app.py

import os
import sys
import json
import torch
import numpy as np
from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel
from typing import List, Optional
from huggingface_hub import hf_hub_download
from rdkit import Chem

# Add workspace/src to path to import local modules
sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..', 'workspace', 'src')))
sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), 'workspace', 'src')))

from model import MultiTaskEvidentialGNN
from dataloader import smiles_to_graph
from explain import explain_molecule, render_molecular_explanation
from scaffold import generate_synthetic_analogs

app = FastAPI(title="Explainable & Uncertainty-Aware Multi-Task GNN API")

# Enable CORS for frontend connection (Vercel)
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

# Global model container
model = None
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

def load_evidential_model():
    global model
    try:
        model = MultiTaskEvidentialGNN(
            node_in_dim=31,
            edge_in_dim=6,
            hidden_dim=64,
            num_reg_tasks=2,
            num_clf_tasks=2
        )
        
        # Try local path first, otherwise download from HF
        checkpoint_path = "workspace/models/best_checkpoint.pt"
        if not os.path.exists(checkpoint_path):
            checkpoint_path = "../workspace/models/best_checkpoint.pt"
            
        if not os.path.exists(checkpoint_path):
            print("[Backend] Local checkpoint not found. Downloading from Hugging Face Hub...")
            token_val = None
            token_path = "/home/anamitra/Downloads/API_Keys_and_Secrets/hf_token"
            if os.path.exists(token_path):
                with open(token_path, 'r') as f:
                    token_val = f.read().strip()
            
            try:
                # Use HF Hub Download (downloads to HF cache)
                checkpoint_path = hf_hub_download(
                    repo_id="Arko007/multi-task-evidential-gnn", 
                    filename="best_checkpoint.pt",
                    token=token_val
                )
                print(f"[Backend] Checkpoint downloaded successfully to: {checkpoint_path}")
            except Exception as hf_err:
                print(f"[Backend] HF download failed: {hf_err}. Attempting to run uninitialized (fallback only).")
                return
                
        # Load weights
        model.load_state_dict(torch.load(checkpoint_path, map_location=device))
        model.to(device)
        model.eval()
        print("[Backend] Evidential multi-task model loaded successfully.")
    except Exception as e:
        print(f"[Backend] ERROR loading GNN model: {e}")

@app.on_event("startup")
def startup_event():
    load_evidential_model()
    # Resolve Groq Key
    if not os.environ.get("GROQ_API_KEY"):
        key_path = "/home/anamitra/Downloads/API_Keys_and_Secrets/groq_api.txt"
        if os.path.exists(key_path):
            with open(key_path, 'r') as f:
                os.environ["GROQ_API_KEY"] = f.read().strip()

# Input validation models
class SmilesRequest(BaseModel):
    smiles: str

class AnalogRequest(BaseModel):
    smiles: str
    num_variants: Optional[int] = 30
    reaction_smarts: Optional[str] = None

class GroqAuditRequest(BaseModel):
    smiles: str
    predictions: dict
    important_atoms: List[str]

# API Route: Predict
@app.post("/api/predict")
def predict_smiles(req: SmilesRequest):
    if model is None:
        raise HTTPException(status_code=500, detail="GNN model is not loaded. Please train it first.")
        
    mol = Chem.MolFromSmiles(req.smiles)
    if not mol:
        raise HTTPException(status_code=400, detail="Invalid SMILES string provided.")
        
    data = smiles_to_graph(req.smiles)
    if data is None:
        raise HTTPException(status_code=400, detail="Graph extraction failed for SMILES.")
        
    data = data.to(device)
    
    with torch.no_grad():
        # Add a batch dimension for single item inference
        batch = torch.zeros(data.x.size(0), dtype=torch.long, device=device)
        pred_reg, pred_clf = model(data.x, data.edge_index, data.edge_attr, batch)
        
    # Unpack Regressions (0: Caco-2, 1: Solubility)
    # pred_reg shape: [1, 2, 4] -> (gamma, nu, alpha, beta)
    caco2_gamma, caco2_nu, caco2_alpha, caco2_beta = pred_reg[0, 0].cpu().numpy()
    sol_gamma, sol_nu, sol_alpha, sol_beta = pred_reg[0, 1].cpu().numpy()
    
    caco2_aleatoric = float(np.sqrt(caco2_beta / (caco2_alpha - 1.0)))
    caco2_epistemic = float(1.0 / np.sqrt(caco2_nu))
    
    sol_aleatoric = float(np.sqrt(sol_beta / (sol_alpha - 1.0)))
    sol_epistemic = float(1.0 / np.sqrt(sol_nu))
    
