fix: python_version, requirements, academic UI redesign

README.md

@@ -4,8 +4,8 @@ emoji: 🧬
 colorFrom: blue
 colorTo: green
 sdk: gradio
-sdk_version: 4.44.0
-python_version: 3.10
+sdk_version: 5.0.0
+python_version: "3.11"
 app_file: app.py
 pinned: false
 license: mit

app.py

@@ -1,552 +1,764 @@
-"""
-app.py — BioInteract Gradio Space
-Interpretable Drug–Target Interaction Prediction
-
-Two tabs:
-  1. Case Studies  — pre-computed, clinically validated pairs
-  2. Custom Prediction — user supplies SMILES + protein sequence
-"""
-import sys
-import json
-import io
-import warnings
-from pathlib import Path
-
-import numpy as np
-import yaml
-import matplotlib
-matplotlib.use('Agg')
-import matplotlib.pyplot as plt
-import seaborn as sns
-import gradio as gr
-from PIL import Image
-
-# ---------- path setup ----------
-ROOT = Path(__file__).parent
-sys.path.insert(0, str(ROOT))
-
-# torch-geometric import guard: provide useful error if wheels missing
-try:
-    from torch_geometric.data import Batch as PyGBatch
-    PYGEOMETRIC_AVAILABLE = True
-except ImportError as _pyg_err:
-    PYGEOMETRIC_AVAILABLE = False
-    _pyg_err_msg = str(_pyg_err)
-
-# protein features are pure-python and safe to import eagerly
-from src.data.protein_feat import residue_physicochemical_features, residue_domain_labels
-
-# Model and heavy-dependency lazy loader
-_model = None
-_DEVICE = None
-_CONFIG_PATH = ROOT / 'configs' / 'default.yaml'
-_CKPT_PATH   = ROOT / 'checkpoints' / 'best.pt'
-_REPORT_PATH = ROOT / 'examples' / 'interpretability_report.json'
-
-def _load_model():
-    """Lazy-load the PyTorch model and return a ready-to-use instance.
-    This delays importing heavy packages (torch, torch_geometric, rdkit)
-    until a user requests a custom prediction.
-    """
-    global _model, _DEVICE
-    if _model is not None:
-        return _model
-
-    try:
-        import torch
-    except Exception as e:
-        raise RuntimeError(f"PyTorch not available: {e}")
-
-    # Import model code only when torch is available
-    from src.models.biointeract import BioInteract
-
-    print("[BioInteract] Loading model config …")
-    with open(_CONFIG_PATH) as f:
-        cfg = yaml.safe_load(f)
-
-    print("[BioInteract] Loading pretrained weights …")
-    device = torch.device('cpu')
-    model = BioInteract(cfg['model']).to(device)
-    ckpt = torch.load(_CKPT_PATH, map_location='cpu', weights_only=False)
-    model.load_state_dict(ckpt['model_state_dict'])
-    model.eval()
-    _model = model
-    _DEVICE = device
-    print(f"[BioInteract] Model ready — epoch {ckpt.get('epoch','?')}, params = {sum(p.numel() for p in _model.parameters()):,}")
-    return _model
-
-# Load pre-computed case study report (safe, JSON-only)
-with open(_REPORT_PATH) as f:
-    _REPORT = json.load(f)
-
-# ============================================================
-# ESM-2 lazy loader (only initialised on first custom prediction)
-# ============================================================
-
-_esm_tokenizer = None
-_esm_model     = None
-ESM_MODEL_NAME = "facebook/esm2_t30_150M_UR50D"
-
-def _get_esm():
-    global _esm_tokenizer, _esm_model
-    if _esm_model is None:
-        from transformers import EsmModel, EsmTokenizer
-        print("[BioInteract] Downloading / loading ESM-2 (150M) …")
-        _esm_tokenizer = EsmTokenizer.from_pretrained(ESM_MODEL_NAME)
-        _esm_model     = EsmModel.from_pretrained(ESM_MODEL_NAME).eval()
-        print("[BioInteract] ESM-2 ready")
-    return _esm_tokenizer, _esm_model
-
-
-def compute_esm2_embedding(sequence: str, max_len: int = 512) -> torch.Tensor:
-    """Run ESM-2 and return per-residue embeddings (L, 640)."""
-    import torch
-    seq = sequence[:max_len]
-    tokenizer, esm = _get_esm()
-    inputs = tokenizer(seq, return_tensors='pt', add_special_tokens=True)
-    with torch.no_grad():
-        outputs = esm(**inputs)
-    # strip [CLS] and [EOS] tokens → (L, 640)
-    embedding = outputs.last_hidden_state[0, 1:-1, :]
-    return embedding[:len(seq)]
-
-
-# ============================================================
-# Plotting helpers
-# ============================================================
-
-def _plot_interaction_heatmap(
-    interaction_map: np.ndarray,
-    sequence: str,
-    title: str = 'Atom–Residue Interaction Map',
-) -> Image.Image:
-    """Render interaction heatmap as a PIL Image."""
-    n_atoms, n_res = interaction_map.shape
-
-    # subsample residues if too long
-    max_show_res = 80
-    if n_res > max_show_res:
-        scores = interaction_map.sum(axis=0)
-        center = int(np.argmax(scores))
-        start  = max(0, center - max_show_res // 2)
-        end    = min(n_res, start + max_show_res)
-        interaction_map = interaction_map[:, start:end]
-        res_labels = [f"{seq[i]}{i+1}" for i, seq in
-                      enumerate([sequence[j] for j in range(start, end)])]
-        # simpler: just residue index labels
-        res_labels = [f"{sequence[i]}{i+1}" if i < len(sequence) else str(i+1)
-                      for i in range(start, end)]
-    else:
-        res_labels = [f"{sequence[i]}{i+1}" if i < len(sequence) else str(i+1)
-                      for i in range(n_res)]
-
-    figw = max(12, len(res_labels) * 0.15)
-    figh = max(5, n_atoms * 0.25)
-    fig, ax = plt.subplots(figsize=(figw, figh))
-    sns.heatmap(
-        interaction_map,
-        xticklabels=res_labels,
-        yticklabels=[f"A{i}" for i in range(n_atoms)],
-        cmap='YlOrRd',
-        ax=ax,
-        cbar_kws={'label': 'Attention Score', 'shrink': 0.8},
-    )
-    ax.set_title(title, fontsize=13, fontweight='bold')
-    ax.set_xlabel('Protein Residue', fontsize=10)
-    ax.set_ylabel('Drug Atom', fontsize=10)
-    plt.xticks(rotation=90, fontsize=6)
-    plt.yticks(fontsize=6)
-    plt.tight_layout()
-
-    buf = io.BytesIO()
-    fig.savefig(buf, format='png', dpi=150, bbox_inches='tight')
-    plt.close(fig)
-    buf.seek(0)
-    return Image.open(buf).copy()
-
-
-def _plot_top_residues(top_residues: list, title: str = 'Top Binding Residues') -> Image.Image:
-    """Bar chart of top residue attention scores."""
