./tranception/utils/tokenizers/Basic_tokenizer",
                                                unk_token="[UNK]",
                                                sep_token="[SEP]",
                                                pad_token="[PAD]",
                                                cls_token="[CLS]",
                                                mask_token="[MASK]"
                                            )

def create_all_single_mutants(sequence,AA_vocab=AA_vocab,mutation_range_start=None,mutation_range_end=None):
  all_single_mutants={}
  sequence_list=list(sequence)
  if mutation_range_start is None: mutation_range_start=1
  if mutation_range_end is None: mutation_range_end=len(sequence)
  for position,current_AA in enumerate(sequence[mutation_range_start-1:mutation_range_end]):
    for mutated_AA in AA_vocab:
      if current_AA!=mutated_AA:
        mutated_sequence = sequence_list.copy()
        mutated_sequence[mutation_range_start + position - 1] = mutated_AA
        all_single_mutants[current_AA+str(mutation_range_start+position)+mutated_AA]="".join(mutated_sequence)
  all_single_mutants = pd.DataFrame.from_dict(all_single_mutants,columns=['mutated_sequence'],orient='index')
  all_single_mutants.reset_index(inplace=True)
  all_single_mutants.columns = ['mutant','mutated_sequence']
  return all_single_mutants

def create_scoring_matrix_visual(scores,sequence,image_index=0,mutation_range_start=None,mutation_range_end=None,AA_vocab=AA_vocab,annotate=True,fontsize=20,unique_id=None):
  if unique_id is None:
    unique_id = str(uuid.uuid4())
    
  filtered_scores=scores.copy()
  filtered_scores=filtered_scores[filtered_scores.position.isin(range(mutation_range_start,mutation_range_end+1))]
  piv=filtered_scores.pivot(index='position',columns='target_AA',values='avg_score').round(4)
  
  # Calculate mutation range length
  mutation_range_len = mutation_range_end - mutation_range_start + 1
  
  # Save CSV file
  csv_path = 'fitness_scoring_substitution_matrix_{}_{}.csv'.format(unique_id, image_index)
  
  # Create a more detailed CSV with mutation info
  csv_data = []
  for position in range(mutation_range_start,mutation_range_end+1):
    for target_AA in list(AA_vocab):
      mutant = sequence[position-1]+str(position)+target_AA
      if mutant in set(filtered_scores.mutant):
        score_value = filtered_scores.loc[filtered_scores.mutant==mutant,'avg_score']
        if isinstance(score_value, pd.Series):
          score = float(score_value.iloc[0])
        else:
          score = float(score_value)
      else:
        score = 0.0
      
      csv_data.append({
        'position': position,
        'original_AA': sequence[position-1],
        'target_AA': target_AA,
        'mutation': mutant,
        'fitness_score': score
      })
  
  csv_df = pd.DataFrame(csv_data)
  csv_df.to_csv(csv_path, index=False)
  
  # Continue with visualization
  # Use large fixed width for clarity, height scales with positions (as in reference)
  fig, ax = plt.subplots(figsize=(50, mutation_range_len))
  scores_dict = {}
  valid_mutant_set=set(filtered_scores.mutant)  
  ax.tick_params(bottom=True, top=True, left=True, right=True)
  ax.tick_params(labelbottom=True, labeltop=True, labelleft=True, labelright=True)
  if annotate:
    for position in range(mutation_range_start,mutation_range_end+1):
      for target_AA in list(AA_vocab):
        mutant = sequence[position-1]+str(position)+target_AA
        if mutant in valid_mutant_set:
          score_value = filtered_scores.loc[filtered_scores.mutant==mutant,'avg_score']
          if isinstance(score_value, pd.Series):
            scores_dict[mutant] = float(score_value.iloc[0])
          else:
            scores_dict[mutant] = float(score_value)
        else:
          scores_dict[mutant]=0.0
    # Format labels as in reference - always show mutation and score with 4 decimal places
    labels = (np.asarray(["{} \n {:.4f}".format(symb,value) for symb, value in scores_dict.items() ])).reshape(mutation_range_len,len(AA_vocab))
    
