{"cells":[{"cell_type":"code","execution_count":null,"id":"af118806","metadata":{"language":"python"},"outputs":[],"source":"from pathlib import Path\n\nfrom IPython.display import Image, Markdown, display\n\n\n\nSTATIC_IMAGE_DIR = Path(\"/kaggle/input/notebook-static-images-public\")\n\ndisplay(Markdown(\"## Saved analysis figures\"))\n\nfor image_path in sorted(STATIC_IMAGE_DIR.glob(\"integration-*.png\")):\n\n display(Markdown(f\"**{image_path.stem}**\"))\n\n display(Image(filename=str(image_path)))"},{"cell_type":"markdown","id":"13d0758d-7b69-4f84-bf66-1bd6c8d18c56","metadata":{"jp-MarkdownHeadingCollapsed":true},"source":"# Data merge : Multiome(scRNA + scATAC)"},{"cell_type":"code","execution_count":null,"id":"4fea546a-0419-4610-8c9f-181ac30d8e47","metadata":{},"outputs":[],"source":"import pandas as pd\n\ninput_path = \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/train/train_multi_inputs.h5\"\ntarget_path = \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/train/train_multi_targets.h5\"\n\ntrain_multi_inputs = pd.read_hdf(input_path, key=\"train_multi_inputs\")\ntrain_multi_targets = pd.read_hdf(target_path, key=\"train_multi_targets\")\n\nprint(train_multi_inputs.shape)\nprint(train_multi_targets.shape)"},{"cell_type":"code","execution_count":null,"id":"469dffd0-6291-4230-a662-6a7908bc22df","metadata":{},"outputs":[],"source":"import pandas as pd\nimport numpy as np\nimport scipy.sparse as sp\nimport anndata as ad\nimport pyranges as pr\n\nbase = \"/home/Data_Drive_8TB/kykim/7. Kaggle/Single-Cell_Perturbations/baseline\"\n\npromoter_path = f\"{base}/promoter.bed\"\nenhancer_path = f\"{base}/enhancer.bed\"\ngtf_path = f\"{base}/gencode.v44.annotation.gtf.gz\"\n\ngtf = pd.read_csv(\n gtf_path,\n sep=\"\\t\",\n comment=\"#\",\n header=None,\n names=[\n \"Chromosome\",\n \"Source\",\n \"Feature\",\n \"Start\",\n \"End\",\n \"Score\",\n \"Strand\",\n \"Frame\",\n \"Attribute\"\n ]\n)\n\ngtf_gene = gtf[gtf[\"Feature\"] == \"gene\"].copy()\n\ngtf_gene[\"gene_id\"] = gtf_gene[\"Attribute\"].str.extract(\n r'gene_id \"([^\"]+)\"'\n)[0]\n\ngtf_gene[\"gene_name\"] = gtf_gene[\"Attribute\"].str.extract(\n r'gene_name \"([^\"]+)\"'\n)[0]\n\ngtf_gene[\"gene_id\"] = gtf_gene[\"gene_id\"].str.replace(\n r\"\\.\\d+$\", \"\", regex=True\n)\n\ntarget = train_multi_targets.copy()\ntarget.index = target.index.astype(str)\n\ntarget_ensg = pd.Index(\n target.columns.astype(str)\n).str.replace(\n r\"\\.\\d+$\", \"\", regex=True\n)\n\nensg_to_symbol = (\n gtf_gene[\n [\"gene_id\", \"gene_name\"]\n ]\n .dropna()\n .drop_duplicates(\"gene_id\")\n .set_index(\"gene_id\")[\"gene_name\"]\n .to_dict()\n)\n\ntarget_symbols = target_ensg.map(ensg_to_symbol)\n\nvalid_target = target_symbols.notna()\n\ntarget = target.loc[:, valid_target]\ntarget.columns = target_symbols[valid_target]\n\ntarget = target.loc[:, ~target.columns.duplicated()]\n\ngene_order = pd.Index(target.columns.astype(str))\n\ngtf_gene = gtf_gene[\n gtf_gene[\"gene_id\"].isin(\n target_ensg[valid_target]\n )\n].copy()\n\ngene_tss = gtf_gene[\n [\n \"Chromosome\",\n \"Start\",\n \"End\",\n \"Strand\",\n \"gene_id\",\n \"gene_name\"\n ]\n].copy()\n\ngene_tss[\"Start\"] = np.where(\n gene_tss[\"Strand\"] == \"+\",\n gene_tss[\"Start\"] - 1,\n gene_tss[\"End\"] - 1\n)\n\ngene_tss[\"End\"] = gene_tss[\"Start\"] + 1\n\ngene_tss_pr = pr.PyRanges(gene_tss)\n\npromoter = pd.read_csv(\n promoter_path,\n sep=\"\\t\",\n header=None,\n names=[\n \"Chromosome\",\n \"Start\",\n \"End\",\n \"ID1\",\n \"ID2\",\n \"Annotation\"\n ]\n)\n\nenhancer = pd.read_csv(\n enhancer_path,\n sep=\"\\t\",\n header=None,\n names=[\n \"Chromosome\",\n \"Start\",\n \"End\",\n \"ID1\",\n \"ID2\",\n \"Annotation\"\n ]\n)\n\npromoter_pr = pr.PyRanges(promoter)\nenhancer_pr = pr.PyRanges(enhancer)\n\npromoter_gene = promoter_pr.join(\n gene_tss_pr\n).df\n\npromoter_gene = promoter_gene[\n [\n \"Chromosome\",\n \"Start\",\n \"End\",\n \"gene_id\"\n ]\n].drop_duplicates()\n\nenhancer_gene = enhancer_pr.nearest(\n gene_tss_pr,\n strandedness=False\n).df\n\nenhancer_gene = enhancer_gene[\n [\n \"Chromosome\",\n \"Start\",\n \"End\",\n \"gene_id\"\n ]\n].drop_duplicates()\n\ninput_cells = train_multi_inputs.index.astype(str)\n\nif \"cell_id\" in train_multi_inputs.columns:\n input_cells = train_multi_inputs[\"cell_id\"].astype(str).values\n peak_cols = train_multi_inputs.columns[1:]\nelse:\n peak_cols = train_multi_inputs.columns\n\npeak_cols = pd.Index(peak_cols.astype(str))\n\npeak_info = pd.DataFrame({\n \"peak\": peak_cols\n})\n\npeak_info[[\"Chromosome\", \"position\"]] = peak_info[\n \"peak\"\n].str.split(\":\", n=1, expand=True)\n\npeak_info[[\"Start\", \"End\"]] = peak_info[\n \"position\"\n].str.split(\"-\", n=1, expand=True)\n\npeak_info[\"Start\"] = peak_info[\"Start\"].astype(np.int64)\npeak_info[\"End\"] = peak_info[\"End\"].astype(np.int64)\n\npeak_pr = pr.PyRanges(\n peak_info[\n [\n \"Chromosome\",\n \"Start\",\n \"End\",\n \"peak\"\n ]\n ]\n)\n\npeak_promoter = peak_pr.join(\n pr.PyRanges(promoter_gene)\n).df\n\npeak_promoter = peak_promoter[\n [\n \"peak\",\n \"gene_id\"\n ]\n].drop_duplicates()\n\npeak_enhancer = peak_pr.join(\n pr.PyRanges(enhancer_gene)\n).df\n\npeak_enhancer = peak_enhancer[\n [\n \"peak\",\n \"gene_id\"\n ]\n].drop_duplicates()\n\npromoter_peak_set = set(\n peak_promoter[\"peak\"]\n)\n\npeak_enhancer = peak_enhancer[\n ~peak_enhancer[\"peak\"].isin(\n promoter_peak_set\n )\n].copy()\n\npromoter_map = peak_promoter.copy()\nenhancer_map = peak_enhancer.copy()\n\npromoter_map[\"gene_symbol\"] = promoter_map[\n \"gene_id\"\n].map(ensg_to_symbol)\n\nenhancer_map[\"gene_symbol\"] = enhancer_map[\n \"gene_id\"\n].map(ensg_to_symbol)\n\npromoter_map = promoter_map[\n promoter_map[\"gene_symbol\"].isin(gene_order)\n].copy()\n\nenhancer_map = enhancer_map[\n enhancer_map[\"gene_symbol\"].isin(gene_order)\n].copy()\n\ngene_to_idx = {\n gene: i\n for i, gene in enumerate(gene_order)\n}\n\npeak_to_idx = {\n peak: i\n for i, peak in enumerate(peak_cols)\n}\n\npromoter_map[\"gene_idx\"] = promoter_map[\n \"gene_symbol\"\n].map(gene_to_idx)\n\npromoter_map[\"peak_idx\"] = promoter_map[\n \"peak\"\n].map(peak_to_idx)\n\nenhancer_map[\"gene_idx\"] = enhancer_map[\n \"gene_symbol\"\n].map(gene_to_idx)\n\nenhancer_map[\"peak_idx\"] = enhancer_map[\n \"peak\"\n].map(peak_to_idx)\n\npromoter_map = promoter_map.dropna(\n subset=[\"gene_idx\", \"peak_idx\"]\n)\n\nenhancer_map = enhancer_map.dropna(\n subset=[\"gene_idx\", \"peak_idx\"]\n)\n\npromoter_map[\"gene_idx\"] = promoter_map[\n \"gene_idx\"\n].astype(np.int64)\n\npromoter_map[\"peak_idx\"] = promoter_map[\n \"peak_idx\"\n].astype(np.int64)\n\nenhancer_map[\"gene_idx\"] = enhancer_map[\n \"gene_idx\"\n].astype(np.int64)\n\nenhancer_map[\"peak_idx\"] = enhancer_map[\n \"peak_idx\"\n].astype(np.int64)\n\nn_genes = len(gene_order)\nn_peaks = len(peak_cols)\n\nP = sp.csr_matrix(\n (\n np.ones(len(promoter_map), dtype=np.float32),\n (\n promoter_map[\"peak_idx\"].values,\n promoter_map[\"gene_idx\"].values\n )\n ),\n shape=(n_peaks, n_genes)\n)\n\nE = sp.csr_matrix(\n (\n np.ones(len(enhancer_map), dtype=np.float32),\n (\n enhancer_map[\"peak_idx\"].values,\n enhancer_map[\"gene_idx\"].values\n )\n ),\n shape=(n_peaks, n_genes)\n)\n\ninput_cell_to_row = {\n cell: i\n for i, cell in enumerate(input_cells)\n}\n\ntarget_cells = target.index.astype(str)\n\ncommon_cells = target_cells[\n target_cells.isin(input_cell_to_row)\n]\n\ntarget = target.loc[common_cells]\n\ninput_rows = np.array([\n input_cell_to_row[cell]\n for cell in common_cells\n])\n\nn_cells = len(common_cells)\n\npromoter_matrix = np.zeros(\n (n_cells, n_genes),\n dtype=np.float32\n)\n\nenhancer_matrix = np.zeros(\n (n_cells, n_genes),\n dtype=np.float32\n)\n\nchunk_size = 256\n\nfor start in range(0, n_cells, chunk_size):\n\n end = min(\n start + chunk_size,\n n_cells\n )\n\n rows = input_rows[start:end]\n\n if \"cell_id\" in train_multi_inputs.columns:\n X = train_multi_inputs.iloc[\n rows, 1:\n ].to_numpy(\n dtype=np.float32\n )\n else:\n X = train_multi_inputs.iloc[\n rows\n ].to_numpy(\n dtype=np.float32\n )\n\n promoter_matrix[start:end] = (\n X @ P\n )\n\n enhancer_matrix[start:end] = (\n X @ E\n )\n\ntarget = target.loc[\n common_cells,\n gene_order\n]\n\nrna_matrix = target.to_numpy(\n dtype=np.float32\n)\n\nadata_multi = ad.AnnData(\n X=sp.csr_matrix(rna_matrix)\n)\n\nadata_multi.obs_names = common_cells\nadata_multi.var_names = gene_order\n\nadata_multi.layers[\"rna\"] = sp.csr_matrix(\n rna_matrix\n)\n\nadata_multi.layers[\"A_promoter\"] = sp.csr_matrix(\n promoter_matrix\n)\n\nadata_multi.layers[\"B_enhancer\"] = sp.csr_matrix(\n enhancer_matrix\n)"},{"cell_type":"code","execution_count":null,"id":"52553e4d-8b08-40b4-8152-5cce17c83aa8","metadata":{},"outputs":[],"source":"print(\"adata shape:\", adata_multi.shape)\n\nprint(\n \"RNA:\",\n adata_multi.layers[\"rna\"].shape\n)\n\nprint(\n \"A_promoter:\",\n adata_multi.layers[\"A_promoter\"].shape\n)\n\nprint(\n \"B_enhancer:\",\n adata_multi.layers[\"B_enhancer\"].shape\n)\n\nprint(\n \"cell 정합:\",\n np.array_equal(\n adata_multi.obs_names,\n target.index\n )\n)\n\nprint(\n \"gene 정합:\",\n np.array_equal(\n adata_multi.var_names,\n target.columns\n )\n)\n\nprint(\n \"RNA-Promoter 정합:\",\n adata_multi.layers[\"rna\"].shape\n == adata_multi.layers[\"A_promoter\"].shape\n)\n\nprint(\n \"RNA-Enhancer 정합:\",\n adata_multi.layers[\"rna\"].shape\n == adata_multi.layers[\"B_enhancer\"].shape\n)\n\nprint(\n \"Promoter-Enhancer 정합:\",\n adata_multi.layers[\"A_promoter\"].shape\n == adata_multi.layers[\"B_enhancer\"].shape\n)"},{"cell_type":"code","execution_count":null,"id":"9c1b69d3-78ca-4d48-8b49-7f1aa7af2224","metadata":{},"outputs":[],"source":"metadata = pd.read_csv(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/metadata.csv\"\n)\n\nmetadata[\"cell_id\"] = metadata[\"cell_id\"].astype(str)\n\nmetadata = metadata.set_index(\"cell_id\")\n\nmetadata = metadata.loc[adata_multi.obs_names]\n\nadata_multi.obs = metadata.copy()"},{"cell_type":"code","execution_count":null,"id":"fb473737-eb98-4475-aa68-0db1e2f17bde","metadata":{},"outputs":[],"source":"output_path = \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/train/train_multi.h5ad\"\nadata_multi.write_h5ad(output_path)"},{"cell_type":"markdown","id":"33514366-04df-448f-a8ba-fd7eda36482c","metadata":{"jp-MarkdownHeadingCollapsed":true},"source":"# Data merge : Multiome(scRNA + CITEseq)\n"},{"cell_type":"code","execution_count":null,"id":"d9c7d5bd-6f86-4b1d-8c9e-606c697d8f93","metadata":{},"outputs":[],"source":"import pandas as pd\n\ninput_path = \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/train/train_cite_inputs.h5\"\ntarget_path = \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/train/train_cite_targets.h5\"\n\ntrain_cite_inputs = pd.read_hdf(\n input_path,\n key=\"train_cite_inputs\"\n)\n\ntrain_cite_targets = pd.read_hdf(\n target_path,\n key=\"train_cite_targets\"\n)\n\nprint(train_cite_inputs.shape)\nprint(train_cite_targets.shape)"},{"cell_type":"code","execution_count":null,"id":"47c022bd-3fc7-420d-8446-3db0af079227","metadata":{},"outputs":[],"source":"train_cite_inputs.columns = (\n train_cite_inputs.columns\n .str.split(\"_\", n=1)\n .str[1]\n)"},{"cell_type":"code","execution_count":null,"id":"6b7e39a5-a43b-493b-909d-0a31e34e61d1","metadata":{},"outputs":[],"source":"import mudata as md\nimport anndata as ad\nimport scipy.sparse as sp\n\ntrain_cite_inputs.index = train_cite_inputs.index.astype(str)\ntrain_cite_targets.index = train_cite_targets.index.astype(str)\n\ncommon_cells = train_cite_inputs.index.intersection(\n train_cite_targets.index\n)\n\nrna = train_cite_inputs.loc[common_cells].copy()\nprotein = train_cite_targets.loc[common_cells].copy()\n\nrna.index = common_cells\nprotein.index = common_cells\n\nadata_rna = ad.AnnData(\n X=sp.csr_matrix(\n rna.to_numpy(dtype=\"float32\")\n )\n)\n\nadata_rna.obs_names = common_cells\nadata_rna.var_names = rna.columns.astype(str)\n\nadata_protein = ad.AnnData(\n X=sp.csr_matrix(\n protein.to_numpy(dtype=\"float32\")\n )\n)\n\nadata_protein.obs_names = common_cells\nadata_protein.var_names = protein.columns.astype(str)\n\nmdata_cite = md.MuData({\n \"rna\": adata_rna,\n \"protein\": adata_protein\n})\n\nprint(mdata_cite)\nprint(mdata_cite[\"rna\"].shape)\nprint(mdata_cite[\"protein\"].shape)"},{"cell_type":"code","execution_count":null,"id":"f1ff3f9b-7301-4c18-8464-7e89f448fa48","metadata":{},"outputs":[],"source":"metadata = pd.read_csv(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/metadata.csv\"\n)\nmdata_cite.obs = metadata.set_index(\"cell_id\").loc[mdata_cite[\"rna\"].obs_names].copy()"},{"cell_type":"code","execution_count":null,"id":"6c296eaf-4209-4c3d-ba0e-589e4b35e5df","metadata":{},"outputs":[],"source":"output_path = \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/train/train_cite.h5ad\"\n\nmdata_cite.write_h5mu(output_path)"},{"cell_type":"markdown","id":"5def0d13-b2b4-47c3-bdba-fa5e1f3fd512","metadata":{"jp-MarkdownHeadingCollapsed":true},"source":"# Test data"},{"cell_type":"code","execution_count":null,"id":"9448501a-342b-4741-ae56-f31d75e7282c","metadata":{},"outputs":[],"source":"import pandas as pd\n\ntest_cite_day2 = pd.read_hdf(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/test/test_cite_inputs_day_2_donor_27678.h5\",\n key=\"test_cite_inputs_day_2_donor_27678\"\n)\n\ntest_cite = pd.read_hdf(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/test/test_cite_inputs.h5\",\n key=\"test_cite_inputs\"\n)\n\ntest_multi = pd.read_hdf(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/test/test_multi_inputs.h5\",\n key=\"test_multi_inputs\"\n)\n\nprint(test_cite_day2.shape)\nprint(test_cite.shape)\nprint(test_multi.shape)"},{"cell_type":"code","execution_count":null,"id":"3c986c1f-91e9-4e1b-9aac-22b41855d680","metadata":{},"outputs":[],"source":"test_cite_day2.columns = (\n test_cite_day2.columns\n .str.split(\"_\", n=1)\n .str[1])\n\ntest_cite.columns = (\n test_cite.columns\n .str.split(\"_\", n=1)\n .str[1])"},{"cell_type":"code","execution_count":null,"id":"74ee1508-a045-4726-b47a-45bae8ded226","metadata":{},"outputs":[],"source":"test_cite_day2 = test_cite_day2.loc[:, ~test_cite_day2.columns.duplicated()]\ntest_cite = test_cite.loc[:, ~test_cite.columns.duplicated()]"},{"cell_type":"code","execution_count":null,"id":"df03d6cf-0848-4339-ba63-b299937d9871","metadata":{},"outputs":[],"source":"common_cols = test_cite_day2.columns.intersection(test_cite.columns)\n\ntest_cite_combined = pd.concat(\n [\n test_cite_day2[common_cols],\n test_cite[common_cols]\n ],\n axis=0,\n ignore_index=True\n)"},{"cell_type":"code","execution_count":null,"id":"1fa5221d-3448-49e4-a98b-03b8f6a143af","metadata":{},"outputs":[],"source":"import pandas as pd\n\ncommon_cols = test_cite_day2.columns.intersection(test_cite.columns)\n\ntest_cite_combined = pd.concat(\n [\n test_cite_day2[common_cols],\n test_cite[common_cols]\n ],\n axis=0,\n join=\"inner\"\n)\n\nprint(\"test_cite_day2:\", test_cite_day2.shape)\nprint(\"test_cite:\", test_cite.shape)\nprint(\"공통 열 수:\", len(common_cols))\nprint(\"매칭되지 않은 test_cite_day2 열:\", len(test_cite_day2.columns.difference(common_cols)))\nprint(\"매칭되지 않은 test_cite 열:\", len(test_cite.columns.difference(common_cols)))\nprint(\"합친 데이터:\", test_cite_combined.shape)"},{"cell_type":"code","execution_count":null,"id":"bed6871b-bb2c-4919-ae78-b59e4ed97b17","metadata":{},"outputs":[],"source":"import pandas as pd\n\nmetadata = pd.read_csv(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/metadata.csv\"\n)\n\nmetadata_cite_day2 = pd.read_csv(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/metadata_cite_day_2_donor_27678.csv\"\n)\n\nmetadata[\"cell_id\"] = metadata[\"cell_id\"].astype(str)\nmetadata_cite_day2[\"cell_id\"] = metadata_cite_day2[\"cell_id\"].astype(str)\ntest_cite_combined.index = test_cite_combined.index.astype(str)\n\nmatched_metadata = test_cite_combined.index.intersection(\n metadata[\"cell_id\"]\n)\n\nmatched_day2 = test_cite_combined.index.intersection(\n metadata_cite_day2[\"cell_id\"]\n)\n\nprint(\"test_cite_combined cell 수:\", len(test_cite_combined))\nprint(\"metadata.csv cell 수:\", len(metadata))\nprint(\"metadata.csv 매칭되는 cell 수:\", len(matched_metadata))\nprint()\nprint(\"metadata_cite_day_2_donor_27678.csv cell 수:\", len(metadata_cite_day2))\nprint(\"metadata_cite_day_2_donor_27678.csv 매칭되는 cell 수:\", len(matched_day2))"},{"cell_type":"code","execution_count":null,"id":"43571573-65fc-4d65-9343-612fbad1a3f9","metadata":{},"outputs":[],"source":"import pandas as pd\nimport anndata as ad\nimport scipy.sparse as sp\n\nmetadata = pd.read_csv(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/metadata.csv\"\n)\n\nmetadata_cite_day2 = pd.read_csv(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/metadata_cite_day_2_donor_27678.csv\"\n)\n\nmetadata[\"cell_id\"] = metadata[\"cell_id\"].astype(str)\nmetadata_cite_day2[\"cell_id\"] = metadata_cite_day2[\"cell_id\"].astype(str)\n\ntest_cite_combined.index = test_cite_combined.index.astype(str)\n\nmetadata_all = pd.concat(\n [\n metadata.set_index(\"cell_id\"),\n metadata_cite_day2.set_index(\"cell_id\")\n ],\n axis=0\n)\n\nmetadata_all = metadata_all.loc[\n ~metadata_all.index.duplicated(keep=\"first\")\n]\n\nmatched_obs = metadata_all.loc[test_cite_combined.index]\n\nadata_cite_test = ad.AnnData(\n X=sp.csr_matrix(test_cite_combined.to_numpy(dtype=\"float32\"))\n)\n\nadata_cite_test.obs_names = test_cite_combined.index\nadata_cite_test.var_names = test_cite_combined.columns.astype(str)\nadata_cite_test.obs = matched_obs.copy()\n\nprint(adata_cite_test)\nprint(adata_cite_test.obs.head())"},{"cell_type":"code","execution_count":null,"id":"a8ed1f66-b56e-4189-af7e-2c57ed1af706","metadata":{},"outputs":[],"source":"output_path = \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/test/test_rna.h5ad\"\n\nadata_cite_test.write_h5ad(output_path)"},{"cell_type":"code","execution_count":null,"id":"5adcebc4-e219-412c-bce9-a3cbfb179857","metadata":{},"outputs":[],"source":"#### ATAC split ####"},{"cell_type":"code","execution_count":null,"id":"0cc20bc6-d660-43ff-8988-61d3f9edfaac","metadata":{},"outputs":[],"source":"import pandas as pd\nimport numpy as np\nimport scipy.sparse as sp\nimport anndata as ad\nimport pyranges as pr\n\nbase = \"/home/Data_Drive_8TB/kykim/7. Kaggle/Single-Cell_Perturbations/baseline\"\n\npromoter_path = f\"{base}/promoter.bed\"\nenhancer_path = f\"{base}/enhancer.bed\"\ngtf_path = f\"{base}/gencode.v44.annotation.gtf.gz\"\n\ngtf = pd.read_csv(\n gtf_path,\n sep=\"\\t\",\n comment=\"#\",\n header=None,\n names=[\n \"Chromosome\", \"Source\", \"Feature\", \"Start\", \"End\",\n \"Score\", \"Strand\", \"Frame\", \"Attribute\"\n ]\n)\n\ngtf_gene = gtf[gtf[\"Feature\"] == \"gene\"].copy()\n\ngtf_gene[\"gene_id\"] = gtf_gene[\"Attribute\"].str.extract(\n r'gene_id \"([^\"]+)\"'\n)[0]\n\ngtf_gene[\"gene_name\"] = gtf_gene[\"Attribute\"].str.extract(\n r'gene_name \"([^\"]+)\"'\n)[0]\n\ngtf_gene[\"gene_id\"] = gtf_gene[\"gene_id\"].str.replace(\n r\"\\.