-    labels = [r[0] for r in top_residues]
-    scores = [r[1] for r in top_residues]
-
-    fig, ax = plt.subplots(figsize=(8, 4))
-    bars = ax.barh(labels[::-1], scores[::-1], color='steelblue', edgecolor='white')
-    ax.set_xlabel('Normalised Attention Score', fontsize=10)
-    ax.set_title(title, fontsize=11, fontweight='bold')
-    ax.set_xlim(0, 1.05)
-    for bar, score in zip(bars, scores[::-1]):
-        ax.text(score + 0.01, bar.get_y() + bar.get_height() / 2,
-                f'{score:.3f}', va='center', fontsize=8)
-    plt.tight_layout()
-
-    buf = io.BytesIO()
-    fig.savefig(buf, format='png', dpi=150, bbox_inches='tight')
-    plt.close(fig)
-    buf.seek(0)
-    return Image.open(buf).copy()
-
-
-def _plot_functional_groups(fg_dict: dict, title: str = 'Pharmacophore Importance') -> Image.Image:
-    """Horizontal bar chart for functional group Grad-CAM scores."""
-    if not fg_dict:
-        return None
-    labels = list(fg_dict.keys())
-    scores = list(fg_dict.values())
-
-    fig, ax = plt.subplots(figsize=(7, max(3, len(labels) * 0.5)))
-    ax.barh(labels, scores, color='coral', edgecolor='white')
-    ax.set_xlabel('Grad-CAM Importance', fontsize=10)
-    ax.set_title(title, fontsize=11, fontweight='bold')
-    ax.set_xlim(0, 1.05)
-    plt.tight_layout()
-
-    buf = io.BytesIO()
-    fig.savefig(buf, format='png', dpi=150, bbox_inches='tight')
-    plt.close(fig)
-    buf.seek(0)
-    return Image.open(buf).copy()
-
-
-# ============================================================
-# Tab 1 — Case Studies
-# ============================================================
-
-_CASE_OPTIONS = {
-    f"{cs['drug_name']} / {cs['target_name']}": cs
-    for cs in _REPORT.get('case_studies', [])
-    if (ROOT / 'examples' / f"{cs['drug_name']}_{cs['target_name'].replace('(','').replace(')','').replace('/','')}.png").exists()
-       or (ROOT / 'examples' / f"{cs['drug_name']}_{cs['target_name']}.png").exists()
-}
-
-# If the report has no exact-match filenames, fall back to the 3 available PNGs
-_FIXED_CASES = {
-    'ABL1(E255K) + Drug 5328940 (Kd = 0.047 nM)': {
-        'png': ROOT / 'examples' / '5328940_ABL1E255K.png',
-        'prob': 0.988,
-        'affinity_nM': 0.047,
-        'top_residues': [['V104', 1.0], ['A648', 0.503], ['S199', 0.417],
-                         ['P649', 0.281], ['P936', 0.237], ['N707', 0.192],
-                         ['P651', 0.180], ['L799', 0.166], ['K796', 0.141], ['P934', 0.079]],
-        'functional_groups': {'Halogen': 0.650, 'Amino': 0.616, 'Ether': 0.345,
-                              'Aromatic Ring': 0.311, 'Heterocycle N': 0.167},
-        'description': (
-            'Drug 5328940 bound to ABL1 E255K resistance mutant. '
-            'Prediction probability 98.8 % (Kd = 0.047 nM). '
-            'Key contacts: V104 (gatekeeper), A648, S199.'
-        ),
-    },
-    'EGFR + Drug 156414': {
-        'png': ROOT / 'examples' / '156414_EGFR.png',
-        'prob': None,
-        'affinity_nM': None,
-        'top_residues': [],
-        'functional_groups': {},
-        'description': (
-            'Drug 156414 bound to wild-type EGFR kinase domain. '
-            'Inhibitor with selectivity for the EGFR tyrosine kinase.'
-        ),
-    },
-    'BRAF + Drug 11717001 (Sorafenib analogue)': {
-        'png': ROOT / 'examples' / '11717001_BRAF.png',
-        'prob': None,
-        'affinity_nM': None,
-        'top_residues': [],
-        'functional_groups': {},
-        'description': (
-            'Drug 11717001 (Sorafenib analogue) bound to BRAF kinase. '
-            'RAF inhibitor used in treatment of melanoma and other cancers.'
-        ),
-    },
-}
-
-
-def show_case_study(case_key: str):
-    """Called when user selects a case from the dropdown."""
-    case = _FIXED_CASES.get(case_key)
-    if case is None:
-        return None, None, None, "Case not found."
-
-    heatmap_img = Image.open(case['png']) if case['png'].exists() else None
-
-    prob_text = (f"**Predicted binding probability:** {case['prob']*100:.1f}%\n\n"
-                 if case['prob'] else "")
-    kd_text   = (f"**Experimental Kd:** {case['affinity_nM']} nM\n\n"
-                 if case['affinity_nM'] else "")
-    info_md   = f"### {case_key}\n\n{prob_text}{kd_text}{case['description']}"
-
-    residue_img = (
-        _plot_top_residues(case['top_residues'], f"Top Binding Residues — {case_key}")
-        if case['top_residues'] else None
-    )
-    fg_img = (
-        _plot_functional_groups(case['functional_groups'], f"Pharmacophore Importance — {case_key}")
-        if case['functional_groups'] else None
-    )
-
-    return heatmap_img, residue_img, fg_img, info_md
-
-
-# ============================================================
-# Tab 2 — Custom Prediction
-# ============================================================
-
-# Example: Imatinib SMILES + first 256 aa of ABL1 kinase domain
-_EXAMPLE_SMILES = (
-    "Cc1ccc(NC(=O)c2ccc(CN3CCN(C)CC3)cc2)cc1Nc1nccc(-c2cccnc2)n1"
-)
-_EXAMPLE_SEQUENCE = (
-    "MGPSENDPNLFVALYDFVASGDNTLSITKGEKLRVLGYNHNGEWCEAQTKNGQGWVPSNYITPVNSLEKHSWYHGPVSRNAAEYLLSSGINGSFLVRESESSPGQRSISLRYEGRVYHYRINTASDGKLYVSSESRFNTLAELVHHHSTLVQHSDSVESAYRSKL"
-    "LNSGVYHYRINTASDGKLYVSSESRFNTLAELVHHHSTLVQ"
-)
-
-MAX_SEQ_LEN = 512
-
-
-def run_prediction(smiles: str, sequence: str, progress=gr.Progress()):
-    """
-    Full inference pipeline: SMILES + sequence → probability + interaction heatmap.
-    Returns: (status_msg, prob_text, heatmap_image, residue_image)
-    """
-    smiles    = (smiles or '').strip()
-    sequence  = (sequence or '').strip().upper()
-
-    if not smiles:
-        return "Please enter a SMILES string.", "", None, None
-    if not sequence:
-        return "Please enter a protein sequence (amino acids).", "", None, None
-
-    try:
-        _load_model()
-    except Exception as e:
-        return (f"Model not available: {e}", "", None, None)
-
-    if not PYGEOMETRIC_AVAILABLE:
-        return (
-            f"torch-geometric is not available in this environment: {_pyg_err_msg}. "
-            "Please use the Case Studies tab for pre-computed results.",
-            "", None, None
-        )
-
-    # --- Drug graph ---
-    progress(0.1, desc="Parsing SMILES …")
-    # import heavy SMILES->graph conversion lazily
-    from src.data.mol_graph import smiles_to_graph
-    drug_graph = smiles_to_graph(smiles)
-    if drug_graph is None:
-        return "Invalid SMILES string — RDKit could not parse it.", "", None, None
-
-    import torch
-    device = _DEVICE
-    drug_batch = PyGBatch.from_data_list([drug_graph]).to(device)
-
-    # --- Protein features ---
-    seq = sequence[:MAX_SEQ_LEN]
-    if len(sequence) > MAX_SEQ_LEN:
-        warnings.warn(f"Sequence truncated to {MAX_SEQ_LEN} residues.")
