{"metadata":{"kernelspec":{"display_name":"Python 3","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":[{"sourceId":118765,"databundleVersionId":15231210,"sourceType":"competition"},{"sourceId":14604295,"sourceType":"datasetVersion","datasetId":9328538}],"dockerImageVersionId":31260,"isInternetEnabled":false,"language":"python","sourceType":"notebook","isGpuEnabled":true},"papermill":{"default_parameters":{},"duration":2354.045049,"end_time":"2026-01-13T06:54:44.524542","environment_variables":{},"exception":null,"input_path":"__notebook__.ipynb","output_path":"__notebook__.ipynb","parameters":{},"start_time":"2026-01-13T06:15:30.479493","version":"2.6.0"}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"Thanks to \"https://www.kaggle.com/code/kami1976/stanford-rna-3d-folding-part-2a18\"","metadata":{}},{"cell_type":"code","source":"!pip install --no-index /kaggle/input/biopython-cp312/biopython-1.86-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-01-31T15:16:34.187731Z","iopub.execute_input":"2026-01-31T15:16:34.188021Z","iopub.status.idle":"2026-01-31T15:16:39.626411Z","shell.execute_reply.started":"2026-01-31T15:16:34.187994Z","shell.execute_reply":"2026-01-31T15:16:39.625419Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import pandas as pd\nimport numpy as np\nimport random\nimport time\nimport warnings\nimport os, sys\n\nwarnings.filterwarnings('ignore')\n\nDATA_PATH = '/kaggle/input/stanford-rna-3d-folding-2/'\ntrain_seqs = pd.read_csv(DATA_PATH + 'train_sequences.csv')\ntest_seqs = pd.read_csv(DATA_PATH + 'test_sequences.csv')\ntrain_labels = pd.read_csv(DATA_PATH + 'train_labels.csv')\n\nsys.path.append(os.path.join(DATA_PATH, \"extra\"))\n\n# --- Robust import for Kaggle's extra/parse_fasta_py.py (it may miss typing imports) ---\ntry:\n    import typing as _typing\n    import builtins as _builtins\n\n    # Make these names available during module import-time annotation evaluation\n    _builtins.Dict  = getattr(_typing, \"Dict\")\n    _builtins.Tuple = getattr(_typing, \"Tuple\")\n    _builtins.List  = getattr(_typing, \"List\")\n\n    from parse_fasta_py import parse_fasta as _parse_fasta_raw\n\n    # Normalize output to: {chain_id: sequence_string}\n    def parse_fasta(fasta_content: str):\n        d = _parse_fasta_raw(fasta_content)\n        out = {}\n        for k, v in d.items():\n            # some variants return (sequence, headers/lines) or similar\n            out[k] = v[0] if isinstance(v, tuple) else v\n        return out\n\nexcept Exception:\n    # Fallback FASTA parser: {chain_id: sequence_string}\n    def parse_fasta(fasta_content: str):\n        out = {}\n        cur = None\n        seq_parts = []\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(seq_parts)\n                header = line[1:]\n                # First token is usually chain id in this dataset\n                cur = header.split()[0]\n                seq_parts = []\n            else:\n                seq_parts.append(line.replace(\" \", \"\"))\n        if cur is not None:\n            out[cur] = \"\".join(seq_parts)\n        return out\n\ndef parse_stoichiometry(stoich: str):\n    if pd.isna(stoich) or str(stoich).strip() == \"\":\n        return []\n    out = []\n    for part in str(stoich).split(';'):\n        ch, cnt = part.split(':')\n        out.append((ch.strip(), int(cnt)))\n    return out\n\ndef get_chain_segments(row):\n    \"\"\"\n    Returns list of (start,end) segments in row['sequence'] corresponding to chain copies in stoichiometry order.\n    Falls back to single segment if parsing fails.