\\d+$\", \"\", regex=True\n)\n\nensg_to_symbol = (\n gtf_gene[\n [\"gene_id\", \"gene_name\"]\n ]\n .dropna()\n .drop_duplicates(\"gene_id\")\n .set_index(\"gene_id\")[\"gene_name\"]\n .to_dict()\n)\n\ngene_order = pd.Index(\n gtf_gene[\"gene_name\"].dropna().drop_duplicates().astype(str)\n)\n\ngene_tss = gtf_gene[\n [\n \"Chromosome\",\n \"Start\",\n \"End\",\n \"Strand\",\n \"gene_id\",\n \"gene_name\"\n ]\n].copy()\n\ngene_tss[\"Start\"] = np.where(\n gene_tss[\"Strand\"] == \"+\",\n gene_tss[\"Start\"] - 1,\n gene_tss[\"End\"] - 1\n)\n\ngene_tss[\"End\"] = gene_tss[\"Start\"] + 1\n\ngene_tss_pr = pr.PyRanges(gene_tss)\n\npromoter = pd.read_csv(\n promoter_path,\n sep=\"\\t\",\n header=None,\n names=[\n \"Chromosome\",\n \"Start\",\n \"End\",\n \"ID1\",\n \"ID2\",\n \"Annotation\"\n ]\n)\n\nenhancer = pd.read_csv(\n enhancer_path,\n sep=\"\\t\",\n header=None,\n names=[\n \"Chromosome\",\n \"Start\",\n \"End\",\n \"ID1\",\n \"ID2\",\n \"Annotation\"\n ]\n)\n\npromoter_gene = pr.PyRanges(promoter).join(\n gene_tss_pr\n).df[\n [\"Chromosome\", \"Start\", \"End\", \"gene_id\"]\n].drop_duplicates()\n\nenhancer_gene = pr.PyRanges(enhancer).nearest(\n gene_tss_pr,\n strandedness=False\n).df[\n [\"Chromosome\", \"Start\", \"End\", \"gene_id\"]\n].drop_duplicates()\n\nif \"cell_id\" in test_multi.columns:\n cell_ids = test_multi[\"cell_id\"].astype(str).values\n peak_cols = pd.Index(test_multi.columns[1:].astype(str))\n X = test_multi.iloc[:, 1:]\nelse:\n cell_ids = test_multi.index.astype(str).values\n peak_cols = pd.Index(test_multi.columns.astype(str))\n X = test_multi\n\npeak_info = pd.DataFrame({\n \"peak\": peak_cols\n})\n\npeak_info[[\"Chromosome\", \"position\"]] = peak_info[\n \"peak\"\n].str.split(\":\", n=1, expand=True)\n\npeak_info[[\"Start\", \"End\"]] = peak_info[\n \"position\"\n].str.split(\"-\", n=1, expand=True)\n\npeak_info[\"Start\"] = peak_info[\"Start\"].astype(np.int64)\npeak_info[\"End\"] = peak_info[\"End\"].astype(np.int64)\n\npeak_pr = pr.PyRanges(\n peak_info[\n [\"Chromosome\", \"Start\", \"End\", \"peak\"]\n ]\n)\n\npeak_promoter = peak_pr.join(\n pr.PyRanges(promoter_gene)\n).df[\n [\"peak\", \"gene_id\"]\n].drop_duplicates()\n\npeak_enhancer = peak_pr.join(\n pr.PyRanges(enhancer_gene)\n).df[\n [\"peak\", \"gene_id\"]\n].drop_duplicates()\n\npromoter_peak_set = set(\n peak_promoter[\"peak\"]\n)\n\npeak_enhancer = peak_enhancer[\n ~peak_enhancer[\"peak\"].isin(promoter_peak_set)\n].copy()\n\npromoter_map = peak_promoter.copy()\nenhancer_map = peak_enhancer.copy()\n\npromoter_map[\"gene_symbol\"] = promoter_map[\"gene_id\"].map(\n ensg_to_symbol\n)\n\nenhancer_map[\"gene_symbol\"] = enhancer_map[\"gene_id\"].map(\n ensg_to_symbol\n)\n\npromoter_map = promoter_map[\n promoter_map[\"gene_symbol\"].isin(gene_order)\n].copy()\n\nenhancer_map = enhancer_map[\n enhancer_map[\"gene_symbol\"].isin(gene_order)\n].copy()\n\ngene_to_idx = {\n gene: i\n for i, gene in enumerate(gene_order)\n}\n\npeak_to_idx = {\n peak: i\n for i, peak in enumerate(peak_cols)\n}\n\npromoter_map[\"gene_idx\"] = promoter_map[\"gene_symbol\"].map(\n gene_to_idx\n)\n\npromoter_map[\"peak_idx\"] = promoter_map[\"peak\"].map(\n peak_to_idx\n)\n\nenhancer_map[\"gene_idx\"] = enhancer_map[\"gene_symbol\"].map(\n gene_to_idx\n)\n\nenhancer_map[\"peak_idx\"] = enhancer_map[\"peak\"].map(\n peak_to_idx\n)\n\npromoter_map = promoter_map.dropna(\n subset=[\"gene_idx\", \"peak_idx\"]\n)\n\nenhancer_map = enhancer_map.dropna(\n subset=[\"gene_idx\", \"peak_idx\"]\n)\n\npromoter_map[\"gene_idx\"] = promoter_map[\"gene_idx\"].astype(np.int64)\npromoter_map[\"peak_idx\"] = promoter_map[\"peak_idx\"].astype(np.int64)\n\nenhancer_map[\"gene_idx\"] = enhancer_map[\"gene_idx\"].astype(np.int64)\nenhancer_map[\"peak_idx\"] = enhancer_map[\"peak_idx\"].astype(np.int64)\n\nn_cells = len(cell_ids)\nn_genes = len(gene_order)\nn_peaks = len(peak_cols)\n\nP = sp.csr_matrix(\n (\n np.ones(len(promoter_map), dtype=np.float32),\n (\n promoter_map[\"peak_idx\"].values,\n promoter_map[\"gene_idx\"].values\n )\n ),\n shape=(n_peaks, n_genes)\n)\n\nE = sp.csr_matrix(\n (\n np.ones(len(enhancer_map), dtype=np.float32),\n (\n enhancer_map[\"peak_idx\"].values,\n enhancer_map[\"gene_idx\"].values\n )\n ),\n shape=(n_peaks, n_genes)\n)\n\npromoter_matrix = np.zeros(\n (n_cells, n_genes),\n dtype=np.float32\n)\n\nenhancer_matrix = np.zeros(\n (n_cells, n_genes),\n dtype=np.float32\n)\n\nchunk_size = 256\n\nfor start in range(0, n_cells, chunk_size):\n end = min(start + chunk_size, n_cells)\n\n X_chunk = X.iloc[\n start:end\n ].to_numpy(\n dtype=np.float32\n )\n\n promoter_matrix[start:end] = X_chunk @ P\n enhancer_matrix[start:end] = X_chunk @ E\n\nadata_multi_test = ad.AnnData(\n X=sp.csr_matrix(promoter_matrix)\n)\n\nadata_multi_test.obs_names = cell_ids\nadata_multi_test.var_names = gene_order\n\nadata_multi_test.layers[\"A_promoter\"] = sp.csr_matrix(\n promoter_matrix\n)\n\nadata_multi_test.layers[\"B_enhancer\"] = sp.csr_matrix(\n enhancer_matrix\n)\n\nprint(adata_multi_test)\nprint(adata_multi_test.layers.keys())\nprint(adata_multi_test.layers[\"A_promoter\"].shape)\nprint(adata_multi_test.layers[\"B_enhancer\"].shape)"},{"cell_type":"code","execution_count":null,"id":"6223efd5-d57f-4338-8a6e-2c49902abf2b","metadata":{},"outputs":[],"source":"metadata = pd.read_csv(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/metadata.csv\"\n)\n\nmetadata[\"cell_id\"] = metadata[\"cell_id\"].astype(str)\nmetadata = metadata.set_index(\"cell_id\")\n\nadata_multi_test.obs = metadata.loc[\n adata_multi_test.obs_names.astype(str)\n].copy()"},{"cell_type":"code","execution_count":null,"id":"652911b7-79a5-4628-bc6a-40b197d822aa","metadata":{},"outputs":[],"source":"output_path = \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/test/test_atac.h5ad\"\n\nadata_multi_test.write_h5ad(output_path)"},{"cell_type":"code","execution_count":null,"id":"20ae2870-8a25-4a24-aa8e-216c4d3b5562","metadata":{},"outputs":[],"source":"import mudata as md\n\ntrain_cite = md.read_h5mu(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/train/train_cite.h5ad\"\n)\n\nprint(train_cite)"},{"cell_type":"code","execution_count":null,"id":"83f3ea96-1e50-4d05-bf93-88f6fd083f31","metadata":{},"outputs":[],"source":"train_cite.var"},{"cell_type":"code","execution_count":null,"id":"670076c3-dbed-4313-94b3-a8c7b2e4eec6","metadata":{},"outputs":[],"source":""},{"cell_type":"markdown","id":"3a874a5f-d4a8-40c5-ae98-0b86290d8e44","metadata":{"jp-MarkdownHeadingCollapsed":true},"source":"# Data merge"},{"cell_type":"code","execution_count":5,"id":"1469fac3-4ce4-47c3-a3c0-0a37742d5639","metadata":{},"outputs":[],"source":"import mudata as md\nimport anndata as ad\n\ntrain_cite = md.read_h5mu(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/train/train_cite.h5ad\"\n)\n\ntrain_multi = ad.read_h5ad(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/train/train_multi.h5ad\"\n)\n\ntest_rna = ad.read_h5ad(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/test/test_rna.h5ad\"\n)\n\ntest_atac = ad.read_h5ad(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/test/test_atac.h5ad\"\n)\n\ngenes = {\n \"train_cite\": set(train_cite[\"rna\"].var_names),\n \"train_multi\": set(train_multi.var_names),\n \"test_rna\": set(test_rna.var_names),\n \"test_atac\": set(test_atac.var_names)\n}\n\nnames = list(genes.keys())\n\nfor i in range(len(names)):\n for j in range(i + 1, len(names)):\n a, b = names[i], names[j]\n\n overlap = genes[a] & genes[b]\n only_a = genes[a] - genes[b]\n only_b = genes[b] - genes[a]\n\n print(f\"{a} vs {b}\")\n print(f\" {a}: {len(genes[a])}\")\n print(f\" {b}: {len(genes[b])}\")\n print(f\" 겹치는 유전자: {len(overlap)}\")\n print(f\" {a}에만 존재: {len(only_a)}\")\n print(f\" {b}에만 존재: {len(only_b)}\")\n print()\n\ncommon_all = set.intersection(*genes.values())\n\nprint(\"=\" * 50)\nprint(f\"4개 데이터셋 공통 유전자: {len(common_all)}\")"},{"cell_type":"code","execution_count":6,"id":"55c96b94-cf1f-423e-ad44-bfe45dd6c370","metadata":{},"outputs":[],"source":"A = train_multi.layers[\"A_promoter\"].toarray()\nB = train_multi.layers[\"B_enhancer\"].toarray()\n\ntrain_multi.layers[\"peak_sum\"] = (1 + A) * (1 + B)\ntrain_multi"},{"cell_type":"code","execution_count":7,"id":"a2fe4096-6e26-4c01-97fd-ac09ca8317a0","metadata":{},"outputs":[],"source":"A = test_atac.layers[\"A_promoter\"].toarray()\nB = test_atac.layers[\"B_enhancer\"].toarray()\n\ntest_atac.layers[\"peak_sum\"] = (1 + A) * (1 + B)\ntest_atac"},{"cell_type":"code","execution_count":8,"id":"bbf57ffb-7f49-46d3-83a4-4291845d7192","metadata":{},"outputs":[],"source":"test_atac = test_atac[:, train_multi.var_names].copy()"},{"cell_type":"code","execution_count":9,"id":"4c4d292f-6f7f-4b63-a2e1-f7589c6053ce","metadata":{},"outputs":[],"source":"print(\"=== train_multi ===\")\nprint(\"Shape:\", train_multi.shape)\nprint(\"Duplicated genes:\", train_multi.var_names.duplicated().sum())\nprint(\n \"Unique duplicated genes:\",\n train_multi.var_names[train_multi.var_names.duplicated(keep=False)].unique().size\n)\n\nprint(\"\\n=== train_cite RNA ===\")\nprint(\"Shape:\", train_cite[\"rna\"].shape)\nprint(\"Duplicated genes:\", train_cite[\"rna\"].var_names.duplicated().sum())\nprint(\n \"Unique duplicated genes:\",\n train_cite[\"rna\"].var_names[\n train_cite[\"rna\"].var_names.duplicated(keep=False)\n ].unique().size\n)\n\nprint(\"\\n=== test_atac ===\")\nprint(\"Shape:\", test_atac.shape)\nprint(\"Duplicated genes:\", test_atac.var_names.duplicated().sum())\nprint(\n \"Unique duplicated genes:\",\n test_atac.var_names[\n test_atac.var_names.duplicated(keep=False)\n ].unique().size\n)\n\nprint(\"\\n=== test_rna ===\")\nprint(\"Shape:\", test_rna.shape)\nprint(\"Duplicated genes:\", test_rna.var_names.duplicated().sum())\nprint(\n \"Unique duplicated genes:\",\n test_rna.var_names[\n test_rna.var_names.duplicated(keep=False)\n ].unique().size\n)"},{"cell_type":"code","execution_count":10,"id":"2dc7a516-0085-4edc-b4a2-23ff3e3a9d78","metadata":{},"outputs":[],"source":"import numpy as np\n\n# CITE RNA: 중복 gene은 첫 번째 occurrence만 유지\ntrain_cite_rna = train_cite[\"rna\"]\n\ncite_keep = ~train_cite_rna.var_names.duplicated(keep=\"first\")\n\ntrain_cite_rna_unique = train_cite_rna[:, cite_keep].copy()\n\n# Multiome: 중복 없음\ntrain_multi_unique = train_multi\n\n# --------------------------------------------------\n# 최종 gene 개수 확인\n# --------------------------------------------------\n\nprint(\"=== Gene count ===\")\nprint(\"train_cite RNA:\", train_cite_rna_unique.n_vars)\nprint(\"test_rna :\", test_rna.n_vars)\n\nprint(\"train_multi :\", train_multi_unique.n_vars)\nprint(\"test_atac :\", test_atac.n_vars)\n\n\n# --------------------------------------------------\n# Gene 순서 확인\n# --------------------------------------------------\n\ncite_order_match = np.array_equal(\n train_cite_rna_unique.var_names.to_numpy(),\n test_rna.var_names.to_numpy()\n)\n\nmulti_order_match = np.array_equal(\n train_multi_unique.var_names.to_numpy(),\n test_atac.var_names.to_numpy()\n)\n\nprint(\"\\n=== Gene count & order check ===\")\nprint(\n \"train_cite RNA ↔ test_rna:\",\n cite_order_match\n)\n\nprint(\n \"train_multi ↔ test_atac:\",\n multi_order_match\n)\n\n\n# --------------------------------------------------\n# 개수와 순서가 모두 일치하는지 최종 확인\n# --------------------------------------------------\n\nassert train_cite_rna_unique.n_vars == test_rna.n_vars\nassert train_multi_unique.n_vars == test_atac.n_vars\n\nassert cite_order_match\nassert multi_order_match\n\nprint(\"\\nAll gene features match.\")"},{"cell_type":"code","execution_count":null,"id":"b0827e56-94c8-4a36-a471-3af88adfb025","metadata":{},"outputs":[],"source":""},{"cell_type":"markdown","id":"54d89029-6802-420d-a134-3a7cfbcbb51a","metadata":{},"source":"# Multimodal diffusion modeling"},{"cell_type":"code","execution_count":11,"id":"ad36d510-274e-46be-a629-bb47d914a4ea","metadata":{},"outputs":[],"source":"import math\nimport itertools\nimport numpy as np\nimport torch\nimport os\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom torch.utils.data import Dataset, DataLoader\n\nprint(\"=\" * 60)\nprint(\"PyTorch version:\", torch.