-    L = len(seq)
-
-    progress(0.2, desc="Computing ESM-2 embeddings (may take 1–3 min on first run) …")
-    try:
-        esm2_emb = compute_esm2_embedding(seq)       # (L, 640)
-    except Exception as e:
-        return f"ESM-2 error: {e}", "", None, None
-
-    esm2_emb = esm2_emb.unsqueeze(0).to(device)      # (1, L, 640)
-    physchem = residue_physicochemical_features(seq).unsqueeze(0).to(device)  # (1, L, 4)
-    domain   = residue_domain_labels(L).unsqueeze(0).to(device)               # (1, L)
-    prot_mask = torch.ones(1, L, dtype=torch.bool, device=device)
-
-    # --- Model forward ---
-    progress(0.85, desc="Running BioInteract model …")
-    with torch.no_grad():
-        logit, attn_data = _model(
-            drug_batch, esm2_emb, physchem, domain, prot_mask,
-            return_attention=True
-        )
-
-    prob = torch.sigmoid(logit).item()
-    interaction_map = attn_data['interaction_map'][0].cpu().numpy()   # (N_atoms, L)
-    drug_mask_np    = attn_data['drug_mask'][0].cpu().numpy()          # (N_atoms,)
-    prot_mask_np    = prot_mask[0].cpu().numpy()                       # (L,)
-
-    # Trim padding
-    n_real_atoms = int(drug_mask_np.sum())
-    imap = interaction_map[:n_real_atoms, :L]
-
-    # Top-10 residues
-    residue_scores = imap.sum(axis=0)
-    residue_scores = residue_scores / (residue_scores.max() + 1e-9)
-    top_idx = np.argsort(residue_scores)[::-1][:10]
-    top_residues = [[f"{seq[i]}{i+1}", float(residue_scores[i])] for i in top_idx]
-
-    # --- Plots ---
-    progress(0.95, desc="Generating plots …")
-    heatmap_img  = _plot_interaction_heatmap(imap, seq, title='Atom–Residue Interaction Map')
-    residue_img  = _plot_top_residues(top_residues, 'Top 10 Binding Residues')
-
-    label = "BINDING" if prob > 0.5 else "NON-BINDING"
-    prob_text = (
-        f"## Predicted: {label}\n\n"
-        f"**Binding probability:** {prob * 100:.1f}%  \n"
-        f"**Atoms analysed:** {n_real_atoms}  \n"
-        f"**Residues analysed:** {L}"
-    )
-
-    return "Prediction complete.", prob_text, heatmap_img, residue_img
-
-
-# ============================================================
-# Global statistics panel (sidebar info)
-# ============================================================
-
-_GLOBAL_STATS = _REPORT.get('global_stats', {})
-_stats_md = (
-    "### Model Statistics (Davis Dataset)\n\n"
-    f"- **AUROC (random split):** 0.921\n"
-    f"- **AUROC (cold-target):** 0.941\n"
-    f"- **Attention sparsity:** {_GLOBAL_STATS.get('attention_sparsity', 0.992)*100:.1f}%\n"
-    f"- **Training samples:** {_GLOBAL_STATS.get('n_samples', 1506):,}\n"
-    f"- **Model params:** {_REPORT.get('model_info', {}).get('params', 2442083):,}\n"
-)
-
-
-# ============================================================
-# Build Gradio UI
-# ============================================================
-
-_HEADER = """
-<div style="text-align:center; padding: 16px 0 8px 0;">
-  <h1 style="margin:0; font-size:2rem;">🧬 BioInteract</h1>
-  <p style="margin:4px 0; color:#555; font-size:1.05rem;">
-    Interpretable Drug–Target Interaction Prediction via Residue-Level Cross-Attention
-  </p>
-  <p style="margin:0; font-size:0.9rem; color:#888;">
-    GINE molecular graph encoder · ESM-2 protein language model · Bidirectional cross-attention
-  </p>
-</div>
-"""
-
-with gr.Blocks(
-    title="BioInteract — Interpretable DTI Prediction",
-    theme=gr.themes.Soft(primary_hue="blue"),
-) as demo:
-
-    gr.HTML(_HEADER)
-
-    with gr.Row():
-        with gr.Column(scale=3):
-            with gr.Tabs():
-
-                # ── Tab 1: Case Studies ────────────────────────────────────
-                with gr.Tab("📊 Case Studies"):
-                    gr.Markdown(
-                        "Select a **pre-computed case study** to explore the model's "
-                        "atom–residue interaction map and binding residue predictions "
-                        "for clinically validated drug–target pairs."
-                    )
-                    case_dropdown = gr.Dropdown(
-                        choices=list(_FIXED_CASES.keys()),
-                        value=list(_FIXED_CASES.keys())[0],
-                        label="Select drug–target pair",
-                        interactive=True,
-                    )
-                    case_info_md = gr.Markdown()
-
-                    with gr.Row():
-                        case_heatmap   = gr.Image(label="Atom–Residue Interaction Heatmap",
-                                                  type='pil', height=400)
-                        with gr.Column():
-                            case_residues  = gr.Image(label="Top Binding Residues",
-                                                      type='pil', height=250)
-                            case_fg        = gr.Image(label="Pharmacophore Importance",
-                                                      type='pil', height=250)
-
-                    case_dropdown.change(
-                        fn=show_case_study,
-                        inputs=case_dropdown,
-                        outputs=[case_heatmap, case_residues, case_fg, case_info_md],
-                    )
-                    # Auto-load first case on startup
-                    demo.load(
-                        fn=show_case_study,
-                        inputs=[gr.State(list(_FIXED_CASES.keys())[0])],
-                        outputs=[case_heatmap, case_residues, case_fg, case_info_md],
-                    )
-
-                # ── Tab 2: Custom Prediction ───────────────────────────────
-                with gr.Tab("🔬 Custom Prediction"):
-                    gr.Markdown(
-                        "Enter any **drug SMILES** and **protein amino acid sequence** "
-                        "to get a binding prediction with an interaction heatmap.  \n"
-                        "> ⏳ First prediction loads ESM-2 (150 M params) on CPU — allow **1–3 minutes**.  \n"
-                        "> Sequences longer than 512 residues are automatically truncated."