\n    \"\"\"\n    seq = row['sequence']\n    stoich = row.get('stoichiometry', '')\n    all_seq = row.get('all_sequences', '')\n\n    if pd.isna(stoich) or pd.isna(all_seq) or str(stoich).strip()==\"\" or str(all_seq).strip()==\"\":\n        return [(0, len(seq))]\n\n    try:\n        chain_dict = parse_fasta(all_seq)  # dict: chain_id -> sequence\n        order = parse_stoichiometry(stoich)\n        segs = []\n        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                L = len(base)\n                segs.append((pos, pos + L))\n                pos += L\n        if pos != len(seq):\n            return [(0, len(seq))]\n        return segs\n    except Exception:\n        return [(0, len(seq))]\n\ndef build_segments_map(df):\n    seg_map = {}\n    stoich_map = {}\n    for _, r in df.iterrows():\n        tid = r['target_id']\n        seg_map[tid] = get_chain_segments(r)\n        stoich_map[tid] = str(r.get('stoichiometry', '') if not pd.isna(r.get('stoichiometry', '')) else '')\n    return seg_map, stoich_map\n\ntrain_segs_map, train_stoich_map = build_segments_map(train_seqs)\ntest_segs_map,  test_stoich_map  = build_segments_map(test_seqs)\n\ndef process_labels(labels_df):\n    coords_dict = {}\n    # Faster + safer prefix extraction\n    prefixes = labels_df['ID'].str.rsplit('_', n=1).str[0]\n    for id_prefix, group in labels_df.groupby(prefixes):\n        coords_dict[id_prefix] = group.sort_values('resid')[['x_1', 'y_1', 'z_1']].values\n    return coords_dict\n\ntrain_coords_dict = process_labels(train_labels)\n\nfrom Bio.Align import PairwiseAligner\n\naligner = PairwiseAligner()\naligner.mode = 'global'\naligner.match_score = 2\naligner.mismatch_score = -1.5\n\n# Stronger gap penalties discourage \"sliding\" (critical: residue numbering must match)\naligner.open_gap_score   = -8\naligner.extend_gap_score = -0.4\n\n# Also penalize terminal gaps (prevents end-gap semi-global behavior)\naligner.query_left_open_gap_score  = -8\naligner.query_left_extend_gap_score = -0.4\naligner.query_right_open_gap_score = -8\naligner.query_right_extend_gap_score = -0.4\naligner.target_left_open_gap_score = -8\naligner.target_left_extend_gap_score = -0.4\naligner.target_right_open_gap_score = -8\naligner.target_right_extend_gap_score = -0.4\n\ndef find_similar_sequences(query_seq, train_seqs_df, train_coords_dict, top_n=5):\n    similar_seqs = []\n    \n    # Pre-filter: Iterate only valid targets\n    # Note: aligner.score is much faster than generating full alignments\n    for _, row in train_seqs_df.iterrows():\n        target_id, train_seq = row['target_id'], row['sequence']\n        if target_id not in train_coords_dict: continue\n        \n        # Length filter (keep your original logic)\n        if abs(len(train_seq) - len(query_seq)) / max(len(train_seq), len(query_seq)) > 0.3: continue\n        \n        # FAST SCORE: Calculates score without traceback overhead\n        raw_score = aligner.score(query_seq, train_seq)\n        \n        normalized_score = raw_score / (2 * min(len(query_seq), len(train_seq)))\n        similar_seqs.append((target_id, train_seq, normalized_score, train_coords_dict[target_id]))\n    \n    similar_seqs.sort(key=lambda x: x[2], reverse=True)\n    return similar_seqs[:top_n]\n\ndef adapt_template_to_query(query_seq, template_seq, template_coords):\n    # Generate the alignment object\n    # aligner.align returns an iterator; we take the first optimal alignment\n    alignment = next(iter(aligner.align(query_seq, template_seq)))\n    \n    new_coords = np.full((len(query_seq), 3), np.nan)\n    \n    # VECTORIZED MAPPING:\n    # alignment.aligned returns lists of (start, end) tuples for matched segments.\n    # This avoids the slow python loop \"for char_q, char_t in zip...