__version__)\nprint(\"CUDA available:\", torch.cuda.is_available())\nprint(\"CUDA version:\", torch.version.cuda)\nprint(\"GPU count:\", torch.cuda.device_count())\n\nif torch.cuda.is_available():\n print(\"GPU:\", torch.cuda.get_device_name(0))\n print(\"Current GPU:\", torch.cuda.current_device())\n print(\n \"GPU memory:\",\n f\"{torch.cuda.get_device_properties(0).total_memory / 1024**3:.2f} GB\"\n )\n device = torch.device(\"cuda\")\nelse:\n print(\"GPU: None\")\n device = torch.device(\"cpu\")\n\nprint(\"Using device:\", device)\nprint(\"=\" * 60)"},{"cell_type":"code","execution_count":12,"id":"8ca76696-1f80-4787-9710-3c55e05239d0","metadata":{},"outputs":[],"source":"import os\nimport math\nimport itertools\nimport numpy as np\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom torch.utils.data import Dataset, DataLoader\n\ntrain_cite_rna = train_cite[\"rna\"]\ncite_keep = ~train_cite_rna.var_names.duplicated(keep=\"first\")\ntrain_cite_rna_unique = train_cite_rna[:, cite_keep].copy()\ntrain_multi_unique = train_multi\n\nprint(\"=== Gene count ===\")\nprint(\"train_cite RNA original:\", train_cite_rna.n_vars)\nprint(\"train_cite RNA unique :\", train_cite_rna_unique.n_vars)\nprint(\"test_rna :\", test_rna.n_vars)\nprint(\"train_multi :\", train_multi_unique.n_vars)\nprint(\"test_atac :\", test_atac.n_vars)\n\ncite_order_match = np.array_equal(\n train_cite_rna_unique.var_names.to_numpy(),\n test_rna.var_names.to_numpy()\n)\n\nmulti_order_match = np.array_equal(\n train_multi_unique.var_names.to_numpy(),\n test_atac.var_names.to_numpy()\n)\n\nprint(\"\\n=== Gene count & order check ===\")\nprint(\"train_cite RNA unique ↔ test_rna:\", cite_order_match)\nprint(\"train_multi ↔ test_atac:\", multi_order_match)\n\nassert train_cite_rna_unique.n_vars == test_rna.n_vars\nassert train_multi_unique.n_vars == test_atac.n_vars\nassert cite_order_match\nassert multi_order_match\n\nprint(\"\\nAll gene features match.\")\n\ndevice = torch.device(\n \"cuda\" if torch.cuda.is_available() else \"cpu\"\n)\n\nmulti_gene_dim = train_multi_unique.layers[\"rna\"].shape[1]\ncite_gene_dim = train_cite_rna_unique.shape[1]\nprotein_dim = train_cite[\"protein\"].shape[1]\n\nlatent_dim = 128\nprotein_latent_dim = 64\nhidden_dim = 512\ntime_dim = 64\nnum_diffusion_steps = 200\n\nbatch_size = 128\nlr = 1e-4\nepochs = 50\npatience = 10\n\n\nclass SinusoidalTimeEmbedding(nn.Module):\n\n def __init__(self, dim):\n super().__init__()\n self.dim = dim\n\n def forward(self, t):\n half = self.dim // 2\n emb = math.log(10000) / (half - 1)\n\n emb = torch.exp(\n torch.arange(\n half,\n device=t.device\n ) * -emb\n )\n\n emb = (\n t.float().unsqueeze(1)\n * emb.unsqueeze(0)\n )\n\n return torch.cat(\n [emb.sin(), emb.cos()],\n dim=1\n )\n\n\nclass Encoder(nn.Module):\n\n def __init__(self, input_dim, latent_dim):\n super().__init__()\n\n self.net = nn.Sequential(\n nn.Linear(input_dim, hidden_dim),\n nn.GELU(),\n nn.LayerNorm(hidden_dim),\n nn.Linear(hidden_dim, hidden_dim),\n nn.GELU(),\n nn.Linear(hidden_dim, latent_dim)\n )\n\n def forward(self, x):\n return self.net(x)\n\n\nclass Decoder(nn.Module):\n\n def __init__(self, latent_dim, output_dim):\n super().__init__()\n\n self.net = nn.Sequential(\n nn.Linear(latent_dim, hidden_dim),\n nn.GELU(),\n nn.LayerNorm(hidden_dim),\n nn.Linear(hidden_dim, hidden_dim),\n nn.GELU(),\n nn.Linear(hidden_dim, output_dim)\n )\n\n def forward(self, z):\n return self.net(z)\n\n\nclass ConditionalDenoiser(nn.Module):\n\n def __init__(\n self,\n latent_dim,\n time_dim\n ):\n super().__init__()\n\n self.time_embedding = nn.Sequential(\n SinusoidalTimeEmbedding(time_dim),\n nn.Linear(time_dim, time_dim),\n nn.GELU()\n )\n\n input_dim = (\n latent_dim +\n latent_dim +\n time_dim\n )\n\n self.net = nn.Sequential(\n nn.Linear(input_dim, hidden_dim),\n nn.GELU(),\n nn.LayerNorm(hidden_dim),\n nn.Linear(hidden_dim, hidden_dim),\n nn.GELU(),\n nn.Linear(hidden_dim, latent_dim)\n )\n\n def forward(\n self,\n x_t,\n condition,\n t\n ):\n t_emb = self.time_embedding(t)\n\n x = torch.cat(\n [\n x_t,\n condition,\n t_emb\n ],\n dim=1\n )\n\n return self.net(x)\n\n\nclass ProteinConditionalDenoiser(nn.Module):\n\n def __init__(\n self,\n protein_latent_dim,\n rna_latent_dim,\n time_dim\n ):\n super().__init__()\n\n self.time_embedding = nn.Sequential(\n SinusoidalTimeEmbedding(time_dim),\n nn.Linear(time_dim, time_dim),\n nn.GELU()\n )\n\n input_dim = (\n protein_latent_dim +\n rna_latent_dim +\n time_dim\n )\n\n self.net = nn.Sequential(\n nn.Linear(input_dim, hidden_dim),\n nn.GELU(),\n nn.LayerNorm(hidden_dim),\n nn.Linear(hidden_dim, hidden_dim),\n nn.GELU(),\n nn.Linear(hidden_dim, protein_latent_dim)\n )\n\n def forward(\n self,\n x_t,\n rna_condition,\n t\n ):\n t_emb = self.time_embedding(t)\n\n x = torch.cat(\n [\n x_t,\n rna_condition,\n t_emb\n ],\n dim=1\n )\n\n return self.net(x)\n\n\nclass DiffusionSchedule:\n\n def __init__(\n self,\n num_steps,\n device\n ):\n self.num_steps = num_steps\n\n self.beta = torch.linspace(\n 1e-4,\n 0.02,\n num_steps,\n device=device\n )\n\n self.alpha = 1.0 - self.beta\n\n self.alpha_bar = torch.cumprod(\n self.alpha,\n dim=0\n )\n\n def q_sample(\n self,\n x0,\n t,\n noise=None\n ):\n if noise is None:\n noise = torch.randn_like(x0)\n\n alpha_bar_t = (\n self.alpha_bar[t]\n .unsqueeze(1)\n )\n\n x_t = (\n torch.sqrt(alpha_bar_t) * x0 +\n torch.sqrt(1.0 - alpha_bar_t) * noise\n )\n\n return x_t, noise\n\n\nclass MultimodalDiffusionModel(nn.Module):\n\n def __init__(\n self,\n multi_gene_dim,\n cite_gene_dim,\n protein_dim\n ):\n super().