-                    )
-                    with gr.Row():
-                        smiles_box = gr.Textbox(
-                            label="Drug SMILES",
-                            placeholder="e.g. Cc1ccc(NC(=O)c2ccc(CN3CCN(C)CC3)cc2)cc1Nc1nccc(-c2cccnc2)n1",
-                            lines=2,
-                        )
-                    sequence_box = gr.Textbox(
-                        label="Protein Amino Acid Sequence",
-                        placeholder="e.g. MGPSENDPNLFVALYDFVASGD...",
-                        lines=4,
-                        max_lines=8,
-                    )
-
-                    with gr.Row():
-                        example_btn = gr.Button("📋 Load Imatinib / ABL1 Example", variant="secondary")
-                        predict_btn = gr.Button("🚀 Predict", variant="primary")
-
-                    status_box = gr.Textbox(label="Status", interactive=False, lines=1)
-                    prob_md    = gr.Markdown()
-
-                    with gr.Row():
-                        pred_heatmap  = gr.Image(label="Atom–Residue Interaction Heatmap",
-                                                 type='pil', height=400)
-                        pred_residues = gr.Image(label="Top 10 Binding Residues",
-                                                 type='pil', height=300)
-
-                    example_btn.click(
-                        fn=lambda: (_EXAMPLE_SMILES, _EXAMPLE_SEQUENCE),
-                        inputs=[],
-                        outputs=[smiles_box, sequence_box],
-                    )
-                    predict_btn.click(
-                        fn=run_prediction,
-                        inputs=[smiles_box, sequence_box],
-                        outputs=[status_box, prob_md, pred_heatmap, pred_residues],
-                    )
-
-        # ── Sidebar ────────────────────────────────────────────────────────
-        with gr.Column(scale=1):
-            gr.Markdown(_stats_md)
-            gr.Markdown(
-                "### Architecture\n\n"
-                "```\n"
-                "Drug SMILES\n"
-                "  → GINE (3-layer)\n"
-                "  → atom repr (N×256)\n\n"
-                "Protein sequence\n"
-                "  → ESM-2 (150M)\n"
-                "  → physicochemical\n"
-                "  → residue repr (L×256)\n\n"
-                "Bidirectional\n"
-                "cross-attention\n"
-                "  → interaction map\n"
-                "  → gated pooling\n"
-                "  → binding score\n"
-                "```"
-            )
-            gr.Markdown(
-                "### Citation\n\n"
-                "Wang S, Zhang Q et al.  \n"
-                "*BioInteract: Interpretable DTI Prediction via "
-                "Residue-Level Cross-Attention with Biological Prior Knowledge.*  \n"
-                "PLOS Computational Biology, 2026."
-            )
-
-
-if __name__ == "__main__":
-    demo.launch(server_name="0.0.0.0", server_port=7860)
+"""
+app.py — BioInteract Gradio Space
+Interpretable Drug–Target Interaction Prediction
+
+Two tabs:
+  1. Case Studies  — pre-computed, clinically validated pairs
+  2. Custom Prediction — user supplies SMILES + protein sequence
+"""
+import sys
+import json
+import io
+import warnings
+from pathlib import Path
+
+import numpy as np
+import torch
+import yaml
+import matplotlib
+matplotlib.use('Agg')
+import matplotlib as mpl
+import matplotlib.pyplot as plt
+import seaborn as sns
+import gradio as gr
+from PIL import Image
+
+# ---------- path setup ----------
+ROOT = Path(__file__).parent
+sys.path.insert(0, str(ROOT))
+
+from src.models.biointeract import BioInteract
+from src.data.mol_graph import smiles_to_graph
+from src.data.protein_feat import residue_physicochemical_features, residue_domain_labels
+
+# ---------- publication-style matplotlib defaults ----------
+mpl.rcParams.update({
+    'font.family': 'DejaVu Serif',
+    'font.size': 10,
+    'axes.titlesize': 11,
+    'axes.titleweight': 'bold',
+    'axes.labelsize': 10,
+    'axes.labelcolor': '#1a1a2e',
+    'axes.edgecolor': '#444',
+    'axes.linewidth': 0.8,
+    'axes.spines.top': False,
+    'axes.spines.right': False,
+    'xtick.direction': 'out',
+    'ytick.direction': 'out',
+    'xtick.color': '#444',
+    'ytick.color': '#444',
+    'figure.facecolor': 'white',
+    'axes.facecolor': '#fafafa',
+    'grid.color': '#e0e0e0',
+    'grid.linewidth': 0.5,
+    'savefig.facecolor': 'white',
+    'savefig.dpi': 150,
+})
+
+# ============================================================
+# Global model loading
+# ============================================================
+
+DEVICE = torch.device('cpu')
+_CONFIG_PATH = ROOT / 'configs' / 'default.yaml'
+_CKPT_PATH   = ROOT / 'checkpoints' / 'best.pt'
+_REPORT_PATH = ROOT / 'examples' / 'interpretability_report.json'
+
+print("[BioInteract] Loading model config …")
+with open(_CONFIG_PATH) as f:
+    _CONFIG = yaml.safe_load(f)
+
+print("[BioInteract] Loading pretrained weights …")
+_model = BioInteract(_CONFIG['model']).to(DEVICE)
+_ckpt = torch.load(_CKPT_PATH, map_location='cpu', weights_only=False)
+_model.load_state_dict(_ckpt['model_state_dict'])
+_model.eval()
+print(f"[BioInteract] Model ready — epoch {_ckpt.get('epoch','?')}, "
+      f"params = {sum(p.numel() for p in _model.parameters()):,}")
+
+with open(_REPORT_PATH) as f:
+    _REPORT = json.load(f)
+
+# ============================================================
+# ESM-2 lazy loader
+# ============================================================
+
+_esm_tokenizer = None
+_esm_model     = None
+ESM_MODEL_NAME = "facebook/esm2_t30_150M_UR50D"
+
+
+def _get_esm():
+    global _esm_tokenizer, _esm_model
+    if _esm_model is None:
+        from transformers import EsmModel, EsmTokenizer
+        print("[BioInteract] Downloading / loading ESM-2 (150M) …")
+        _esm_tokenizer = EsmTokenizer.from_pretrained(ESM_MODEL_NAME)
+        _esm_model     = EsmModel.from_pretrained(ESM_MODEL_NAME).eval()
+        print("[BioInteract] ESM-2 ready")
+    return _esm_tokenizer, _esm_model
+
+
+def compute_esm2_embedding(sequence: str, max_len: int = 512) -> torch.Tensor:
+    """Run ESM-2 and return per-residue embeddings (L, 640)."""
+    seq = sequence[:max_len]
+    tokenizer, esm = _get_esm()
+    inputs = tokenizer(seq, return_tensors='pt', add_special_tokens=True)
+    with torch.no_grad():
+        outputs = esm(**inputs)
+    embedding = outputs.last_hidden_state[0, 1:-1, :]
+    return embedding[:len(seq)]
+
+
+# ============================================================
+# Publication-quality plotting helpers
+# ============================================================
+
+_HEATMAP_CMAP  = 'Blues'
+_BAR_COLOR     = '#1a4a7a'
+_BAR_ACCENT    = '#2e7cbf'
+_FG_COLOR      = '#8b2635'
+
+
+def _plot_interaction_heatmap(
+    interaction_map: np.ndarray,
+    sequence: str,
+    title: str = 'Atom–Residue Interaction Map',
+) -> Image.Image:
+    """Render interaction heatmap with publication-quality styling."""
