{"metadata":{"kernelspec":{"display_name":"Python 3 (ipykernel)","language":"python","name":"python3"},"language_info":{"codemirror_mode":{"name":"ipython","version":3},"file_extension":".py","mimetype":"text/x-python","name":"python","nbconvert_exporter":"python","pygments_lexer":"ipython3","version":"3.12.12"},"kaggle":{"accelerator":"gpu","dataSources":[{"sourceType":"competition","sourceId":118765,"databundleVersionId":15231210},{"sourceType":"datasetVersion","sourceId":14604295,"datasetId":9328538,"databundleVersionId":15440074},{"sourceType":"datasetVersion","sourceId":14874339,"datasetId":9502242,"databundleVersionId":15736806},{"sourceType":"datasetVersion","sourceId":10855324,"datasetId":6742586,"databundleVersionId":11219268}],"dockerImageVersionId":31260,"isInternetEnabled":false,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"Reference\n\nhttps://www.kaggle.com/code/qiweiyin/protenix-v1-inference-2026\n\nhttps://www.kaggle.com/code/nihilisticneuralnet/0-409-stanford-rna-folding-2-protenix-template\n\nhttps://www.kaggle.com/code/alexxanderlarko/protenix-v1","metadata":{"execution":{"iopub.execute_input":"2026-02-16T19:38:36.011993Z","iopub.status.busy":"2026-02-16T19:38:36.011656Z","iopub.status.idle":"2026-02-16T19:38:36.016065Z","shell.execute_reply":"2026-02-16T19:38:36.015433Z","shell.execute_reply.started":"2026-02-16T19:38:36.011965Z"}}},{"cell_type":"markdown","source":"| version | description             | LB    \n|---------|-------------------------|-------\n| 3       | protenix+TBM            | 0.408 \n| 4       | pure protenix           | 0.249 \n| 5       | USE_MSA and USE_RNA_MSA | TBC   \n| 6       | USE_MSA                 | TBC\n| 7       | USE_RNA_MSA             | TBC","metadata":{}},{"cell_type":"code","source":"# !pip install /kaggle/input/datasets/ogurtsov/biopython/biopython-1.85-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\nimport sys\nimport pandas as pd\n\n# ── Local vs Kaggle mode ─────────────────────────────────────────────────────\n# On Kaggle competition rerun, KAGGLE_IS_COMPETITION_RERUN is set to a truthy value.\n# When running locally we do NOT exit — instead we cap the test set to a small\n# number of samples so the notebook finishes quickly.\n\nIS_KAGGLE = bool(os.environ.get(\"KAGGLE_IS_COMPETITION_RERUN\", \"\"))\n\n# How many test samples to use when running locally\nLOCAL_N_SAMPLES = 2\n\nif IS_KAGGLE:\n    print(\"Running in KAGGLE COMPETITION mode — all test targets will be processed.\")\nelse:\n    print(f\"Running in LOCAL mode — only the first {LOCAL_N_SAMPLES} test targets \"\n          f\"will be processed to save time.\")\n","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import gc\nimport json\nimport os\nimport time\n\nos.environ[\"LAYERNORM_TYPE\"] = \"torch\"\nos.environ.setdefault(\"RNA_MSA_DEPTH_LIMIT\", \"512\")\n\nimport sys\nfrom pathlib import Path\n\nimport numpy as np\nimport pandas as pd\nimport torch\nfrom Bio.Align import PairwiseAligner\nfrom tqdm import tqdm","metadata":{},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def get_c1_mask(data: dict, atom_array) -> torch.Tensor:\n    # 1. Try atom_array attributes first\n    if atom_array is not None:\n        try:\n            if hasattr(atom_array, \"centre_atom_mask\"):\n                m = atom_array.centre_atom_mask == 1\n                if hasattr(atom_array, \"is_rna\"):\n                    m = m & atom_array.is_rna\n                return torch.from_numpy(m).bool()\n            \n            if hasattr(atom_array, \"atom_name\"):\n                base = atom_array.atom_name == \"C1'\"\n                if hasattr(atom_array, \"is_rna\"):\n                    base = base & atom_array.is_rna\n                return torch.from_numpy(base).bool()\n        except Exception:\n            pass\n\n    # 2. Fallback to feature dict\n    f = data[\"input_feature_dict\"]\n    \n    if \"centre_atom_mask\" in f:\n        return (f[\"centre_atom_mask\"] == 1).bool()\n    if \"center_atom_mask\" in f:\n        return (f[\"center_atom_mask\"] == 1).bool()\n        \n    # Heuristic fallback: check which index gives us roughly N_token atoms\n    n_tokens = data.get(\"N_token\", torch.tensor(0)).item()\n    mask11 = (f[\"atom_to_tokatom_idx\"] == 11).bool()\n    mask12 = (f[\"atom_to_tokatom_idx\"] == 12).bool()\n    \n    c11 = mask11.sum().item()\n    c12 = mask12.sum().item()\n    \n    # Return the one closer to N_tokens (likely one per residue)\n    if abs(c11 - n_tokens) < abs(c12 - n_tokens):\n        return mask11\n    else:\n        return mask12\n","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"\n# ─────────────── Paths & Constants ───────────────────────────────────────────\nDATA_BASE              = \"/kaggle/input/stanford-rna-3d-folding-2\"\nDEFAULT_TEST_CSV       = f\"{DATA_BASE}/test_sequences.csv\"\nDEFAULT_TRAIN_CSV      = f\"{DATA_BASE}/train_sequences.csv\"\nDEFAULT_TRAIN_LBLS     = f\"{DATA_BASE}/train_labels.csv\"\nDEFAULT_VAL_CSV        = f\"{DATA_BASE}/validation_sequences.csv\"\nDEFAULT_VAL_LBLS       = f\"{DATA_BASE}/validation_labels.csv\"\nDEFAULT_OUTPUT         = \"/kaggle/working/submission.csv\"\n\nDEFAULT_CODE_DIR = (\n    \"/kaggle/input/datasets/qiweiyin/protenix-v1-adjusted\"\n    \"/Protenix-v1-adjust-v2/Protenix-v1-adjust-v2/Protenix-v1\"\n)\nDEFAULT_ROOT_DIR = DEFAULT_CODE_DIR\n\nMODEL_NAME    = \"protenix_base_20250630_v1.0.0\"\nN_SAMPLE      = 5\nSEED          = 42\nMAX_SEQ_LEN   = int(os.environ.get(\"MAX_SEQ_LEN\",   \"512\"))\nCHUNK_OVERLAP = int(os.environ.get(\"CHUNK_OVERLAP\",  \"64\"))\n\n# TBM quality thresholds — sequences below these get routed to Protenix\nMIN_SIMILARITY       = float(os.environ.get(\"MIN_SIMILARITY\",       \"0.0\"))\nMIN_PERCENT_IDENTITY = float(os.environ.get(\"MIN_PERCENT_IDENTITY\", \"50.0\"))\n\n# Set False to skip Protenix and use de-novo fallback instead\nUSE_PROTENIX = True\n\n\ndef parse_bool(value: str, default: bool = False) -> str:\n    v = str(value).strip().lower()\n    if v in {\"1\", \"true\", \"t\", \"yes\", \"y\", \"on\"}:\n        return \"true\"\n    if v in {\"0\", \"false\", \"f\", \"no\", \"n\", \"off\"}:\n        return \"false\"\n    return \"true\" if default else \"false\"\n\n\nUSE_MSA      = parse_bool(os.environ.get(\"USE_MSA\",      \"false\"))\nUSE_TEMPLATE = parse_bool(os.environ.get(\"USE_TEMPLATE\", \"false\"))\nUSE_RNA_MSA  = parse_bool(os.environ.get(\"USE_RNA_MSA\",  \"true\"))\n\nMODEL_N_SAMPLE = int(os.environ.get(\"MODEL_N_SAMPLE\", str(N_SAMPLE)))\n\n\n# ─────────────── General Utilities ───────────────────────────────────────────\ndef seed_everything(seed: int) -> None:\n    os.environ[\"CUBLAS_WORKSPACE_CONFIG\"] = \":4096:8\"\n    torch.manual_seed(seed)\n    torch.cuda.manual_seed(seed)\n    torch.cuda.manual_seed_all(seed)\n    np.random.seed(seed)\n    torch.backends.cudnn.benchmark = False\n    torch.backends.cudnn.deterministic = True\n    torch.backends.cudnn.enabled = True\n    torch.use_deterministic_algorithms(True)\n\n\ndef resolve_paths():\n    test_csv   = os.environ.get(\"TEST_CSV\",           DEFAULT_TEST_CSV)\n    output_csv = os.environ.get(\"SUBMISSION_CSV\",     DEFAULT_OUTPUT)\n    code_dir   = os.environ.get(\"PROTENIX_CODE_DIR\",  DEFAULT_CODE_DIR)\n    root_dir   = os.environ.get(\"PROTENIX_ROOT_DIR\",  DEFAULT_ROOT_DIR)\n    return test_csv, output_csv, code_dir, root_dir\n\n\ndef ensure_required_files(root_dir: str) -> None:\n    for p, name in [\n        (Path(root_dir) / \"checkpoint\" / f\"{MODEL_NAME}.pt\",          \"checkpoint\"),\n        (Path(root_dir) / \"common\" / \"components.cif\",                \"CCD file\"),\n        (Path(root_dir) / \"common\" / \"components.cif.rdkit_mol.pkl\",  \"CCD cache\"),\n    ]:\n        if not p.exists():\n            raise FileNotFoundError(f\"Missing {name}: {p}\")\n\n\n# ─────────────── Protenix Input / Config Helpers ─────────────────────────────\ndef build_input_json(df: pd.DataFrame, json_path: str) -> None:\n    data = [\n        {\n            \"name\": row[\"target_id\"],\n            \"covalent_bonds\": [],\n            \"sequences\": [{\"rnaSequence\": {\"sequence\": row[\"sequence\"], \"count\": 1}}],\n        }\n        for _, row in df.iterrows()\n    ]\n    with open(json_path, \"w\", encoding=\"utf-8\") as f:\n        json.dump(data, f)\n\n\ndef build_configs(input_json_path: str, dump_dir: str, model_name: str):\n    from configs.configs_base import configs as configs_base\n    from configs.configs_data import data_configs\n    from configs.configs_inference import inference_configs\n    from configs.configs_model_type import model_configs\n    from protenix.config.config import parse_configs\n\n    base = {**configs_base, **{\"data\": data_configs}, **inference_configs}\n\n    def deep_update(t, p):\n        for k, v in p.items():\n            if isinstance(v, dict) and k in t and isinstance(t[k], dict):\n                deep_update(t[k], v)\n            else:\n                t[k] = v\n\n    deep_update(base, model_configs[model_name])\n    arg_str = \" \".join([\n        f\"--model_name {model_name}\",\n        f\"--input_json_path {input_json_path}\",\n        f\"--dump_dir {dump_dir}\",\n        f\"--use_msa {USE_MSA}\",\n        f\"--use_template {USE_TEMPLATE}\",\n        f\"--use_rna_msa {USE_RNA_MSA}\",\n        