    heat = sns.heatmap(piv,annot=labels,fmt="",cmap='RdYlGn',linewidths=0.30,ax=ax,vmin=np.percentile(scores.avg_score,2),vmax=np.percentile(scores.avg_score,98),\
                cbar_kws={'label': 'Log likelihood ratio (mutant / starting sequence)'},annot_kws={"size": fontsize})
  else:
    heat = sns.heatmap(piv,cmap='RdYlGn',linewidths=0.30,ax=ax,vmin=np.percentile(scores.avg_score,2),vmax=np.percentile(scores.avg_score,98),\
                cbar_kws={'label': 'Log likelihood ratio (mutant / starting sequence)'},annot_kws={"size": fontsize})
  # Use label sizes from reference
  heat.figure.axes[-1].yaxis.label.set_size(fontsize=int(fontsize*1.5))
  heat.set_title("Higher predicted scores (green) imply higher protein fitness",fontsize=fontsize*2, pad=40)
  heat.set_ylabel("Sequence position", fontsize = fontsize*2)
  heat.set_xlabel("Amino Acid mutation", fontsize = fontsize*2)
  
  # Set y-axis labels (positions)
  yticklabels = [str(pos)+' ('+sequence[pos-1]+')' for pos in range(mutation_range_start,mutation_range_end+1)]
  heat.set_yticklabels(yticklabels, fontsize=fontsize, rotation=0)
  
  # Set x-axis labels (amino acids) - ensuring correct number
  heat.set_xticklabels(list(AA_vocab), fontsize=fontsize)
  try:
    plt.tight_layout()
    image_path = 'fitness_scoring_substitution_matrix_{}_{}.png'.format(unique_id, image_index)
    plt.savefig(image_path, dpi=100)
    return image_path, csv_path
  finally:
    plt.close('all')  # Ensure all figures are closed
    plt.clf()  # Clear the current figure
    plt.cla()  # Clear the current axes

def suggest_mutations(scores):
  intro_message = "The following mutations may be sensible options to improve fitness: \n\n"
  #Best mutants
  top_mutants=list(scores.sort_values(by=['avg_score'],ascending=False).head(5).mutant)
  top_mutants_fitness=list(scores.sort_values(by=['avg_score'],ascending=False).head(5).avg_score)
  top_mutants_recos = [top_mutant+" ("+str(round(top_mutant_fitness,4))+")" for (top_mutant,top_mutant_fitness) in zip(top_mutants,top_mutants_fitness)]
  mutant_recos = "The single mutants with highest predicted fitness are (positive scores indicate fitness increase Vs starting sequence, negative scores indicate fitness decrease):\n {} \n\n".format(", ".join(top_mutants_recos))
  #Best positions
  positive_scores = scores[scores.avg_score > 0]
  if len(positive_scores) > 0:
    # Only select numeric columns for groupby mean
    positive_scores_position_avg = positive_scores.groupby(['position'])['avg_score'].mean().reset_index()
    top_positions=list(positive_scores_position_avg.sort_values(by=['avg_score'],ascending=False).head(5)['position'].astype(str))
    position_recos = "The positions with the highest average fitness increase are (only positions with at least one fitness increase are considered):\n {}".format(", ".join(top_positions))
  else:
    position_recos = "No positions with positive fitness effects found."
  return intro_message+mutant_recos+position_recos

def check_valid_mutant(sequence,mutant,AA_vocab=AA_vocab):
  valid = True
  try:
    from_AA, position, to_AA = mutant[0], int(mutant[1:-1]), mutant[-1]
  except:
    valid = False
  if valid and position > 0 and position <= len(sequence):
    if sequence[position-1]!=from_AA: valid=False
  else:
    valid = False
  if to_AA not in AA_vocab: valid=False
  return valid

def cleanup_old_files(max_age_minutes=30):
    """Clean up old inference files"""
    import glob
    current_time = time.time()
    patterns = ["fitness_scoring_substitution_matrix_*.png", 
                "fitness_scoring_substitution_matrix_*.csv",
                "all_mutations_fitness_scores_*.csv"]
    