+    n_atoms, n_res = interaction_map.shape
+
+    max_show_res = 80
+    if n_res > max_show_res:
+        scores = interaction_map.sum(axis=0)
+        center = int(np.argmax(scores))
+        start  = max(0, center - max_show_res // 2)
+        end    = min(n_res, start + max_show_res)
+        interaction_map = interaction_map[:, start:end]
+        res_labels = [f"{sequence[i]}{i+1}" if i < len(sequence) else str(i+1)
+                      for i in range(start, end)]
+    else:
+        res_labels = [f"{sequence[i]}{i+1}" if i < len(sequence) else str(i+1)
+                      for i in range(n_res)]
+
+    figw = max(13, len(res_labels) * 0.16)
+    figh = max(5, n_atoms * 0.28)
+    fig, ax = plt.subplots(figsize=(figw, figh))
+
+    sns.heatmap(
+        interaction_map,
+        xticklabels=res_labels,
+        yticklabels=[f"a{i+1}" for i in range(n_atoms)],
+        cmap=_HEATMAP_CMAP,
+        ax=ax,
+        linewidths=0,
+        cbar_kws={'label': 'Normalised Attention Score', 'shrink': 0.75,
+                  'aspect': 20},
+    )
+    ax.set_title(title, pad=10)
+    ax.set_xlabel('Protein Residue', labelpad=6)
+    ax.set_ylabel('Drug Atom', labelpad=6)
+    plt.xticks(rotation=90, fontsize=5.5)
+    plt.yticks(fontsize=6, rotation=0)
+
+    for spine in ax.spines.values():
+        spine.set_visible(False)
+
+    fig.text(0.02, 0.01,
+             'Colour intensity encodes normalised cross-attention weight',
+             fontsize=7, color='#666', style='italic')
+    plt.tight_layout(rect=[0, 0.03, 1, 1])
+
+    buf = io.BytesIO()
+    fig.savefig(buf, format='png', bbox_inches='tight')
+    plt.close(fig)
+    buf.seek(0)
+    return Image.open(buf).copy()
+
+
+def _plot_top_residues(top_residues: list, title: str = 'Top Binding Residues') -> Image.Image:
+    """Horizontal bar chart with academic styling."""
+    labels = [r[0] for r in top_residues]
+    scores = [r[1] for r in top_residues]
+
+    fig, ax = plt.subplots(figsize=(8, 4.2))
+    colors = [_BAR_COLOR if s >= 0.5 else _BAR_ACCENT for s in scores[::-1]]
+    bars = ax.barh(labels[::-1], scores[::-1],
+                   color=colors, edgecolor='none', height=0.65)
+
+    ax.set_xlabel('Normalised Attention Score', labelpad=6)
+    ax.set_title(title, pad=8)
+    ax.set_xlim(0, 1.12)
+    ax.axvline(0.5, color='#aaa', linewidth=0.8, linestyle='--', alpha=0.7)
+    ax.text(0.51, -0.6, 'threshold', fontsize=7, color='#888', style='italic')
+
+    for bar, score in zip(bars, scores[::-1]):
+        ax.text(score + 0.015, bar.get_y() + bar.get_height() / 2,
+                f'{score:.3f}', va='center', fontsize=8, color='#222')
+
+    ax.set_axisbelow(True)
+    ax.yaxis.set_tick_params(labelsize=9)
+    plt.tight_layout()
+
+    buf = io.BytesIO()
+    fig.savefig(buf, format='png', bbox_inches='tight')
+    plt.close(fig)
+    buf.seek(0)
+    return Image.open(buf).copy()
+
+
+def _plot_functional_groups(fg_dict: dict, title: str = 'Pharmacophore Importance') -> Image.Image:
+    """Horizontal bar chart for functional group Grad-CAM scores."""
+    if not fg_dict:
+        return None
+    labels = list(fg_dict.keys())
+    scores = list(fg_dict.values())
+
+    fig, ax = plt.subplots(figsize=(7, max(3, len(labels) * 0.55)))
+    ax.barh(labels, scores, color=_FG_COLOR, edgecolor='none', height=0.6, alpha=0.85)
+    ax.set_xlabel('Grad-CAM Importance Score', labelpad=6)
+    ax.set_title(title, pad=8)
+    ax.set_xlim(0, 1.12)
+    for i, (label, score) in enumerate(zip(labels, scores)):
+        ax.text(score + 0.015, i, f'{score:.3f}', va='center', fontsize=8, color='#222')
+    ax.set_axisbelow(True)
+    plt.tight_layout()
+
+    buf = io.BytesIO()
+    fig.savefig(buf, format='png', bbox_inches='tight')
+    plt.close(fig)
+    buf.seek(0)
+    return Image.open(buf).copy()
+
+
+# ============================================================
+# Tab 1 — Case Studies
+# ============================================================
+
+_FIXED_CASES = {
+    'ABL1(E255K) + Drug 5328940  (Kd = 0.047 nM)': {
+        'png': ROOT / 'examples' / '5328940_ABL1E255K.png',
+        'prob': 0.988,
+        'affinity_nM': 0.047,
+        'top_residues': [['V104', 1.0], ['A648', 0.503], ['S199', 0.417],
+                         ['P649', 0.281], ['P936', 0.237], ['N707', 0.192],
+                         ['P651', 0.180], ['L799', 0.166], ['K796', 0.141], ['P934', 0.079]],
+        'functional_groups': {'Halogen': 0.650, 'Amino': 0.616, 'Ether': 0.345,
+                              'Aromatic Ring': 0.311, 'Heterocycle N': 0.167},
+        'description': (
+            'Drug 5328940 binds the ABL1 E255K resistance mutant with extremely high '
+            'affinity (Kd = 0.047 nM). The model predicts binding with 98.8% probability. '
+            'Key contacts include V104 (gatekeeper residue), A648, and S199, consistent '
+            'with known structural data for Type II kinase inhibitors.'
+        ),
+    },
+    'EGFR + Drug 156414': {
+        'png': ROOT / 'examples' / '156414_EGFR.png',
+        'prob': None,
+        'affinity_nM': None,
+        'top_residues': [],
+        'functional_groups': {},
+        'description': (
+            'Drug 156414 targets the wild-type EGFR kinase domain. '
+            'EGFR inhibitors are first-line treatments for non-small cell lung cancer '
+            'with activating mutations. The interaction map highlights the ATP-binding cleft.'
+        ),
+    },
+    'BRAF + Drug 11717001  (Sorafenib analogue)': {
+        'png': ROOT / 'examples' / '11717001_BRAF.png',
+        'prob': None,
+        'affinity_nM': None,
+        'top_residues': [],
+        'functional_groups': {},
+        'description': (
+            'Drug 11717001 is a Sorafenib analogue targeting BRAF kinase, a driver '
+            'oncogene in ~50% of cutaneous melanomas (V600E mutation). RAF inhibitors '
+            'block the MAPK/ERK signalling cascade that promotes uncontrolled proliferation.'
+        ),
+    },
+}
+
+
+def show_case_study(case_key: str):
+    """Called when user selects a case from the dropdown."""