f\"--sample_diffusion.N_sample {MODEL_N_SAMPLE}\",\n        f\"--seeds {SEED}\",\n    ])\n    return parse_configs(configs=base, arg_str=arg_str, fill_required_with_null=True)\n\n\ndef get_c1_mask(data: dict, atom_array) -> torch.Tensor:\n    # 1. Try atom_array attributes first\n    if atom_array is not None:\n        try:\n            if hasattr(atom_array, \"centre_atom_mask\"):\n                m = atom_array.centre_atom_mask == 1\n                if hasattr(atom_array, \"is_rna\"):\n                    m = m & atom_array.is_rna\n                return torch.from_numpy(m).bool()\n            \n            if hasattr(atom_array, \"atom_name\"):\n                base = atom_array.atom_name == \"C1'\"\n                if hasattr(atom_array, \"is_rna\"):\n                    base = base & atom_array.is_rna\n                return torch.from_numpy(base).bool()\n        except Exception:\n            pass\n\n    # 2. Fallback to feature dict\n    f = data[\"input_feature_dict\"]\n    \n    # CASE A: center_atom_mask exists\n    if \"center_atom_mask\" in f:\n        return (f[\"center_atom_mask\"] == 1).bool()\n    if \"centre_atom_mask\" in f:\n        return (f[\"centre_atom_mask\"] == 1).bool()\n        \n    # CASE B: Use atom_name\n    if \"atom_name\" in f:\n        # Check against \"C1'\" (byte encoded or string?)\n        # For now assume typical behavior is center_atom_mask is present.\n        pass\n\n    # CASE C: atom_to_tokatom_idx fallback\n    # The index for C1' is typically 11 or 12 depending on featurizer.\n    # Let's try to match exactly C1' if possible.\n    # But usually 'centre_atom_mask' should be there.\n    \n    # If we fall through, assume standard mask\n    return (f[\"atom_to_tokatom_idx\"] == 11).bool()\n\n\ndef get_feature_c1_mask(data: dict) -> torch.Tensor:\n    f = data[\"input_feature_dict\"]\n    if \"centre_atom_mask\" in f:\n        return f[\"centre_atom_mask\"].long() == 1\n    return f[\"atom_to_tokatom_idx\"].long() == 12\n\n\ndef coords_to_rows(target_id: str, seq: str, coords: np.ndarray) -> list:\n    \"\"\"coords shape: (N_SAMPLE, seq_len, 3)\"\"\"\n    rows = []\n    for i in range(len(seq)):\n        row = {\"ID\": f\"{target_id}_{i + 1}\", \"resname\": seq[i], \"resid\": i + 1}\n        for s in range(N_SAMPLE):\n            if s < coords.shape[0] and i < coords.shape[1]:\n                x, y, z = coords[s, i]\n            else:\n                x, y, z = 0.0, 0.0, 0.0\n            row[f\"x_{s + 1}\"] = float(x)\n            row[f\"y_{s + 1}\"] = float(y)\n            row[f\"z_{s + 1}\"] = float(z)\n        rows.append(row)\n    return rows\n\n\ndef pad_samples(coords: np.ndarray, n: int) -> np.ndarray:\n    if coords.shape[0] >= n:\n        return coords[:n]\n    if coords.shape[0] == 0:\n        return np.zeros((n, coords.shape[1], 3), dtype=coords.dtype)\n    extra = np.repeat(coords[:1], n - coords.shape[0], axis=0)\n    return np.concatenate([coords, extra], axis=0)\n\n\n# ─────────────── TBM Core Functions ──────────────────────────────────────────\ndef _make_aligner() -> PairwiseAligner:\n    al = PairwiseAligner()\n    al.mode                           = \"global\"\n    al.match_score                    = 2\n    al.mismatch_score                 = -1.5\n    al.open_gap_score                 = -8\n    al.extend_gap_score               = -0.4\n    al.query_left_open_gap_score      = -8\n    al.query_left_extend_gap_score    = -0.4\n    al.query_right_open_gap_score     = -8\n    al.query_right_extend_gap_score   = -0.4\n    al.target_left_open_gap_score     = -8\n    al.target_left_extend_gap_score   = -0.4\n    al.target_right_open_gap_score    = -8\n    al.target_right_extend_gap_score  = -0.4\n    return al\n\n\n_aligner = _make_aligner()\n\n\ndef parse_stoichiometry(stoich: str) -> list:\n    if pd.isna(stoich) or str(stoich).strip() == \"\":\n        return []\n    return [(ch.strip(), int(cnt)) for part in str(stoich).split(\";\")\n            for ch, cnt in [part.split(\":\")]]\n\n\ndef parse_fasta(fasta_content: str) -> dict:\n    out, cur, parts = {}, None, []\n    for line in str(fasta_content).splitlines():\n        line = line.strip()\n        if not line:\n            continue\n        if line.startswith(\">\"):\n            if cur is not None:\n                out[cur] = \"\".join(parts)\n            cur = line[1:].split()[0]\n            parts = []\n        else:\n            parts.append(line.replace(\" \", \"\"))\n    if cur is not None:\n        out[cur] = \"\".join(parts)\n    return out\n\n\ndef get_chain_segments(row) -> list:\n    seq    = row[\"sequence\"]\n    stoich = row.get(\"stoichiometry\", \"\")\n    