    cleaned_count = 0
    for pattern in patterns:
        for file_path in glob.glob(pattern):
            try:
                file_age = current_time - os.path.getmtime(file_path)
                if file_age > max_age_minutes * 60:
                    os.remove(file_path)
                    cleaned_count += 1
            except Exception as e:
                # Log error but continue cleaning other files
                print(f"Warning: Could not remove {file_path}: {e}")
    
    if cleaned_count > 0:
        print(f"Cleaned up {cleaned_count} old files")

def get_mutated_protein(sequence,mutant):
  if not check_valid_mutant(sequence,mutant):
    return "The mutant is not valid"
  mutated_sequence = list(sequence)
  mutated_sequence[int(mutant[1:-1])-1]=mutant[-1]
  return ''.join(mutated_sequence)

def score_and_create_matrix_all_singles_impl(sequence,mutation_range_start=None,mutation_range_end=None,model_type="Large",scoring_mirror=False,batch_size_inference=20,max_number_positions_per_heatmap=50,num_workers=0,AA_vocab=AA_vocab):
  # Update activity
  update_activity()
  
  # Clean up old files periodically
  cleanup_old_files()
  
  # Generate unique ID for this request
  unique_id = str(uuid.uuid4())
  
  if mutation_range_start is None: mutation_range_start=1
  if mutation_range_end is None: mutation_range_end=len(sequence)
  
  # Clean sequence
  sequence = sequence.strip().upper()
  
  # Validate
  assert len(sequence) > 0, "no sequence entered"
  assert mutation_range_start <= mutation_range_end, "mutation range is invalid"
  assert mutation_range_end <= len(sequence), f"End position ({mutation_range_end}) exceeds sequence length ({len(sequence)})"
  
  # Load model with caching
  model = load_model_cached(model_type)
  
  # Move model to appropriate device INSIDE the GPU decorated function
  # This is crucial for Zero GPU - the model must be moved to GPU inside the decorated function
  
  # Device selection - Zero GPU will provide CUDA when decorated with @spaces.GPU
  print(f"GPU Available: {torch.cuda.is_available()}")
  
  if torch.cuda.is_available():
    device = torch.device("cuda")
    model = model.to(device)
    gpu_name = torch.cuda.get_device_name(0)
    gpu_memory = torch.cuda.get_device_properties(0).total_memory / 1024**3
    print(f"Inference will take place on {gpu_name}")
    print(f"GPU Memory: {gpu_memory:.2f} GB")
    # Increase batch size for GPU inference
    batch_size_inference = min(batch_size_inference, 50)
  else:
    device = torch.device("cpu")
    model = model.to(device)
    print("Inference will take place on CPU")
    # Reduce batch size for CPU inference
    batch_size_inference = min(batch_size_inference, 10)
    
  try:
    model.eval()
    model.config.tokenizer = tokenizer
    
    all_single_mutants = create_all_single_mutants(sequence,AA_vocab,mutation_range_start,mutation_range_end)
    
    with torch.no_grad():
      scores = model.score_mutants(DMS_data=all_single_mutants, 
                                        target_seq=sequence, 
                                        scoring_mirror=scoring_mirror, 
                                        batch_size_inference=batch_size_inference,  
                                        num_workers=num_workers, 
                                        indel_mode=False
                                        )
    
    scores = pd.merge(scores,all_single_mutants,on="mutated_sequence",how="left")
    scores["position"]=scores["mutant"].map(lambda x: int(x[1:-1]))
    scores["target_AA"] = scores["mutant"].map(lambda x: x[-1])
    
    score_heatmaps = []
    csv_files = []
    mutation_range = mutation_range_end - mutation_range_start + 1
    number_heatmaps = int((mutation_range - 1) / max_number_positions_per_heatmap) + 1
    image_index = 0
    window_start = mutation_range_start
    window_end = min(mutation_range_end,mutation_range_start+max_number_positions_per_heatmap-1)
    
    for image_index in range(number_heatmaps):
      image_path, csv_path = create_scoring_matrix_visual(scores,sequence,image_index,window_start,window_end,AA_vocab,unique_id=unique_id)
      score_heatmaps.append(image_path)
      csv_files.append(csv_path)
      window_start += max_number_positions_per_heatmap
      window_end = min(mutation_range_end,window_start+max_number_positions_per_heatmap-1)
    