+    case = _FIXED_CASES.get(case_key)
+    if case is None:
+        return None, None, None, "Case not found."
+
+    heatmap_img = Image.open(case['png']) if case['png'].exists() else None
+
+    prob_text = (f"**Predicted binding probability:** {case['prob']*100:.1f}%\n\n"
+                 if case['prob'] else "")
+    kd_text   = (f"**Experimental affinity (Kd):** {case['affinity_nM']} nM\n\n"
+                 if case['affinity_nM'] else "")
+    info_md   = (
+        f"#### {case_key.strip()}\n\n"
+        f"{prob_text}{kd_text}"
+        f"**Clinical context:** {case['description']}"
+    )
+
+    residue_img = (
+        _plot_top_residues(case['top_residues'],
+                           f'Top Binding Residues — {case_key.split("+")[0].strip()}')
+        if case['top_residues'] else None
+    )
+    fg_img = (
+        _plot_functional_groups(case['functional_groups'],
+                                f'Pharmacophore Importance — {case_key.split("+")[0].strip()}')
+        if case['functional_groups'] else None
+    )
+
+    return heatmap_img, residue_img, fg_img, info_md
+
+
+# ============================================================
+# Tab 2 — Custom Prediction
+# ============================================================
+
+_EXAMPLE_SMILES = (
+    "Cc1ccc(NC(=O)c2ccc(CN3CCN(C)CC3)cc2)cc1Nc1nccc(-c2cccnc2)n1"
+)
+_EXAMPLE_SEQUENCE = (
+    "MGPSENDPNLFVALYDFVASGDNTLSITKGEKLRVLGYNHNGEWCEAQTKNGQGWVPSNYITPVNSLEKHSWYHGPVSRNAAEYLLSSGINGSFLVRESESSPGQRSISLRYEGRVYHYRINTASDGKLYVSSESRFNTLAELVHHHSTLVQHSDSVESAYRSKL"
+    "LNSGVYHYRINTASDGKLYVSSESRFNTLAELVHHHSTLVQ"
+)
+
+MAX_SEQ_LEN = 512
+
+
+def run_prediction(smiles: str, sequence: str, progress=gr.Progress()):
+    """
+    Full inference pipeline: SMILES + sequence → probability + interaction heatmap.
+    Returns: (status_msg, prob_text, heatmap_image, residue_image)
+    """
+    smiles    = (smiles or '').strip()
+    sequence  = (sequence or '').strip().upper()
+
+    if not smiles:
+        return "Input required: please provide a SMILES string.", "", None, None
+    if not sequence:
+        return "Input required: please provide an amino acid sequence.", "", None, None
+
+    progress(0.1, desc="Parsing SMILES string via RDKit …")
+    drug_graph = smiles_to_graph(smiles)
+    if drug_graph is None:
+        return "Parse error: RDKit could not interpret the SMILES string.", "", None, None
+
+    from torch_geometric.data import Batch
+    drug_batch = Batch.from_data_list([drug_graph]).to(DEVICE)
+
+    seq = sequence[:MAX_SEQ_LEN]
+    if len(sequence) > MAX_SEQ_LEN:
+        warnings.warn(f"Sequence truncated to {MAX_SEQ_LEN} residues.")
+    L = len(seq)
+
+    progress(0.2, desc="Computing ESM-2 residue embeddings (first run: ~2 min) …")
+    try:
+        esm2_emb = compute_esm2_embedding(seq)
+    except Exception as e:
+        return f"ESM-2 error: {e}", "", None, None
+
+    esm2_emb  = esm2_emb.unsqueeze(0).to(DEVICE)
+    physchem  = residue_physicochemical_features(seq).unsqueeze(0).to(DEVICE)
+    domain    = residue_domain_labels(L).unsqueeze(0).to(DEVICE)
+    prot_mask = torch.ones(1, L, dtype=torch.bool, device=DEVICE)
+
+    progress(0.85, desc="Running BioInteract cross-attention inference …")
+    with torch.no_grad():
+        logit, attn_data = _model(
+            drug_batch, esm2_emb, physchem, domain, prot_mask,
+            return_attention=True
+        )
+
+    prob = torch.sigmoid(logit).item()
+    interaction_map = attn_data['interaction_map'][0].cpu().numpy()
+    drug_mask_np    = attn_data['drug_mask'][0].cpu().numpy()
+
+    n_real_atoms = int(drug_mask_np.sum())
+    imap = interaction_map[:n_real_atoms, :L]
+
+    residue_scores = imap.sum(axis=0)
+    residue_scores = residue_scores / (residue_scores.max() + 1e-9)
+    top_idx = np.argsort(residue_scores)[::-1][:10]
+    top_residues = [[f"{seq[i]}{i+1}", float(residue_scores[i])] for i in top_idx]
+
+    progress(0.95, desc="Generating publication-quality figures …")
+    heatmap_img = _plot_interaction_heatmap(imap, seq, title='Atom–Residue Cross-Attention Map')
+    residue_img = _plot_top_residues(top_residues, 'Top 10 Predicted Binding Residues')
+
+    label = "**BINDING**" if prob > 0.5 else "**NON-BINDING**"
+    conf  = "High confidence" if abs(prob - 0.5) > 0.3 else "Moderate confidence"
+    prob_text = (
+        f"### Prediction Result: {label}\n\n"
+        f"| Metric | Value |\n"
+        f"|--------|-------|\n"
+        f"| Binding probability | **{prob * 100:.1f}%** |\n"
+        f"| Confidence | {conf} |\n"
+        f"| Drug atoms analysed | {n_real_atoms} |\n"
+        f"| Protein residues analysed | {L} |\n"
+    )
+
+    return "Inference complete.", prob_text, heatmap_img, residue_img
+
+
+# ============================================================
+# Global statistics panel
+# ============================================================
+
+_GLOBAL_STATS = _REPORT.get('global_stats', {})
+
+_SIDEBAR_HTML = f"""
+<div style="background:#f8f9fa; border:1px solid #dee2e6; border-radius:6px; padding:16px; font-family:'Georgia',serif;">
+
+  <div style="border-bottom:2px solid #1a4a7a; margin-bottom:12px; padding-bottom:6px;">
+    <strong style="color:#1a4a7a; font-size:0.9rem; text-transform:uppercase; letter-spacing:0.04em;">
+      Performance Metrics
+    </strong><br>
+    <span style="color:#666; font-size:0.78rem;">Davis Kinase Benchmark Dataset</span>
+  </div>
+
+  <table style="width:100%; border-collapse:collapse; font-size:0.82rem; margin-bottom:14px;">
+    <thead>
+      <tr style="background:#1a4a7a; color:white;">
+        <th style="padding:6px 8px; text-align:left; font-weight:600;">Split</th>
+        <th style="padding:6px 8px; text-align:center; font-weight:600;">AUROC</th>
+        <th style="padding:6px 8px; text-align:center; font-weight:600;">AUPRC</th>
+      </tr>
+    </thead>
+    <tbody>
+      <tr style="background:#eef2f7;">
+        <td style="padding:5px 8px;">Random</td>
+        <td style="padding:5px 8px; text-align:center;">0.921</td>
+        <td style="padding:5px 8px; text-align:center;">0.608</td>
+      </tr>
+      <tr>
+        <td style="padding:5px 8px;">Cold-Drug</td>
+        <td style="padding:5px 8px; text-align:center;">0.739</td>
+        <td style="padding:5px 8px; text-align:center;">0.169</td>
+      </tr>
+      <tr style="background:#eef2f7;">