__init__()\n\n self.atac_encoder = Encoder(\n multi_gene_dim,\n latent_dim\n )\n\n self.multi_rna_encoder = Encoder(\n multi_gene_dim,\n latent_dim\n )\n\n self.cite_rna_encoder = Encoder(\n cite_gene_dim,\n latent_dim\n )\n\n self.protein_encoder = Encoder(\n protein_dim,\n protein_latent_dim\n )\n\n self.multi_rna_decoder = Decoder(\n latent_dim,\n multi_gene_dim\n )\n\n self.cite_rna_decoder = Decoder(\n latent_dim,\n cite_gene_dim\n )\n\n self.protein_decoder = Decoder(\n protein_latent_dim,\n protein_dim\n )\n\n self.atac_to_rna_diffusion = ConditionalDenoiser(\n latent_dim,\n time_dim\n )\n\n self.rna_to_protein_diffusion = ProteinConditionalDenoiser(\n protein_latent_dim,\n latent_dim,\n time_dim\n )\n\n def encode_atac(self, x):\n return self.atac_encoder(x)\n\n def encode_multi_rna(self, x):\n return self.multi_rna_encoder(x)\n\n def encode_cite_rna(self, x):\n return self.cite_rna_encoder(x)\n\n def encode_protein(self, x):\n return self.protein_encoder(x)\n\n\nclass MultiomeDataset(Dataset):\n\n def __init__(\n self,\n peak,\n rna\n ):\n self.peak = peak\n self.rna = rna\n\n def __len__(self):\n return self.peak.shape[0]\n\n def __getitem__(self, idx):\n\n peak = self.peak[idx]\n rna = self.rna[idx]\n\n if hasattr(peak, \"toarray\"):\n peak = peak.toarray().ravel()\n\n if hasattr(rna, \"toarray\"):\n rna = rna.toarray().ravel()\n\n peak = np.asarray(\n peak,\n dtype=np.float32\n )\n\n rna = np.asarray(\n rna,\n dtype=np.float32\n )\n\n return (\n torch.from_numpy(peak),\n torch.from_numpy(rna)\n )\n\n\nclass CiteDataset(Dataset):\n\n def __init__(\n self,\n rna,\n protein\n ):\n self.rna = rna\n self.protein = protein\n\n def __len__(self):\n return self.rna.shape[0]\n\n def __getitem__(self, idx):\n\n rna = self.rna[idx]\n protein = self.protein[idx]\n\n if hasattr(rna, \"toarray\"):\n rna = rna.toarray().ravel()\n\n if hasattr(protein, \"toarray\"):\n protein = protein.toarray().ravel()\n\n rna = np.asarray(\n rna,\n dtype=np.float32\n )\n\n protein = np.asarray(\n protein,\n dtype=np.float32\n )\n\n return (\n torch.from_numpy(rna),\n torch.from_numpy(protein)\n )\n\n\nmulti_peak = train_multi_unique.layers[\"peak_sum\"]\nmulti_rna = train_multi_unique.layers[\"rna\"]\n\ncite_rna = train_cite_rna_unique.X\ncite_protein = train_cite[\"protein\"].X\n\nprint(\"\\n=== Training data ===\")\nprint(\"Multiome peak:\", multi_peak.shape)\nprint(\"Multiome RNA:\", multi_rna.shape)\nprint(\"CITE RNA:\", cite_rna.shape)\nprint(\"CITE protein:\", cite_protein.shape)\n\n\nmulti_dataset = MultiomeDataset(\n multi_peak,\n multi_rna\n)\n\ncite_dataset = CiteDataset(\n cite_rna,\n cite_protein\n)\n\n\nmulti_loader = DataLoader(\n multi_dataset,\n batch_size=batch_size,\n shuffle=True,\n num_workers=4,\n pin_memory=True,\n drop_last=True,\n persistent_workers=True\n)\n\ncite_loader = DataLoader(\n cite_dataset,\n batch_size=batch_size,\n shuffle=True,\n num_workers=4,\n pin_memory=True,\n drop_last=True,\n persistent_workers=True\n)\n\n\nmodel = MultimodalDiffusionModel(\n multi_gene_dim=multi_gene_dim,\n cite_gene_dim=cite_gene_dim,\n protein_dim=protein_dim\n).to(device)\n\n\nschedule = DiffusionSchedule(\n num_diffusion_steps,\n device\n)\n\n\noptimizer = torch.optim.AdamW(\n model.parameters(),\n lr=lr,\n weight_decay=1e-5\n)\n\n\ndef multiome_loss(\n model,\n peak,\n rna\n):\n\n peak = peak.to(\n device,\n non_blocking=True\n )\n\n rna = rna.to(\n device,\n non_blocking=True\n )\n\n z_atac = model.encode_atac(peak)\n z_rna = model.encode_multi_rna(rna)\n\n t = torch.randint(\n 0,\n num_diffusion_steps,\n (rna.shape[0],),\n device=device\n )\n\n z_noisy, noise = schedule.q_sample(\n z_rna,\n t\n )\n\n noise_pred = model.atac_to_rna_diffusion(\n z_noisy,\n z_atac,\n t\n )\n\n diffusion_loss = F.mse_loss(\n noise_pred,\n noise\n )\n\n rna_reconstructed = model.multi_rna_decoder(\n z_rna\n )\n\n reconstruction_loss = F.mse_loss(\n rna_reconstructed,\n rna\n )\n\n loss = (\n diffusion_loss +\n 0.1 * reconstruction_loss\n )\n\n return (\n loss,\n diffusion_loss,\n reconstruction_loss\n )\n\n\ndef cite_loss(\n model,\n rna,\n protein\n):\n\n rna = rna.to(\n device,\n non_blocking=True\n )\n\n protein = protein.to(\n device,\n non_blocking=True\n )\n\n z_rna = model.encode_cite_rna(rna)\n z_protein = model.encode_protein(protein)\n\n t = torch.randint(\n 0,\n num_diffusion_steps,\n (protein.shape[0],),\n device=device\n )\n\n z_noisy, noise = schedule.q_sample(\n z_protein,\n t\n )\n\n noise_pred = model.rna_to_protein_diffusion(\n z_noisy,\n z_rna,\n t\n )\n\n diffusion_loss = F.mse_loss(\n noise_pred,\n noise\n )\n\n protein_reconstructed = model.protein_decoder(\n z_protein\n )\n\n reconstruction_loss = F.mse_loss(\n protein_reconstructed,\n protein\n )\n\n loss = (\n diffusion_loss +\n 0.1 * reconstruction_loss\n )\n\n return (\n loss,\n diffusion_loss,\n reconstruction_loss\n )\n\n\nmulti_iterator = itertools.cycle(\n multi_loader\n)\n\ncite_iterator = itertools.cycle(\n cite_loader\n)\n\n\ntotal_losses = []\nmultiome_losses = []\ncite_losses = []\n\nbest_loss = float(\"inf\")\nbest_epoch = 0\npatience_counter = 0\n\n\nfor epoch in range(epochs):\n\n model.train()\n\n epoch_loss = 0.0\n epoch_multi = 0.0\n epoch_cite = 0.0\n\n steps = max(\n len(multi_loader),\n len(cite_loader)\n )\n\n for step in range(steps):\n\n multi_batch = next(\n multi_iterator\n )\n\n cite_batch = next(\n cite_iterator\n )\n\n peak, rna_multi = multi_batch\n rna_cite, protein = cite_batch\n\n optimizer.zero_grad(\n set_to_none=True\n )\n\n loss_multi, diff_multi, rec_multi = multiome_loss(\n model,\n peak,\n rna_multi\n )\n\n loss_cite, diff_cite, rec_cite = cite_loss(\n model,\n rna_cite,\n protein\n )\n\n loss = (\n loss_multi +\n loss_cite\n )\n\n loss.backward()\n\n torch.nn.utils.clip_grad_norm_(\n model.parameters(),\n 1.0\n )\n\n optimizer.step()\n\n epoch_loss += loss.item()\n epoch_multi += loss_multi.item()\n epoch_cite += loss_cite.item()\n\n epoch_loss /= steps\n epoch_multi /= steps\n epoch_cite /= steps\n\n total_losses.append(epoch_loss)\n multiome_losses.append(epoch_multi)\n cite_losses.append(epoch_cite)\n\n print(\n f\"Epoch {epoch + 1:03d} | \"\n f\"Total {epoch_loss:.6f} | \"\n f\"Multiome {epoch_multi:.6f} | \"\n f\"CITE {epoch_cite:.6f}\"\n )\n\n if epoch_loss < best_loss:\n\n best_loss = epoch_loss\n best_epoch = epoch + 1\n patience_counter = 0\n\n torch.save(\n model.state_dict(),\n os.path.join(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/Diffusion\",\n \"unified_diffusion_model_best.pt\"\n )\n )\n\n else:\n\n patience_counter += 1\n\n if patience_counter >= patience:\n\n print(\n f\"Early stopping at epoch {epoch + 1}\"\n )\n\n print(\n f\"Best epoch: {best_epoch}\"\n )\n\n print(\n f\"Best total loss: {best_loss:.6f}\"\n )\n\n break\n\n\ntorch.save(\n model.state_dict(),\n os.path.join(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/Diffusion\",\n \"unified_diffusion_model.pt\"\n )\n)\n\n\n@torch.no_grad()\ndef sample_rna_latent(\n model,\n z_atac\n):\n\n model.eval()\n\n batch_size_current = z_atac.shape[0]\n\n z = torch.randn(\n batch_size_current,\n latent_dim,\n device=device\n )\n\n for step in reversed(\n range(num_diffusion_steps)\n ):\n\n