all_sq = row.get(\"all_sequences\", \"\")\n    if (pd.isna(stoich) or pd.isna(all_sq)\n            or str(stoich).strip() == \"\" or str(all_sq).strip() == \"\"):\n        return [(0, len(seq))]\n    try:\n        chain_dict = parse_fasta(all_sq)\n        order = parse_stoichiometry(stoich)\n        segs, pos = [], 0\n        for ch, cnt in order:\n            base = chain_dict.get(ch)\n            if base is None:\n                return [(0, len(seq))]\n            for _ in range(cnt):\n                segs.append((pos, pos + len(base)))\n                pos += len(base)\n        return segs if pos == len(seq) else [(0, len(seq))]\n    except Exception:\n        return [(0, len(seq))]\n\n\ndef build_segments_map(df: pd.DataFrame) -> tuple:\n    seg_map, stoich_map = {}, {}\n    for _, r in df.iterrows():\n        tid               = r[\"target_id\"]\n        seg_map[tid]      = get_chain_segments(r)\n        raw_s             = r.get(\"stoichiometry\", \"\")\n        stoich_map[tid]   = \"\" if pd.isna(raw_s) else str(raw_s)\n    return seg_map, stoich_map\n\n\ndef process_labels(labels_df: pd.DataFrame) -> dict:\n    coords = {}\n    prefixes = labels_df[\"ID\"].str.rsplit(\"_\", n=1).str[0]\n    for prefix, grp in labels_df.groupby(prefixes):\n        coords[prefix] = grp.sort_values(\"resid\")[[\"x_1\", \"y_1\", \"z_1\"]].values\n    return coords\n\n\ndef _build_aligned_strings(query_seq, template_seq, alignment):\n    q_segs, t_segs = alignment.aligned\n    aq, at, qi, ti = [], [], 0, 0\n    for (qs, qe), (ts, te) in zip(q_segs, t_segs):\n        while qi < qs: aq.append(query_seq[qi]);    at.append(\"-\");              qi += 1\n        while ti < ts: aq.append(\"-\");              at.append(template_seq[ti]); ti += 1\n        for qp, tp in zip(range(qs, qe), range(ts, te)):\n            aq.append(query_seq[qp]); at.append(template_seq[tp])\n        qi, ti = qe, te\n    while qi < len(query_seq):    aq.append(query_seq[qi]);    at.append(\"-\");              qi += 1\n    while ti < len(template_seq): aq.append(\"-\");              at.append(template_seq[ti]); ti += 1\n    return \"\".join(aq), \"\".join(at)\n\n\ndef find_similar_sequences_detailed(query_seq, train_seqs_df, train_coords_dict, top_n=30):\n    results = []\n    for _, row in train_seqs_df.iterrows():\n        tid, tseq = row[\"target_id\"], row[\"sequence\"]\n        if tid not in train_coords_dict:\n            continue\n        if abs(len(tseq) - len(query_seq)) / max(len(tseq), len(query_seq)) > 0.3:\n            continue\n        aln       = next(iter(_aligner.align(query_seq, tseq)))\n        norm_s    = aln.score / (2 * min(len(query_seq), len(tseq)))\n        identical = sum(\n            1 for (qs, qe), (ts, te) in zip(*aln.aligned)\n            for qp, tp in zip(range(qs, qe), range(ts, te))\n            if query_seq[qp] == tseq[tp]\n        )\n        pct_id = 100 * identical / len(query_seq)\n        aq, at = _build_aligned_strings(query_seq, tseq, aln)\n        results.append((tid, tseq, norm_s, train_coords_dict[tid], pct_id, aq, at))\n    results.sort(key=lambda x: x[2], reverse=True)\n    return results[:top_n]\n\n\ndef adapt_template_to_query(query_seq, template_seq, template_coords) -> np.ndarray:\n    aln        = next(iter(_aligner.align(query_seq, template_seq)))\n    new_coords = np.full((len(query_seq), 3), np.nan)\n    for (qs, qe), (ts, te) in zip(*aln.aligned):\n        chunk = template_coords[ts:te]\n        if len(chunk) == (qe - qs):\n            new_coords[qs:qe] = chunk\n    for i in range(len(new_coords)):\n        if np.isnan(new_coords[i, 0]):\n            pv = next((j for j in range(i - 1, -1, -1) if not np.isnan(new_coords[j, 0])), -1)\n            nv = next((j for j in range(i + 1, len(new_coords)) if not np.isnan(new_coords[j, 0])), -1)\n            if pv >= 0 and nv >= 0:\n                w = (i - pv) / (nv - pv)\n                new_coords[i] = (1 - w) * new_coords[pv] + w * new_coords[nv]\n            elif pv >= 0:\n                new_coords[i] = new_coords[pv] + [3, 0, 0]\n            elif nv >= 0:\n                new_coords[i] = new_coords[nv] + [3, 0, 0]\n            else:\n                new_coords[i] = [i * 3, 0, 0]\n    return np.nan_to_num(new_coords)\n\n\ndef adaptive_rna_constraints(coords, target_id, segments_map, confidence=1.0, passes=2) -> np.ndarray:\n    X        = coords.copy()\n    segments = segments_map.get(target_id, [(0, len(X))])\n    strength = max(0.75 * (1.0 - min(confidence, 0.97)), 0.02)\n    for _ in range(passes):\n        for s, e in segments:\n            C = X[s:e]; L = e - s\n            if L < 3:\n                