\"\n    for (q_start, q_end), (t_start, t_end) in zip(*alignment.aligned):\n        # Map the coordinate chunk directly\n        t_chunk = template_coords[t_start:t_end]\n        \n        # Safety check to ensure shapes match (handles edge cases)\n        if len(t_chunk) == (q_end - q_start):\n            new_coords[q_start:q_end] = t_chunk\n\n    # --- Interpolation Logic (Unchanged) ---\n    for i in range(len(new_coords)):\n        if np.isnan(new_coords[i, 0]):\n            prev_v = next((j for j in range(i-1, -1, -1) if not np.isnan(new_coords[j, 0])), -1)\n            next_v = next((j for j in range(i+1, len(new_coords)) if not np.isnan(new_coords[j, 0])), -1)\n            if prev_v >= 0 and next_v >= 0:\n                w = (i - prev_v) / (next_v - prev_v)\n                new_coords[i] = (1-w)*new_coords[prev_v] + w*new_coords[next_v]\n            elif prev_v >= 0: new_coords[i] = new_coords[prev_v] + [3, 0, 0]\n            elif next_v >= 0: new_coords[i] = new_coords[next_v] + [3, 0, 0]\n            else: new_coords[i] = [i*3, 0, 0]\n            \n    return np.nan_to_num(new_coords)\n\ndef adaptive_rna_constraints(coordinates, sequence, confidence=1.0):\n    refined_coords = coordinates.copy()\n    n = len(sequence)\n    strength = 0.68 * (1.0 - min(confidence, 0.96))\n\n    for _ in range(2):\n        for i in range(n - 1):\n            p1, p2 = refined_coords[i], refined_coords[i+1]\n            dist = np.linalg.norm(p2 - p1)\n            if dist > 0:\n                adj = (5.95 - dist) * strength * 0.45\n                refined_coords[i+1] += (p2 - p1) / dist * adj\n\n            if i < n - 2:\n                p3 = refined_coords[i+2]\n                dist2 = np.linalg.norm(p3 - p1)\n                if dist2 > 0:\n                    adj2 = (10.2 - dist2) * strength * 0.25\n                    refined_coords[i+2] += (p3 - p1) / dist2 * adj2\n\n    return refined_coords\n\ndef predict_rna_structures(sequence, target_id, train_seqs_df, train_coords_dict, n_predictions=5):\n    predictions = []\n    similar_seqs = find_similar_sequences(sequence, train_seqs_df, train_coords_dict, top_n=n_predictions)\n    \n    for i in range(n_predictions):\n        if i < len(similar_seqs):\n            t_id, t_seq, sim, t_coords = similar_seqs[i]\n            adapted = adapt_template_to_query(sequence, t_seq, t_coords)\n            refined = adaptive_rna_constraints(adapted, sequence, confidence=sim)\n            \n            # ШУМ: Слот 0 - чистый, остальные - микро-шум\n            noise = 0.0 if i == 0 else max(0.006, (0.38 - sim) * 0.07)\n            if noise > 0: refined += np.random.normal(0, noise, refined.shape)\n            predictions.append(refined)\n        else:\n            n = len(sequence)\n            coords = np.zeros((n, 3))\n            for j in range(1, n): coords[j] = coords[j-1] + [4.0, 0, 0]\n            predictions.append(coords)\n    return predictions\n\nall_predictions = []\nstart_time = time.time()\nfor idx, row in test_seqs.iterrows():\n    if idx % 10 == 0: print(f\"Processing {idx} | {time.time()-start_time:.1f}s\")\n    tid, seq = row['target_id'], row['sequence']\n    preds = predict_rna_structures(seq, tid, train_seqs, train_coords_dict)\n    for j in range(len(seq)):\n        res = {'ID': f\"{tid}_{j+1}\", 'resname': seq[j], 'resid': j+1}\n        for i in range(5):\n            res[f'x_{i+1}'], res[f'y_{i+1}'], res[f'z_{i+1}'] = preds[i][j]\n        all_predictions.append(res)\n\nsub = pd.DataFrame(all_predictions)\ncols = ['ID', 'resname', 'resid'] + [f'{c}_{i}' for i in range(1,6) for c in ['x','y','z']]\nsub[cols].to_csv('submission.csv', index=False)\nprint(\"submission.csv! saved\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2026-01-31T15:16:39.628246Z","iopub.execute_input":"2026-01-31T15:16:39.628557Z","iopub.status.idle":"2026-01-31T15:19:10.011069Z","shell.execute_reply.started":"2026-01-31T15:16:39.628524Z","shell.execute_reply":"2026-01-31T15:19:10.010344Z"}},"outputs":[],"execution_count":null}]}