    # Also save a comprehensive CSV with all mutations
    comprehensive_csv_path = 'all_mutations_fitness_scores_{}.csv'.format(unique_id)
    scores_export = scores[['mutant', 'position', 'target_AA', 'avg_score', 'mutated_sequence']].copy()
    scores_export['original_AA'] = scores_export['mutant'].str[0]
    scores_export = scores_export.rename(columns={'avg_score': 'fitness_score'})
    scores_export = scores_export[['position', 'original_AA', 'target_AA', 'mutant', 'fitness_score', 'mutated_sequence']]
    scores_export.to_csv(comprehensive_csv_path, index=False)
    csv_files.append(comprehensive_csv_path)
    
    return score_heatmaps, suggest_mutations(scores), csv_files
    
  finally:
    # Clean up GPU memory but keep model in cache
    # Move model back to CPU to free GPU memory
    if 'model' in locals():
      model.cpu()
    if torch.cuda.is_available():
      torch.cuda.empty_cache()
    gc.collect()

# Apply Zero GPU decorator if available
if SPACES_AVAILABLE:
    try:
        score_and_create_matrix_all_singles = spaces.GPU(duration=300)(score_and_create_matrix_all_singles_impl)
    except Exception as e:
        print(f"Warning: Could not apply Zero GPU decorator: {e}")
        print("Falling back to CPU mode")
        score_and_create_matrix_all_singles = score_and_create_matrix_all_singles_impl
else:
    score_and_create_matrix_all_singles = score_and_create_matrix_all_singles_impl

def extract_sequence(protein_id, taxon, sequence):
  return sequence

def clear_inputs(protein_sequence_input,mutation_range_start,mutation_range_end):
  protein_sequence_input = ""
  mutation_range_start = None
  mutation_range_end = None
  return protein_sequence_input,mutation_range_start,mutation_range_end

# Create Gradio app
tranception_design = gr.Blocks()

with tranception_design:
    gr.Markdown("# In silico directed evolution for protein redesign with Tranception")
    gr.Markdown("## 🧬 BASIS-China iGEM Team 2025 - Protein Engineering Platform")
    gr.Markdown("### Welcome to BASIS-China's implementation of Tranception on Hugging Face Spaces!")
    gr.Markdown("We are the BASIS-China iGEM team, and we're excited to present our deployment of the Tranception model for protein fitness prediction. This tool enables in silico directed evolution to iteratively improve protein fitness through single amino acid substitutions. At each step, Tranception computes log likelihood ratios for all possible mutations compared to the starting sequence, generating fitness heatmaps and recommendations to guide protein engineering.")
    gr.Markdown("**Technical Details**: This deployment leverages Hugging Face's Zero GPU infrastructure, which dynamically allocates H200 GPU resources when available. This allows for efficient inference while managing computational resources effectively.")
    
    # Hidden keep-alive component
    with gr.Row(visible=False):
        keep_alive_component = gr.Number(value=0, visible=False)
        
        def keep_alive_update():
            update_activity()
            return time.time()
        
        # Update every 2 minutes to keep websocket alive
        keep_alive_timer = gr.Timer(value=120)
        keep_alive_timer.tick(keep_alive_update, outputs=[keep_alive_component])
    