+        <td style="padding:5px 8px;"><strong>Cold-Target</strong></td>
+        <td style="padding:5px 8px; text-align:center;"><strong>0.941</strong></td>
+        <td style="padding:5px 8px; text-align:center;"><strong>0.549</strong></td>
+      </tr>
+    </tbody>
+  </table>
+
+  <div style="border-bottom:1px solid #dee2e6; margin-bottom:10px; padding-bottom:4px;">
+    <strong style="color:#1a4a7a; font-size:0.85rem;">Model Specifications</strong>
+  </div>
+  <table style="width:100%; border-collapse:collapse; font-size:0.81rem; margin-bottom:14px;">
+    <tr><td style="padding:4px 0; color:#555;">Parameters</td>
+        <td style="padding:4px 0; text-align:right;">{_REPORT.get('model_info', {}).get('params', 2_442_083):,}</td></tr>
+    <tr><td style="padding:4px 0; color:#555;">Training samples</td>
+        <td style="padding:4px 0; text-align:right;">{_GLOBAL_STATS.get('n_samples', 1506):,}</td></tr>
+    <tr><td style="padding:4px 0; color:#555;">Attention sparsity</td>
+        <td style="padding:4px 0; text-align:right;">{_GLOBAL_STATS.get('attention_sparsity', 0.992)*100:.1f}%</td></tr>
+    <tr><td style="padding:4px 0; color:#555;">GNN layers</td>
+        <td style="padding:4px 0; text-align:right;">3 × GINE</td></tr>
+    <tr><td style="padding:4px 0; color:#555;">Attention heads</td>
+        <td style="padding:4px 0; text-align:right;">8</td></tr>
+    <tr><td style="padding:4px 0; color:#555;">Hidden dimension</td>
+        <td style="padding:4px 0; text-align:right;">256</td></tr>
+  </table>
+
+  <div style="border-bottom:1px solid #dee2e6; margin-bottom:10px; padding-bottom:4px;">
+    <strong style="color:#1a4a7a; font-size:0.85rem;">Architecture Overview</strong>
+  </div>
+  <pre style="background:#1a2a3a; color:#c8d8e8; padding:10px; border-radius:4px; font-size:0.72rem; line-height:1.5; margin:0; overflow:auto;">
+Drug SMILES
+  → GINE (3 layers, dim=256)
+  → N × atom vectors
+
+Protein sequence
+  → ESM-2 (150M params)
+  → physicochemical (4-dim)
+  → L × residue vectors
+
+Bidirectional cross-attention
+  → N×L interaction map
+  → gated pooling
+  → binding score</pre>
+
+  <div style="margin-top:14px; padding:10px; background:#fff8e1; border-left:3px solid #f9a825; border-radius:0 4px 4px 0; font-size:0.8rem; color:#555; line-height:1.5;">
+    <strong style="color:#e65100;">Reference</strong><br>
+    Wang S, Zhang Q <em>et al.</em> BioInteract: Interpretable DTI Prediction via
+    Residue-Level Cross-Attention with Biological Prior Knowledge.
+    <em>PLOS Computational Biology</em>, 2026.
+  </div>
+</div>
+"""
+
+# ============================================================
+# Custom CSS
+# ============================================================
+
+_CSS = """
+.gradio-container {
+    font-family: 'Georgia', 'Times New Roman', serif !important;
+    max-width: 1440px !important;
+}
+.tab-nav {
+    border-bottom: 2px solid #1a4a7a !important;
+}
+.tab-nav button {
+    font-size: 0.82rem !important;
+    font-weight: 600 !important;
+    letter-spacing: 0.03em !important;
+    text-transform: uppercase !important;
+    color: #555 !important;
+}
+.tab-nav button.selected {
+    color: #1a4a7a !important;
+    border-bottom: 2px solid #1a4a7a !important;
+}
+label span {
+    font-size: 0.82rem !important;
+    font-weight: 600 !important;
+    color: #1a2a3a !important;
+    text-transform: uppercase !important;
+    letter-spacing: 0.03em !important;
+}
+.gr-button-primary {
+    background: #1a4a7a !important;
+    border-color: #1a4a7a !important;
+}
+.gr-button-primary:hover {
+    background: #0d2137 !important;
+}
+footer { display: none !important; }
+"""
+
+# ============================================================
+# Build Gradio UI
+# ============================================================
+
+_HEADER_HTML = """
+<div style="
+  background: linear-gradient(135deg, #0d2137 0%, #1a4a7a 100%);
+  padding: 24px 32px 20px;
+  border-radius: 8px;
+  margin-bottom: 4px;
+">
+  <h1 style="
+    color: #ffffff;
+    margin: 0 0 6px;
+    font-size: 1.65rem;
+    font-weight: 700;
+    font-family: 'Georgia', serif;
+    letter-spacing: -0.3px;
+  ">BioInteract</h1>
+  <p style="
+    color: #b0cee8;
+    margin: 0 0 14px;
+    font-size: 0.95rem;
+    font-style: italic;
+    font-family: 'Georgia', serif;
+    line-height: 1.4;
+  ">
+    Interpretable Drug–Target Interaction Prediction via
+    Residue-Level Cross-Attention with Biological Prior Knowledge
+  </p>
+  <div style="display: flex; gap: 8px; flex-wrap: wrap;">
+    <span style="background:rgba(255,255,255,0.13); color:#d0e8ff;
+                 padding:3px 12px; border-radius:20px; font-size:0.76rem;
+                 font-family:monospace; letter-spacing:0.02em;">
+      GINE Graph Encoder
+    </span>
+    <span style="background:rgba(255,255,255,0.13); color:#d0e8ff;
+                 padding:3px 12px; border-radius:20px; font-size:0.76rem;
+                 font-family:monospace; letter-spacing:0.02em;">
+      ESM-2 (150 M)
+    </span>
+    <span style="background:rgba(255,255,255,0.13); color:#d0e8ff;
+                 padding:3px 12px; border-radius:20px; font-size:0.76rem;
+                 font-family:monospace; letter-spacing:0.02em;">
+      Bidirectional Cross-Attention
+    </span>
+    <span style="background:rgba(200,230,80,0.2); color:#d4f0a0;
+                 padding:3px 12px; border-radius:20px; font-size:0.76rem;
+                 font-family:monospace; letter-spacing:0.02em;">
+      AUROC 0.941 (cold-target)
+    </span>
+  </div>
+</div>
+"""
+
+_ABSTRACT_HTML = """
+<div style="
+  background:#f4f7fb;
+  border-left:4px solid #1a4a7a;
+  padding:12px 18px;
+  margin:8px 0 4px;
+  border-radius:0 5px 5px 0;
+  font-family:'Georgia',serif;
+">
+  <strong style="color:#1a4a7a; font-size:0.78rem; text-transform:uppercase;
+                 letter-spacing:0.06em;">Abstract</strong>
+  <p style="margin:6px 0 0; font-size:0.88rem; color:#2a2a3e; line-height:1.65;">
+    BioInteract couples a pharmacophore-aware Graph Isomorphism Network with Edge features
+    (GINE) for molecular encoding with ESM-2 protein language model representations for
+    residue encoding. A bidirectional cross-attention mechanism generates an interpretable
+    atom–residue interaction map, enabling simultaneous prediction of binding affinity and
+    mechanistic insight into which drug substructures engage specific protein residues.