continue\n            # bond i–i+1  ~5.95 Å\n            d    = C[1:] - C[:-1]; dist = np.linalg.norm(d, axis=1) + 1e-6\n            adj  = d * ((5.95 - dist) / dist)[:, None] * (0.22 * strength)\n            C[:-1] -= adj; C[1:] += adj\n            # soft i–i+2  ~10.2 Å\n            d2   = C[2:] - C[:-2]; d2n = np.linalg.norm(d2, axis=1) + 1e-6\n            adj2 = d2 * ((10.2 - d2n) / d2n)[:, None] * (0.10 * strength)\n            C[:-2] -= adj2; C[2:] += adj2\n            # Laplacian smoothing\n            C[1:-1] += (0.06 * strength) * (0.5 * (C[:-2] + C[2:]) - C[1:-1])\n            # self-avoidance\n            if L >= 25:\n                idx  = np.linspace(0, L - 1, min(L, 160)).astype(int) if L > 220 else np.arange(L)\n                P    = C[idx]; diff = P[:, None, :] - P[None, :, :]\n                dm   = np.linalg.norm(diff, axis=2) + 1e-6\n                sep  = np.abs(idx[:, None] - idx[None, :])\n                mask = (sep > 2) & (dm < 3.2)\n                if np.any(mask):\n                    vec = (diff * ((3.2 - dm) / dm)[:, :, None] * mask[:, :, None]).sum(axis=1)\n                    C[idx] += (0.015 * strength) * vec\n            X[s:e] = C\n    return X\n\n\ndef _rotmat(axis, ang):\n    a = np.asarray(axis, float); a /= np.linalg.norm(a) + 1e-12\n    x, y, z = a; c, s = np.cos(ang), np.sin(ang); CC = 1 - c\n    return np.array([[c+x*x*CC, x*y*CC-z*s, x*z*CC+y*s],\n                     [y*x*CC+z*s, c+y*y*CC, y*z*CC-x*s],\n                     [z*x*CC-y*s, z*y*CC+x*s, c+z*z*CC]])\n\n\ndef apply_hinge(coords, seg, rng, deg=22):\n    s, e = seg; L = e - s\n    if L < 30: return coords\n    pivot = s + int(rng.integers(10, L - 10))\n    R = _rotmat(rng.normal(size=3), np.deg2rad(float(rng.uniform(-deg, deg))))\n    X = coords.copy(); p0 = X[pivot].copy()\n    X[pivot+1:e] = (X[pivot+1:e] - p0) @ R.T + p0\n    return X\n\n\ndef jitter_chains(coords, segs, rng, deg=12, trans=1.5):\n    X = coords.copy(); gc_ = X.mean(0, keepdims=True)\n    for s, e in segs:\n        R     = _rotmat(rng.normal(size=3), np.deg2rad(float(rng.uniform(-deg, deg))))\n        shift = rng.normal(size=3); shift = shift / (np.linalg.norm(shift) + 1e-12) * float(rng.uniform(0, trans))\n        c     = X[s:e].mean(0, keepdims=True)\n        X[s:e] = (X[s:e] - c) @ R.T + c + shift\n    X -= X.mean(0, keepdims=True) - gc_\n    return X\n\n\ndef smooth_wiggle(coords, segs, rng, amp=0.8):\n    X = coords.copy()\n    for s, e in segs:\n        L = e - s\n        if L < 20: continue\n        ctrl = np.linspace(0, L - 1, 6); disp = rng.normal(0, amp, (6, 3)); t = np.arange(L)\n        X[s:e] += np.vstack([np.interp(t, ctrl, disp[:, k]) for k in range(3)]).T\n    return X\n\n\ndef generate_rna_structure(sequence: str, seed=None) -> np.ndarray:\n    \"\"\"Idealized A-form RNA helix — last-resort de-novo fallback.\"\"\"\n    if seed is not None:\n        np.random.seed(seed)\n    n = len(sequence); coords = np.zeros((n, 3))\n    for i in range(n):\n        ang = i * 0.6\n        coords[i] = [10.0 * np.cos(ang), 10.0 * np.sin(ang), i * 2.5]\n    return coords\n\n\n# ─────────────── TBM Phase ───────────────────────────────────────────────────\ndef tbm_phase(test_df, train_seqs_df, train_coords_dict, segments_map):\n    \"\"\"\n    Phase 1 — Template-Based Modeling.\n\n    Returns\n    -------\n    template_predictions : {target_id: [np.ndarray(seq_len, 3), ...]}\n        0 to N_SAMPLE predictions per target, from real templates.\n    protenix_queue : {target_id: (n_needed, full_sequence)}\n        Targets that still need more predictions.\n    \"\"\"\n    print(f\"\\n{'='*60}\")\n    print(f\"PHASE 1: Template-Based Modeling\")\n    print(f\"  MIN_SIMILARITY = {MIN_SIMILARITY}  |  MIN_PCT_IDENTITY = {MIN_PERCENT_IDENTITY}\")\n    print(f\"{'='*60}\")\n    t0 = time.time()\n\n    template_predictions: dict = {}\n    protenix_queue:       dict = {}\n\n    for _, row in test_df.iterrows():\n        tid = row[\"target_id\"]\n        seq = row[\"sequence\"]\n        segs = segments_map.get(tid, [(0, len(seq))])\n\n        similar = find_similar_sequences_detailed(seq, train_seqs_df, train_coords_dict, top_n=30)\n        preds   = []\n        used    = set()\n\n        for i, (tmpl_id, tmpl_seq, sim, tmpl_coords, pct_id, _, _) in enumerate(similar):\n            if len(preds) >= N_SAMPLE:\n                break\n            if sim < MIN_SIMILARITY or pct_id < MIN_PERCENT_IDENTITY:\n                break           # list is sorted by sim, so no point continuing\n            if tmpl_id in used:\n                continue\n\n            rng     = np.random.default_rng((row.name * 10000000000 + i * 10007) % (2**32))\n            