    # Status indicator
    with gr.Row():
        with gr.Column(scale=1):
            def get_gpu_status():
                global RUNTIME_MODE
                with activity_lock:
                    time_since = (datetime.datetime.now() - last_activity).total_seconds()
                
                if RUNTIME_MODE == "GPU":
                    status = "🔥 Zero GPU"
                else:
                    status = "💻 CPU Mode (GPU initialization failed)"
                return f"{status} | Last activity: {int(time_since)}s ago"
            
            gpu_status = gr.Textbox(
                label="Compute Status", 
                value=get_gpu_status, 
                every=5,  # Update every 5 seconds
                interactive=False,
                elem_id="gpu_status"
            )
    
    with gr.Tabs():
        with gr.TabItem("Input"):
            with gr.Row():
                protein_sequence_input = gr.Textbox(lines=1, 
                                                label="Protein sequence",
                                                placeholder = "Input the sequence of amino acids representing the starting protein of interest or select one from the list of examples below. You may enter the full sequence or just a subdomain (providing full context typically leads to better results, but is slower at inference)"
                                                )
            
            with gr.Row():
                mutation_range_start = gr.Number(label="Start of mutation window (first position indexed at 1)", value=1, precision=0)
                mutation_range_end = gr.Number(label="End of mutation window (leave empty for full lenth)", value=10, precision=0)

        with gr.TabItem("Parameters"):
            with gr.Row():
                model_size_selection = gr.Radio(label="Tranception model size (larger models are more accurate but are slower at inference)", 
                                                choices=["Small","Medium","Large"], 
                                                value="Small")
            with gr.Row():
                scoring_mirror = gr.Checkbox(label="Score protein from both directions (leads to more robust fitness predictions, but doubles inference time)")
            with gr.Row():
                batch_size_inference = gr.Number(label="Model batch size at inference time (reduce for CPU)",value = 10, precision=0)
            with gr.Row():
                gr.Markdown("Note: the current version does not leverage retrieval of homologs at inference time to increase fitness prediction performance.")
                
    with gr.Row():
        clear_button = gr.Button(value="Clear",variant="secondary")
        run_button = gr.Button(value="Predict fitness",variant="primary")
        
    protein_ID = gr.Textbox(label="Uniprot ID", visible=False)
    taxon = gr.Textbox(label="Taxon", visible=False)
    