    # Unpack Classifications (0: hERG, 1: AMES)
    # pred_clf shape: [1, 2, 2] -> Dirichlet parameters alpha_1, alpha_2
    herg_alpha = pred_clf[0, 0].cpu().numpy()
    ames_alpha = pred_clf[0, 1].cpu().numpy()
    
    herg_S = float(np.sum(herg_alpha))
    herg_prob = float(herg_alpha[1] / herg_S)
    herg_epistemic = float(2.0 / herg_S) # u = K / S, K=2
    
    ames_S = float(np.sum(ames_alpha))
    ames_prob = float(ames_alpha[1] / ames_S)
    ames_epistemic = float(2.0 / ames_S) # u = K / S, K=2
    
    return {
        "smiles": req.smiles,
        "caco2": float(caco2_gamma),
        "caco2_aleatoric": caco2_aleatoric,
        "caco2_epistemic": caco2_epistemic,
        "solubility": float(sol_gamma),
        "solubility_aleatoric": sol_aleatoric,
        "solubility_epistemic": sol_epistemic,
        "herg": herg_prob,
        "herg_epistemic": herg_epistemic,
        "ames": ames_prob,
        "ames_epistemic": ames_epistemic
    }

# API Route: GNNExplainer highlights SVG
@app.post("/api/explain")
def explain_smiles(req: SmilesRequest):
    if model is None:
        raise HTTPException(status_code=500, detail="GNN model is not loaded.")

    mol = Chem.MolFromSmiles(req.smiles)
    if not mol:
        raise HTTPException(status_code=400, detail="Invalid SMILES.")

    try:
        data = smiles_to_graph(req.smiles)
        data = data.to(device)

        # We explain hERG toxicity by default (classification task index 0)
        # If the molecule is not hERG toxic, we explain AMES (index 1)
        atom_w, edge_w = explain_molecule(model, data, task_type='clf', task_idx=0, epochs=80)

        svg_dir = "workspace/explanations"
        os.makedirs(svg_dir, exist_ok=True)
        temp_svg_path = os.path.join(svg_dir, "temp_explain.svg")

        # Render SVG (hERG is toxic/bad endpoint, so is_toxic_or_bad=True)
        render_molecular_explanation(req.smiles, atom_w, edge_w, data, temp_svg_path, is_toxic_or_bad=True)

        # Read SVG content
        with open(temp_svg_path, 'r') as f:
            svg_content = f.read()

        # Get top important atoms to return symbols
        atom_symbols = []
        for idx, w in enumerate(atom_w):
            atom_symbol = mol.GetAtomWithIdx(idx).GetSymbol()
            atom_symbols.append({"index": idx, "symbol": atom_symbol, "weight": float(w)})

        # Sort by weight descending
        atom_symbols = sorted(atom_symbols, key=lambda x: x["weight"], reverse=True)

        return {
            "svg": svg_content,
            "attributions": atom_symbols[:5] # Top 5 atoms
        }
    except Exception as e:
        import traceback
        print(f"[Backend] /api/explain failed: {e}\n{traceback.format_exc()}")
        raise HTTPException(status_code=500, detail=f"Explanation generation failed: {e}")

# API Route: Generate Structural Analogs
@app.post("/api/generate")
def generate_analogs_route(req: AnalogRequest):
    if model is None:
        raise HTTPException(status_code=500, detail="GNN model is not loaded.")
        
    analogs = generate_synthetic_analogs(req.smiles, req.reaction_smarts, req.num_variants)
    if not analogs:
        return {"analogs": []}
        
    results = []
    # Predict for each analog
    for smiles in analogs:
        data = smiles_to_graph(smiles)
        if data is None:
            continue
        data = data.to(device)
        
        with torch.no_grad():
            batch = torch.zeros(data.x.size(0), dtype=torch.long, device=device)
            pred_reg, pred_clf = model(data.x, data.edge_index, data.edge_attr, batch)
            
        caco2_gamma = float(pred_reg[0, 0, 0].cpu().item())
        sol_gamma = float(pred_reg[0, 1, 0].cpu().item())
        
        herg_alpha = pred_clf[0, 0].cpu().numpy()
        herg_prob = float(herg_alpha[1] / np.sum(herg_alpha))
        
        ames_alpha = pred_clf[0, 1].cpu().numpy()
        ames_prob = float(ames_alpha[1] / np.sum(ames_alpha))
        
        results.append({
            "smiles": smiles,
            "caco2": caco2_gamma,
            "solubility": sol_gamma,
            "herg": herg_prob,
            "ames": ames_prob,
            # Safety score (higher is safer): 1 - max(herg_prob, ames_prob)
            "safety_score": 1.0 - max(herg_prob, ames_prob)
        })
        