adapted = adapt_template_to_query(seq, tmpl_seq, tmpl_coords)\n\n            # Diversity transforms (same strategy as the 0-409 TBM notebook)\n            slot = len(preds)\n            if slot == 0:\n                X = adapted\n            elif slot == 1:\n                X = adapted + rng.normal(0, max(0.01, (0.40 - sim) * 0.06), adapted.shape)\n            elif slot == 2:\n                longest = max(segs, key=lambda se: se[1] - se[0])\n                X = apply_hinge(adapted, longest, rng)\n            elif slot == 3:\n                X = jitter_chains(adapted, segs, rng)\n            else:\n                X = smooth_wiggle(adapted, segs, rng)\n\n            refined = adaptive_rna_constraints(X, tid, segments_map, confidence=sim)\n            preds.append(refined)\n            used.add(tmpl_id)\n\n        template_predictions[tid] = preds\n        n_needed = N_SAMPLE - len(preds)\n        if n_needed > 0:\n            protenix_queue[tid] = (n_needed, seq)\n            print(f\"  {tid} ({len(seq)} nt): {len(preds)} TBM → need {n_needed} from Protenix\")\n        else:\n            print(f\"  {tid} ({len(seq)} nt): all {N_SAMPLE} from TBM ✓\")\n\n    elapsed = time.time() - t0\n    n_full  = len(test_df) - len(protenix_queue)\n    print(f\"\\nPhase 1 done in {elapsed:.1f}s\")\n    print(f\"  Fully covered by TBM : {n_full}\")\n    print(f\"  Need Protenix        : {len(protenix_queue)}\")\n    return template_predictions, protenix_queue\n\n\n# ─────────────── Main ────────────────────────────────────────────────────────\ndef main() -> None:\n    test_csv, output_csv, code_dir, root_dir = resolve_paths()\n\n    if not os.path.isdir(code_dir):\n        raise FileNotFoundError(\n            f\"Missing PROTENIX_CODE_DIR: {code_dir}. \"\n            \"Set PROTENIX_CODE_DIR to the repo path.\"\n        )\n\n    os.environ[\"PROTENIX_ROOT_DIR\"] = root_dir\n    sys.path.append(code_dir)\n    ensure_required_files(root_dir)\n    seed_everything(SEED)\n\n    # ── Load test data ──────────────────────────────────────────────────────\n    test_df_full = pd.read_csv(test_csv)\n    test_df      = (test_df_full.head(LOCAL_N_SAMPLES) if not IS_KAGGLE\n                    else test_df_full).reset_index(drop=True)\n    print(f\"Test targets : {len(test_df)}\"\n          + (\" (LOCAL MODE)\" if not IS_KAGGLE else \"\"))\n\n    seq_by_id = dict(zip(test_df[\"target_id\"], test_df[\"sequence\"]))\n\n    # Truncated copy for Protenix (Protenix has token limits)\n    test_df_trunc = test_df.copy()\n    test_df_trunc[\"sequence\"] = test_df_trunc[\"sequence\"].str[:MAX_SEQ_LEN]\n\n    # ── Load training data for TBM ──────────────────────────────────────────\n    print(\"\\nLoading training data for TBM …\")\n    train_seqs   = pd.read_csv(DEFAULT_TRAIN_CSV)\n    val_seqs     = pd.read_csv(DEFAULT_VAL_CSV)\n    train_labels = pd.read_csv(DEFAULT_TRAIN_LBLS)\n    val_labels   = pd.read_csv(DEFAULT_VAL_LBLS)\n\n    combined_seqs   = pd.concat([train_seqs,   val_seqs],    ignore_index=True)\n    combined_labels = pd.concat([train_labels, val_labels],  ignore_index=True)\n    train_coords    = process_labels(combined_labels)\n    segments_map, _ = build_segments_map(test_df)\n\n    print(f\"Template pool: {len(combined_seqs)} sequences, {len(train_coords)} structures\")\n\n    # ─── PHASE 1: TBM ──────────────────────────────────────────────────────\n    template_preds, protenix_queue = tbm_phase(\n        test_df, combined_seqs, train_coords, segments_map\n    )\n\n    # ─── PHASE 2: Protenix (only for targets that need extra predictions) ──\n    protenix_preds: dict = {}   # target_id -> np.ndarray (n_needed, seq_len, 3)\n\n    if protenix_queue and USE_PROTENIX:\n        print(f\"\\n{'='*60}\")\n        print(f\"PHASE 2: Protenix for {len(protenix_queue)} targets\")\n        print(f\"{'='*60}\")\n\n        work_dir = Path(\"/kaggle/working\")\n        work_dir.mkdir(parents=True, exist_ok=True)\n\n        # Build input JSON only for queued targets\n        queue_df = (test_df_trunc[test_df_trunc[\"target_id\"].isin(protenix_queue)]\n                    .reset_index(drop=True))\n        input_json_path = str(work_dir / \"protenix_queue_input.json\")\n        build_input_json(queue_df, input_json_path)\n\n        from protenix.data.inference.infer_dataloader import InferenceDataset\n        from runner.inference import (InferenceRunner,\n                                      update_gpu_compatible_configs,\n                                      update_inference_configs)\n\n        configs = build_configs(input_json_path, str(work_dir / \"outputs\"), MODEL_NAME)\n        configs = update_gpu_compatible_configs(configs)\n        runner  = InferenceRunner(configs)\n        dataset = InferenceDataset(configs)\n\n        for i in tqdm(range(len(dataset)), desc=\"Protenix\"):\n            data, atom_array, error_message = dataset[i]\n            target_id = data.get(\"sample_name\", f\"sample_{i}\")\n\n            if target_id not in protenix_queue:\n                continue\n\n            n_needed, full_seq = protenix_queue[target_id]\n\n            if error_message:\n                print(f\"  {target_id}: data error — {error_message}\")\n                protenix_preds[target_id] = None\n                del data, atom_array, error_message\n                gc.collect(); torch.cuda.empty_cache(); gc.collect()\n                continue\n\n            try:\n                new_cfg = update_inference_configs(configs, data[\"N_token\"].item())\n                # Only generate as many samples as we actually need to fill the slots\n                new_cfg.sample_diffusion.N_sample = n_needed\n                runner.update_model_configs(new_cfg)\n\n                prediction = runner.predict(data)\n                raw_coords = prediction[\"coordinate\"] # Shape: [N_sample, all_atoms, 3]\n\n                # -----------------------------------------------------------\n                # DEBUG PRINT START\n                # -----------------------------------------------------------\n                print(f\"\\n[DEBUG] {target_id} | n_needed: {n_needed} | SeqLen: {len(full_seq)}\")\n                print(f\"[DEBUG] raw_coords shape: {raw_coords.shape}\")\n                \n                feat = data[\"input_feature_dict\"]\n                \n                # Check potential masks\n                mask_candidates = {}\n                if \"centre_atom_mask\" in feat:\n                    m = feat[\"centre_atom_mask\"]\n                    mask_candidates['centre_atom_mask'] = (m.sum().item(), m.shape)\n                \n                if \"atom_to_tokatom_idx\" in feat:\n                    idx_11 = (feat[\"atom_to_tokatom_idx\"] == 11).sum().item()\n                    idx_12 = (feat[\"atom_to_tokatom_idx\"] == 12).sum().item()\n                    mask_candidates['idx_11'] = idx_11\n                    mask_candidates['idx_12'] = idx_12\n                \n                print(f\"[DEBUG] Mask candidates counts: {mask_candidates}\")\n                # -----------------------------------------------------------\n                # DEBUG PRINT END\n                # -----------------------------------------------------------\n\n                # ─────────────────────────────────────────────────────────────\n                # DEBUG / FIX: Explicit C1' masking logic\n                # ─────────────────────────────────────────────────────────────\n                # Try to use 'centre_atom_mask' from features if possible\n                if \"centre_atom_mask\" in feat:\n                    mask = (feat[\"centre_atom_mask\"] == 1).to(raw_coords.device)\n                elif \"atom_to_tokatom_idx\" in feat:\n                    # Heuristic: pick the one closest to sequence length\n                    m11 = (feat[\"atom_to_tokatom_idx\"] == 11).to(raw_coords.device)\n                    m12 = (feat[\"atom_to_tokatom_idx\"] == 12).to(raw_coords.device)\n                    \n                    c11, c12 = m11.sum(), m12.sum()\n                    target_len = len(full_seq) # closer to N_token usually\n                    \n                    if abs(c11 - target_len) < abs(c12 - target_len):\n                         mask = m11\n                         print(f\"[DEBUG] Selected idx 11 mask (count={c11})\")\n                    else:\n                         mask = m12\n                         print(f\"[DEBUG] Selected idx 12 mask (count={c12})\")\n                else:\n                    # Should not happen\n                    mask = torch.zeros(raw_coords.shape[1], dtype=torch.bool, device=raw_coords.device)\n                \n                # Extract\n                coords = raw_coords[:, mask, :].detach().cpu().numpy()\n                print(f\"[DEBUG] Extracted coords shape: {coords.shape}\")\n\n                # If we get duplicate coordinates (collapsed), this is bad.\n                # Check for duplications in first sample\n                if coords.shape[1] > 1:\n                     diffs = np.linalg.norm(coords[0, 1:] - coords[0, :-1], axis=-1)\n                     if np.all(diffs < 1e-4):\n                         print(f\"  WARNING: {target_id} has identical coordinates for all residues! (Model collapse?)