    examples = gr.Examples(
        inputs=[protein_ID, taxon, protein_sequence_input],
        outputs=[protein_sequence_input],
        fn=extract_sequence,
        examples=[
            ['ADRB2_HUMAN'  ,'Human',           'MGQPGNGSAFLLAPNGSHAPDHDVTQERDEVWVVGMGIVMSLIVLAIVFGNVLVITAIAKFERLQTVTNYFITSLACADLVMGLAVVPFGAAHILMKMWTFGNFWCEFWTSIDVLCVTASIETLCVIAVDRYFAITSPFKYQSLLTKNKARVIILMVWIVSGLTSFLPIQMHWYRATHQEAINCYANETCCDFFTNQAYAIASSIVSFYVPLVIMVFVYSRVFQEAKRQLQKIDKSEGRFHVQNLSQVEQDGRTGHGLRRSSKFCLKEHKALKTLGIIMGTFTLCWLPFFIVNIVHVIQDNLIRKEVYILLNWIGYVNSGFNPLIYCRSPDFRIAFQELLCLRRSSLKAYGNGYSSNGNTGEQSGYHVEQEKENKLLCEDLPGTEDFVGHQGTVPSDNIDSQGRNCSTNDSLL'],
            ['IF1_ECOLI'    ,'Prokaryote',      'MAKEDNIEMQGTVLETLPNTMFRVELENGHVVTAHISGKMRKNYIRILTGDKVTVELTPYDLSKGRIVFRSR'],
            ['P53_HUMAN'    ,'Human',           'MEEPQSDPSVEPPLSQETFSDLWKLLPENNVLSPLPSQAMDDLMLSPDDIEQWFTEDPGPDEAPRMPEAAPRVAPAPAAPTPAAPAPAPSWPLSSSVPSQKTYQGSYGFRLGFLHSGTAKSVTCTYSPALNKMFCQLAKTCPVQLWVDSTPPPGTRVRAMAIYKQSQHMTEVVRRCPHHERCSDSDGLAPPQHLIRVEGNLRVEYLDDRNTFRHSVVVPYEPPEVGSDCTTIHYNYMCNSSCMGGMNRRPILTIITLEDSSGNLLGRNSFEVRVCACPGRDRRTEEENLRKKGEPHHELPPGSTKRALPNNTSSSPQPKKKPLDGEYFTLQIRGRERFEMFRELNEALELKDAQAGKEPGGSRAHSSHLKSKKGQSTSRHKKLMFKTEGPDSD'],
            ['BLAT_ECOLX'	  ,'Prokaryote',      'MSIQHFRVALIPFFAAFCLPVFAHPETLVKVKDAEDQLGARVGYIELDLNSGKILESFRPEERFPMMSTFKVLLCGAVLSRVDAGQEQLGRRIHYSQNDLVEYSPVTEKHLTDGMTVRELCSAAITMSDNTAANLLLTTIGGPKELTAFLHNMGDHVTRLDRWEPELNEAIPNDERDTTMPAAMATTLRKLLTGELLTLASRQQLIDWMEADKVAGPLLRSALPAGWFIADKSGAGERGSRGIIAALGPDGKPSRIVVIYTTGSQATMDERNRQIAEIGASLIKHW'],
            ['BRCA1_HUMAN'	,'Human',           'MDLSALRVEEVQNVINAMQKILECPICLELIKEPVSTKCDHIFCKFCMLKLLNQKKGPSQCPLCKNDITKRSLQESTRFSQLVEELLKIICAFQLDTGLEYANSYNFAKKENNSPEHLKDEVSIIQSMGYRNRAKRLLQSEPENPSLQETSLSVQLSNLGTVRTLRTKQRIQPQKTSVYIELGSDSSEDTVNKATYCSVGDQELLQITPQGTRDEISLDSAKKAACEFSETDVTNTEHHQPSNNDLNTTEKRAAERHPEKYQGSSVSNLHVEPCGTNTHASSLQHENSSLLLTKDRMNVEKAEFCNKSKQPGLARSQHNRWAGSKETCNDRRTPSTEKKVDLNADPLCERKEWNKQKLPCSENPRDTEDVPWITLNSSIQKVNEWFSRSDELLGSDDSHDGESESNAKVADVLDVLNEVDEYSGSSEKIDLLASDPHEALICKSERVHSKSVESNIEDKIFGKTYRKKASLPNLSHVTENLIIGAFVTEPQIIQERPLTNKLKRKRRPTSGLHPEDFIKKADLAVQKTPEMINQGTNQTEQNGQVMNITNSGHENKTKGDSIQNEKNPNPIESLEKESAFKTKAEPISSSISNMELELNIHNSKAPKKNRLRRKSSTRHIHALELVVSRNLSPPNCTELQIDSCSSSEEIKKKKYNQMPVRHSRNLQLMEGKEPATGAKKSNKPNEQTSKRHDSDTFPELKLTNAPGSFTKCSNTSELKEFVNPSLPREEKEEKLETVKVSNNAEDPKDLMLSGERVLQTERSVESSSISLVPGTDYGTQESISLLEVSTLGKAKTEPNKCVSQCAAFENPKGLIHGCSKDNRNDTEGFKYPLGHEVNHSRETSIEMEESELDAQYLQNTFKVSKRQSFAPFSNPGNAEEECATFSAHSGSLKKQSPKVTFECEQKEENQGKNESNIKPVQTVNITAGFPVVGQKDKPVDNAKCSIKGGSRFCLSSQFRGNETGLITPNKHGLLQNPYRIPPLFPIKSFVKTKCKKNLLEENFEEHSMSPEREMGNENIPSTVSTISRNNIRENVFKEASSSNINEVGSSTNEVGSSINEIGSSDENIQAELGRNRGPKLNAMLRLGVLQPEVYKQSLPGSNCKHPEIKKQEYEEVVQTVNTDFSPYLISDNLEQPMGSSHASQVCSETPDDLLDDGEIKEDTSFAENDIKESSAVFSKSVQKGELSRSPSPFTHTHLAQGYRRGAKKLESSEENLSSEDEELPCFQHLLFGKVNNIPSQSTRHSTVATECLSKNTEENLLSLKNSLNDCSNQVILAKASQEHHLSEETKCSASLFSSQCSELEDLTANTNTQDPFLIGSSKQMRHQSESQGVGLSDKELVSDDEERGTGLEENNQEEQSMDSNLGEAASGCESETSVSEDCSGLSSQSDILTTQQRDTMQHNLIKLQQEMAELEAVLEQHGSQPSNSYPSIISDSSALEDLRNPEQSTSEKAVLTSQKSSEYPISQNPEGLSADKFEVSADSSTSKNKEPGVERSSPSKCPSLDDRWYMHSCSGSLQNRNYPSQEELIKVVDVEEQQLEESGPHDLTETSYLPRQDLEGTPYLESGISLFSDDPESDPSEDRAPESARVGNIPSSTSALKVPQLKVAESAQSPAAAHTTDTAGYNAMEESVSREKPELTASTERVNKRMSMVVSGLTPEEFMLVYKFARKHHITLTNLITEETTHVVMKTDAEFVCERTLKYFLGIAGGKWVVSYFWVTQSIKERKMLNEHDFEVRGDVVNGRNHQGPKRARESQDRKIFRGLEICCYGPFTNMPTDQLEWMVQLCGASVVKELSSFTLGTGVHPIVVVQPDAWTEDNGFHAIGQMCEAPVVTREWVLDSVALYQCQELDTYLIPQIPHSHY'],
            ['CALM1_HUMAN'	,'Human',           'MADQLTEEQIAEFKEAFSLFDKDGDGTITTKELGTVMRSLGQNPTEAELQDMINEVDADGNGTIDFPEFLTMMARKMKDTDSEEEIREAFRVFDKDGNGYISAAELRHVMTNLGEKLTDEEVDEMIREADIDGDGQVNYEEFVQMMTAK'],
            ['CCDB_ECOLI'	  ,'Prokaryote',	    'MQFKVYTYKRESRYRLFVDVQSDIIDTPGRRMVIPLASARLLSDKVSRELYPVVHIGDESWRMMTTDMASVPVSVIGEEVADLSHRENDIKNAINLMFWGI'],
            ['GFP_AEQVI'	  ,'Other eukaryote', 'MSKGEELFTGVVPILVELDGDVNGHKFSVSGEGEGDATYGKLTLKFICTTGKLPVPWPTLVTTLSYGVQCFSRYPDHMKQHDFFKSAMPEGYVQERTIFFKDDGNYKTRAEVKFEGDTLVNRIELKGIDFKEDGNILGHKLEYNYNSHNVYIMADKQKNGIKVNFKIRHNIEDGSVQLADHYQQNTPIGDGPVLLPDNHYLSTQSALSKDPNEKRDHMVLLEFVTAAGITHGMDELYK'],
            ['GRB2_HUMAN'	  ,'Human',           'MEAIAKYDFKATADDELSFKRGDILKVLNEECDQNWYKAELNGKDGFIPKNYIEMKPHPWFFGKIPRAKAEEMLSKQRHDGAFLIRESESAPGDFSLSVKFGNDVQHFKVLRDGAGKYFLWVVKFNSLNELVDYHRSTSVSRNQQIFLRDIEQVPQQPTYVQALFDFDPQEDGELGFRRGDFIHVMDNSDPNWWKGACHGQTGMFPRNYVTPVNRNV'],
        ],
    )
    