+    Evaluated on the Davis kinase dataset, the model achieves AUROC&nbsp;0.941 on the
+    cold-target split, demonstrating strong generalisation to unseen protein targets.
+  </p>
+</div>
+"""
+
+with gr.Blocks(
+    title="BioInteract — Interpretable DTI Prediction",
+    theme=gr.themes.Base(
+        primary_hue=gr.themes.colors.blue,
+        neutral_hue=gr.themes.colors.slate,
+        font=[gr.themes.GoogleFont("Source Serif 4"), "Georgia", "serif"],
+    ),
+    css=_CSS,
+) as demo:
+
+    gr.HTML(_HEADER_HTML)
+    gr.HTML(_ABSTRACT_HTML)
+
+    with gr.Row(equal_height=False):
+        # ── Main content area ──────────────────────────────────────────────
+        with gr.Column(scale=3):
+            with gr.Tabs():
+
+                # ── Tab 1: Case Studies ────────────────────────────────────
+                with gr.Tab("Case Studies"):
+                    gr.Markdown(
+                        "Select a pre-computed case study to examine the model's atom–residue "
+                        "interaction map and predicted binding residues for clinically validated "
+                        "drug–target pairs. All cases are drawn from the Davis kinase benchmark."
+                    )
+                    case_dropdown = gr.Dropdown(
+                        choices=list(_FIXED_CASES.keys()),
+                        value=list(_FIXED_CASES.keys())[0],
+                        label="Drug–Target Pair",
+                        interactive=True,
+                    )
+                    case_info_md = gr.Markdown(
+                        container=True,
+                        min_height=80,
+                    )
+
+                    with gr.Row():
+                        case_heatmap = gr.Image(
+                            label="Atom–Residue Cross-Attention Heatmap",
+                            type='pil',
+                            height=420,
+                        )
+                        with gr.Column():
+                            case_residues = gr.Image(
+                                label="Top Predicted Binding Residues",
+                                type='pil',
+                                height=280,
+                            )
+                            case_fg = gr.Image(
+                                label="Pharmacophore Group Importance (Grad-CAM)",
+                                type='pil',
+                                height=240,
+                            )
+
+                    gr.Markdown(
+                        "_Figure caption:_ The heatmap encodes normalised cross-attention "
+                        "weights between each drug atom (rows) and protein residue (columns). "
+                        "Darker cells indicate stronger predicted interactions. "
+                        "The bar chart ranks residues by aggregated attention score."
+                    )
+
+                    case_dropdown.change(
+                        fn=show_case_study,
+                        inputs=case_dropdown,
+                        outputs=[case_heatmap, case_residues, case_fg, case_info_md],
+                    )
+                    demo.load(
+                        fn=lambda: show_case_study(list(_FIXED_CASES.keys())[0]),
+                        inputs=[],
+                        outputs=[case_heatmap, case_residues, case_fg, case_info_md],
+                    )
+
+                # ── Tab 2: Custom Prediction ───────────────────────────────
+                with gr.Tab("Custom Prediction"):
+                    gr.Markdown(
+                        "Provide a drug SMILES string and a protein amino acid sequence "
+                        "to obtain a binding prediction with an interpretable cross-attention map.\n\n"
+                        "> **Note:** ESM-2 (150 M parameters) initialises on the first request; "
+                        "please allow 1–3 minutes on CPU. Sequences exceeding 512 residues are "
+                        "automatically truncated to the first 512 positions."
+                    )
+
+                    with gr.Row():
+                        smiles_box = gr.Textbox(
+                            label="Drug SMILES",
+                            placeholder=(
+                                "e.g. Cc1ccc(NC(=O)c2ccc(CN3CCN(C)CC3)cc2)cc1"
+                                "Nc1nccc(-c2cccnc2)n1  (Imatinib)"
+                            ),
+                            lines=2,
+                        )
+                    sequence_box = gr.Textbox(
+                        label="Protein Amino Acid Sequence (single-letter code)",
+                        placeholder="e.g. MGPSENDPNLFVALYDFVASGDNTLS…",
+                        lines=4,
+                        max_lines=8,
+                    )
+
+                    with gr.Row():
+                        example_btn = gr.Button(
+                            "Load Imatinib / ABL1 Example",
+                            variant="secondary",
+                            size="sm",
+                        )
+                        predict_btn = gr.Button(
+                            "Run Prediction",
+                            variant="primary",
+                            size="lg",
+                        )
+
+                    status_box = gr.Textbox(
+                        label="Status",
+                        interactive=False,
+                        lines=1,
+                        placeholder="Awaiting input …",
+                    )
+                    prob_md = gr.Markdown(min_height=80)
+
+                    with gr.Row():
+                        pred_heatmap = gr.Image(
+                            label="Atom–Residue Cross-Attention Heatmap",
+                            type='pil',
+                            height=420,
+                        )
+                        pred_residues = gr.Image(
+                            label="Top 10 Predicted Binding Residues",
+                            type='pil',
+                            height=320,
+                        )
+
+                    gr.Markdown(
+                        "_Interpretation:_ Rows correspond to heavy atoms of the drug molecule; "
+                        "columns to protein residues. High-intensity cells indicate residues "
+                        "predicted to form key contacts with the respective drug atoms. "
+                        "The bar chart aggregates attention over all atoms for each residue."
+                    )
+
+                    example_btn.click(
+                        fn=lambda: (_EXAMPLE_SMILES, _EXAMPLE_SEQUENCE),
+                        inputs=[],
+                        outputs=[smiles_box, sequence_box],
+                    )
+                    predict_btn.click(
+                        fn=run_prediction,
+                        inputs=[smiles_box, sequence_box],
+                        outputs=[status_box, prob_md, pred_heatmap, pred_residues],
+                    )
+
+        # ── Sidebar ────────────────────────────────────────────────────────
+        with gr.Column(scale=1, min_width=260):
+            gr.HTML(_SIDEBAR_HTML)
+
+
+if __name__ == "__main__":
+    demo.launch(server_name="0.0.0.0", server_port=7860)

requirements.txt

@@ -1,9 +1,15 @@
-audioop-lts
-numpy
-matplotlib
-seaborn
-pandas
-scipy
-scikit-learn
-PyYAML
-Pillow
+--extra-index-url https://download.pytorch.org/whl/cpu
+torch==2.4.0
+torch-geometric==2.6.1
+rdkit
+transformers==5.6.1
+numpy==1.26.4
+pandas==2.0.3
+scipy==1.13.1
+scikit-learn==1.2.1
+PyYAML==6.0.2
+matplotlib==3.8.4
+seaborn==0.12.2
+tqdm==4.67.3
+Pillow==9.4.0
+gradio==5.49.1