    # Sort by safety score and solubility (higher is better)
    results = sorted(results, key=lambda x: (x["safety_score"], x["solubility"]), reverse=True)
    
    return {"analogs": results}

# API Route: Groq Chemical Audit
@app.post("/api/explain_text")
def explain_predictions_groq_route(req: GroqAuditRequest):
    groq_key = os.environ.get("GROQ_API_KEY")
    if not groq_key:
        raise HTTPException(status_code=500, detail="Groq API Key is not set in the server environment.")
        
    try:
        from groq import Groq
        client = Groq(api_key=groq_key)
        
        prompt = f"""
        You are an expert computational toxicologist and medicinal chemist.
        A multi-task Graph Neural Network (GNN) trained on Tox21 and ADME datasets has evaluated the following molecule:
        SMILES: {req.smiles}
        
        Predictions:
        - Cardiac Toxicity (hERG Blockade): {"Toxic (hERG blocker)" if req.predictions['herg'] > 0.5 else "Safe (Non-blocker)"} (Confidence probability: {req.predictions['herg']:.2f}, Epistemic Uncertainty: {req.predictions['herg_epistemic']:.3f})
        - Mutagenicity (AMES test): {"Mutagenic" if req.predictions['ames'] > 0.5 else "Non-mutagenic"} (Confidence probability: {req.predictions['ames']:.2f}, Epistemic Uncertainty: {req.predictions['ames_epistemic']:.3f})
        - Intestinal Permeability (Caco-2 LogPapp): {req.predictions['caco2']:.3f} cm/s (Epistemic Uncertainty: {req.predictions['caco2_epistemic']:.3f})
        - Aqueous Solubility (LogS): {req.predictions['solubility']:.3f} (Epistemic Uncertainty: {req.predictions['solubility_epistemic']:.3f})
        
        GNNExplainer highlighted the following atoms/regions as highly influential for predictions: {req.important_atoms}
        
        Based on these predictions, provide a brief, professional chemical audit report explaining:
        1. The biological implications of these ADMET predictions.
        2. How the highlighted structural features (e.g. aromatic rings, hydrogen bond donors/acceptors, polar groups) correlate with the predicted toxicity and ADME profiles.
        3. Recommendations for modifying the chemical structure to improve solubility or reduce toxicity (e.g. adding specific polar groups, reducing lipophilicity, or using bioisosteres).
        
        Be concise (maximum 3 paragraphs). Use professional medicinal chemistry terms. Do not repeat the prompt.
        """
        response = client.chat.completions.create(
            model="llama-3.3-70b-versatile",
            messages=[{"role": "user", "content": prompt}],
            temperature=0.2,
            max_tokens=800
        )
        report = response.choices[0].message.content
        return {"report": report}
    except Exception as e:
        raise HTTPException(status_code=500, detail=f"Groq generation failed: {e}")

# API Route: Diagnostics
@app.get("/api/diagnostics")
def get_diagnostics():
    metrics_path = "workspace/models/test_metrics.json"
    if not os.path.exists(metrics_path):
        metrics_path = "../workspace/models/test_metrics.json"
        
    if os.path.exists(metrics_path):
        with open(metrics_path, 'r') as f:
            metrics = json.load(f)
        return metrics
    else:
        # Return fallback mock/default metrics if no training has run yet
        return {
            "reg_Caco-2_mae": 0.354,
            "reg_Caco-2_spearman": 0.684,
            "reg_Caco-2_mca": 0.082,
            "reg_Solubility_mae": 0.412,
            "reg_Solubility_spearman": 0.725,
            "reg_Solubility_mca": 0.076,
            "clf_hERG_accuracy": 0.845,
            "clf_hERG_mca": 0.048,
            "clf_AMES_accuracy": 0.812,
            "clf_AMES_mca": 0.061,
            "info": "Placeholder values (Model has not been trained locally yet. Run training to generate local metrics)."
        }

from fastapi.staticfiles import StaticFiles
dist_path = os.path.abspath(os.path.join(os.path.dirname(__file__), 'dist'))
if os.path.exists(dist_path):
    app.mount("/", StaticFiles(directory=dist_path, html=True), name="static")