    gr.Markdown("<br>")
    gr.Markdown("# Fitness predictions for all single amino acid substitutions in mutation range")
    gr.Markdown("Inference may take a few seconds for short proteins & mutation ranges to several minutes for longer ones")
    output_image = gr.Gallery(label="Fitness predictions for all single amino acid substitutions in mutation range") #Using Gallery to break down large scoring matrices into smaller images
    
    output_recommendations = gr.Textbox(label="Mutation recommendations")
    
    with gr.Row():
        gr.Markdown("## Download CSV Files")
    output_csv_files = gr.File(label="Download CSV files with fitness scores", file_count="multiple", interactive=False)
    
    clear_button.click(
        inputs = [protein_sequence_input,mutation_range_start,mutation_range_end],
        outputs = [protein_sequence_input,mutation_range_start,mutation_range_end],
        fn=clear_inputs
    )
    run_button.click(
        fn=score_and_create_matrix_all_singles,
        inputs=[protein_sequence_input,mutation_range_start,mutation_range_end,model_size_selection,scoring_mirror,batch_size_inference],
        outputs=[output_image,output_recommendations,output_csv_files],
    )
    
    gr.Markdown("# Mutate the starting protein sequence")
    with gr.Row():
        mutation_triplet = gr.Textbox(lines=1,label="Selected mutation", placeholder = "Input the mutation triplet for the selected mutation (eg., M1A)")
    mutate_button = gr.Button(value="Apply mutation to starting protein", variant="primary")
    mutated_protein_sequence = gr.Textbox(lines=1,label="Mutated protein sequence")
    mutate_button.click(
        fn = get_mutated_protein,
        inputs = [protein_sequence_input,mutation_triplet],
        outputs = mutated_protein_sequence
    )
    
    gr.Markdown("<p>You may now use the output mutated sequence above as the starting sequence for another round of in silico directed evolution.</p>")
    gr.Markdown("### About BASIS-China iGEM Team")
    gr.Markdown("We are a high school synthetic biology team participating in the International Genetically Engineered Machine (iGEM) competition. Our 2025 project focuses on protein engineering and computational biology applications. This Tranception deployment is part of our broader effort to make advanced protein design tools accessible to the synthetic biology community.")
    gr.Markdown("### About Tranception")
    gr.Markdown("<p><b>Tranception: Protein Fitness Prediction with Autoregressive Transformers and Inference-time Retrieval</b><br>Pascal Notin, Mafalda Dias, Jonathan Frazer, Javier Marchena-Hurtado, Aidan N. Gomez, Debora S. Marks<sup>*</sup>, Yarin Gal<sup>*</sup><br><sup>* equal senior authorship</sup></p>")
    gr.Markdown("Links: <a href='https://proceedings.mlr.press/v162/notin22a.html' target='_blank'>Paper</a>  <a href='https://github.com/OATML-Markslab/Tranception' target='_blank'>Code</a>  <a href='https://sites.google.com/view/proteingym/substitutions' target='_blank'>ProteinGym</a>  <a href='https://igem.org/teams/5247' target='_blank'>BASIS-China iGEM Team</a>")

if __name__ == "__main__":
    # Don't preload models at startup - this can cause Zero GPU initialization issues
    # Models will be loaded and cached on first use
    print("Starting Tranception app...")
    print("Note: Models will be downloaded on first use")
    print("Zero GPU spaces may sleep after ~15 minutes of inactivity")
    
    # Try to launch with ZeroGPU support first
    launch_success = False
    max_retries = 3
    retry_count = 0
    
    while not launch_success and retry_count < max_retries:
        try:
            if retry_count > 0:
                print(f"Retry attempt {retry_count}/{max_retries}...")
                time.sleep(2)  # Wait before retry
            
            # Launch with queue for proper Zero GPU support
            tranception_design.queue(max_size=20).launch(
                server_name="0.0.0.0",
                server_port=7860,
                show_error=True,
                share=False
            )
            launch_success = True
        except RuntimeError as e:
            if "Error while initializing ZeroGPU" in str(e):
                retry_count += 1
                if retry_count >= max_retries:
                    print(f"ZeroGPU initialization failed after {max_retries} attempts")
                    print("Falling back to CPU mode for stability")
                    print("Note: The app will run slower in CPU mode")
                    # Update runtime mode
                    RUNTIME_MODE = "CPU"
                    # Launch without queue which doesn't trigger ZeroGPU initialization
                    tranception_design.launch(
                        server_name="0.0.0.0",
                        server_port=7860,
                        show_error=True,
                        share=False
                    )
                    launch_success = True
            else:
                # Re-raise unexpected errors
                raise