\")\n                \n                # Pad/trim to full (un-truncated) sequence length\n                if coords.shape[1] != len(full_seq):\n                    # Check for broadcast issue or model collapse\n                    if coords.shape[1] == 1 and len(full_seq) > 1:\n                        # Model outputted only 1 residue/atom but we need many?\n                        # Broadcast the single coord to all positions just in case (though highly suspicious)\n                        # Or perhaps mask was wrong and selected only 1 atom.\n                        # Do NOT broadcast, fill with zeros to be safe.\n                        print(f\"[DEBUG] WARNING: {target_id}: mask selected only 1 atom, but sequence is {len(full_seq)}\")\n                        # padded = np.zeros(...) -> kept as zeros\n                    else:\n                        padded  = np.zeros((coords.shape[0], len(full_seq), 3), dtype=np.float32)\n                        min_len = min(coords.shape[1], len(full_seq))\n                        if min_len > 0:\n                            padded[:, :min_len, :] = coords[:, :min_len, :]\n                        coords = padded\n\n                # Final check for identical coordinates (indicative of model failure)\n                if coords.shape[1] > 1:\n                     diffs = np.linalg.norm(coords[0, 1:] - coords[0, :-1], axis=-1)\n                     if np.all(diffs < 1e-4):\n                         print(f\"  WARNING: {target_id}: Identical coordinates detected! Resetting to zeros.\")\n                         coords = np.zeros_like(coords)\n\n                protenix_preds[target_id] = coords\n                print(f\"  {target_id}: {coords.shape[0]} Protenix predictions generated\")\n\n            except Exception as exc:\n                print(f\"  {target_id}: Protenix FAILED — {exc}\")\n                import traceback\n                traceback.print_exc()\n                protenix_preds[target_id] = None\n\n            finally:\n                del prediction, raw_coords, mask, data, atom_array\n                gc.collect(); torch.cuda.empty_cache(); gc.collect()\n# ...existing code...\n\n    elif protenix_queue and not USE_PROTENIX:\n        print(f\"\\nPHASE 2 skipped (USE_PROTENIX=False). \"\n              f\"De-novo fallback will cover {len(protenix_queue)} targets.\")\n\n    # ─── PHASE 3: Combine everything ───────────────────────────────────────\n    print(f\"\\n{'='*60}\")\n    print(\"PHASE 3: Combine TBM + Protenix + de-novo fallback\")\n    print(f\"{'='*60}\")\n\n    all_rows = []\n\n    for _, row in test_df.iterrows():\n        tid = row[\"target_id\"]\n        seq = row[\"sequence\"]\n\n        combined: list = list(template_preds.get(tid, []))  # TBM predictions\n\n        # Append Protenix predictions to fill remaining slots\n        ptx = protenix_preds.get(tid)\n        if ptx is not None and ptx.ndim == 3:\n            for j in range(ptx.shape[0]):\n                if len(combined) >= N_SAMPLE:\n                    break\n                combined.append(ptx[j])  # (seq_len, 3)\n\n        # De-novo fallback for any still-empty slots\n        n_denovo = 0\n        while len(combined) < N_SAMPLE:\n            seed_val = row.name * 1000000 + len(combined) * 1000\n            dn       = generate_rna_structure(seq, seed=seed_val)\n            combined.append(adaptive_rna_constraints(dn, tid, segments_map, confidence=0.2))\n            n_denovo += 1\n\n        if n_denovo:\n            print(f\"  {tid}: {n_denovo} slot(s) filled with de-novo fallback\")\n\n        # Stack to (N_SAMPLE, seq_len, 3) and write rows\n        stacked = np.stack(combined[:N_SAMPLE], axis=0)\n        all_rows.extend(coords_to_rows(tid, seq, stacked))\n\n    # ── Save ───────────────────────────────────────────────────────────────\n    sub = pd.DataFrame(all_rows)\n    cols = [\"ID\", \"resname\", \"resid\"] + [\n        f\"{c}_{i}\" for i in range(1, N_SAMPLE + 1) for c in [\"x\", \"y\", \"z\"]\n    ]\n    coord_cols = [c for c in cols if c.startswith((\"x_\", \"y_\", \"z_\"))]\n    sub[coord_cols] = sub[coord_cols].clip(-999.999, 9999.999)\n    sub[cols].to_csv(output_csv, index=False)\n\n    print(f\"\\n✓ Saved submission to {output_csv}  ({len(sub):,} rows)\")\n\n","metadata":{},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Main","metadata":{}},{"cell_type":"code","source":"\nif __name__ == \"__main__\":\n    main()\n","metadata":{},"outputs":[],"execution_count":null},{"cell_type":"code","source":"#read submission.csv\nsubmission_path = \"/kaggle/working/submission.csv\"\nsubmission_df = pd.read_csv(submission_path)\nprint(submission_df.head(20))","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}