t = torch.full(\n (batch_size_current,),\n step,\n device=device,\n dtype=torch.long\n )\n\n beta_t = schedule.beta[step]\n alpha_t = schedule.alpha[step]\n alpha_bar_t = schedule.alpha_bar[step]\n\n noise_pred = model.atac_to_rna_diffusion(\n z,\n z_atac,\n t\n )\n\n z = (\n 1.0 / torch.sqrt(alpha_t)\n ) * (\n z -\n (\n beta_t /\n torch.sqrt(\n 1.0 - alpha_bar_t\n )\n ) * noise_pred\n )\n\n if step > 0:\n\n noise = torch.randn_like(z)\n\n z = (\n z +\n torch.sqrt(beta_t) * noise\n )\n\n return z\n\n\n@torch.no_grad()\ndef sample_protein_latent(\n model,\n z_rna\n):\n\n model.eval()\n\n batch_size_current = z_rna.shape[0]\n\n z = torch.randn(\n batch_size_current,\n protein_latent_dim,\n device=device\n )\n\n for step in reversed(\n range(num_diffusion_steps)\n ):\n\n t = torch.full(\n (batch_size_current,),\n step,\n device=device,\n dtype=torch.long\n )\n\n beta_t = schedule.beta[step]\n alpha_t = schedule.alpha[step]\n alpha_bar_t = schedule.alpha_bar[step]\n\n noise_pred = model.rna_to_protein_diffusion(\n z,\n z_rna,\n t\n )\n\n z = (\n 1.0 / torch.sqrt(alpha_t)\n ) * (\n z -\n (\n beta_t /\n torch.sqrt(\n 1.0 - alpha_bar_t\n )\n ) * noise_pred\n )\n\n if step > 0:\n\n noise = torch.randn_like(z)\n\n z = (\n z +\n torch.sqrt(beta_t) * noise\n )\n\n return z\n\n\n@torch.no_grad()\ndef predict_rna_from_atac(\n peak\n):\n\n if hasattr(peak, \"toarray\"):\n peak = peak.toarray()\n\n peak = torch.tensor(\n np.asarray(\n peak,\n dtype=np.float32\n ),\n device=device\n )\n\n z_atac = model.encode_atac(\n peak\n )\n\n z_rna = sample_rna_latent(\n model,\n z_atac\n )\n\n rna = model.multi_rna_decoder(\n z_rna\n )\n\n return rna.cpu().numpy()\n\n\n@torch.no_grad()\ndef predict_protein_from_rna(\n rna\n):\n\n if hasattr(rna, \"toarray\"):\n rna = rna.toarray()\n\n rna = torch.tensor(\n np.asarray(\n rna,\n dtype=np.float32\n ),\n device=device\n )\n\n z_rna = model.encode_cite_rna(\n rna\n )\n\n z_protein = sample_protein_latent(\n model,\n z_rna\n )\n\n protein = model.protein_decoder(\n z_protein\n )\n\n return protein.cpu().numpy()\n\n\n@torch.no_grad()\ndef predict_from_atac(\n peak\n):\n\n if hasattr(peak, \"toarray\"):\n peak = peak.toarray()\n\n peak = torch.tensor(\n np.asarray(\n peak,\n dtype=np.float32\n ),\n device=device\n )\n\n z_atac = model.encode_atac(\n peak\n )\n\n z_rna = sample_rna_latent(\n model,\n z_atac\n )\n\n rna = model.multi_rna_decoder(\n z_rna\n )\n\n z_protein = sample_protein_latent(\n model,\n z_rna\n )\n\n protein = model.protein_decoder(\n z_protein\n )\n\n return (\n rna.cpu().numpy(),\n protein.cpu().numpy()\n )"},{"cell_type":"code","execution_count":11,"id":"9961b74a-3807-4600-8508-8ff0abb1d330","metadata":{},"outputs":[],"source":"import os\nimport matplotlib.pyplot as plt\nimport torch\n\nsave_dir = \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/Diffusion\"\nos.makedirs(save_dir, exist_ok=True)\n\ntorch.save(\n model.state_dict(),\n os.path.join(save_dir, \"unified_diffusion_model.pt\")\n)\n\nepochs = range(1, 51)\n\ntotal_losses = [\n 0.807554, 0.373368, 0.332057, 0.308909, 0.294314,\n 0.281139, 0.273483, 0.268358, 0.263471, 0.259960,\n 0.257332, 0.254955, 0.253008, 0.251682, 0.249867,\n 0.248614, 0.247936, 0.247016, 0.246345, 0.245966,\n 0.244969, 0.244344, 0.244296, 0.243466, 0.243119,\n 0.242382, 0.242341, 0.241777, 0.241481, 0.240982,\n 0.240541, 0.240589, 0.240313, 0.240028, 0.239808,\n 0.239695, 0.239228, 0.239116, 0.238853, 0.238454,\n 0.238575, 0.238040, 0.238185, 0.237862, 0.237632,\n 0.237694, 0.237413, 0.237325, 0.237317, 0.237112\n]\n\nmultiome_losses = [\n 0.399912, 0.225686, 0.219658, 0.215587, 0.213098,\n 0.208748, 0.206747, 0.205170, 0.204008, 0.202967,\n 0.202140, 0.201194, 0.200676, 0.200428, 0.199500,\n 0.199013, 0.198944, 0.198814, 0.198453, 0.198447,\n 0.197906, 0.197783, 0.197786, 0.197577, 0.197140,\n 0.197196, 0.197235, 0.196765, 0.196805, 0.196609,\n 0.196402, 0.196596, 0.196123, 0.196208, 0.196317,\n 0.196151, 0.195969, 0.195927, 0.195829, 0.195740,\n 0.195714, 0.195576, 0.195401, 0.195467, 0.195311,\n 0.195365, 0.195259, 0.195254, 0.195219, 0.195271\n]\n\ncite_losses = [\n 0.407642, 0.147682, 0.112399, 0.093322, 0.081216,\n 0.072391, 0.066736, 0.063188, 0.059464, 0.056994,\n 0.055192, 0.053761, 0.052332, 0.051254, 0.050368,\n 0.049601, 0.048992, 0.048203, 0.047892, 0.047519,\n 0.047063, 0.046561, 0.046511, 0.045888, 0.045979,\n 0.045187, 0.045106, 0.045012, 0.044676, 0.044373,\n 0.044140, 0.043994, 0.044190, 0.043821, 0.043492,\n 0.043544, 0.043259, 0.043189, 0.043025, 0.042714,\n 0.042861, 0.042465, 0.042784, 0.042396, 0.042322,\n 0.042330, 0.042154, 0.042071, 0.042097, 0.041841\n]\n\nplt.figure(figsize=(10, 8))\n\nplt.plot(\n epochs,\n total_losses,\n marker=\"o\",\n label=\"Total Loss\"\n)\n\nplt.plot(\n epochs,\n multiome_losses,\n marker=\"o\",\n label=\"Multiome Loss\"\n)\n\nplt.plot(\n epochs,\n cite_losses,\n marker=\"o\",\n label=\"CITE Loss\"\n)\n\nplt.xlabel(\"Epoch\")\nplt.ylabel(\"Loss\")\nplt.title(\"Training Loss\")\nplt.xticks(list(epochs))\nplt.legend()\nplt.grid(True, alpha=0.3)\nplt.tight_layout()\n\nplt.savefig(\n os.path.join(save_dir, \"training_loss.png\"),\n dpi=300,\n bbox_inches=\"tight\"\n)\n\nplt.show()"},{"cell_type":"code","execution_count":14,"id":"c10a83e2-eb51-4377-9d2a-0e7cf04b8de8","metadata":{},"outputs":[],"source":"import inspect\n\nprint(inspect.signature(MultimodalDiffusionModel))"},{"cell_type":"code","execution_count":47,"id":"e7387117-8ebc-40be-942c-e52a5cd812b1","metadata":{},"outputs":[],"source":"print(\"RNA\")\nprint(z_rna.mean().item())\nprint(z_rna.std().item())\nprint(z_rna.min().item())\nprint(z_rna.max().item())\n\nprint(\"\\nProtein\")\nprint(z_protein_denoised.mean().item())\nprint(z_protein_denoised.std().item())\nprint(z_protein_denoised.min().item())\nprint(z_protein_denoised.max().item())\n\nrna_var = z_rna.detach().cpu().numpy().var(axis=0)\nprotein_var = z_protein_denoised.detach().cpu().numpy().var(axis=0)\n\nprint(\"RNA latent variance\")\nprint(\"mean:\", rna_var.mean())\nprint(\"median:\", np.median(rna_var))\nprint(\"max:\", rna_var.max())\n\nprint(\"\\nProtein latent variance\")\nprint(\"mean:\", protein_var.mean())\nprint(\"median:\", np.median(protein_var))\nprint(\"max:\", protein_var.max())"},{"cell_type":"code","execution_count":65,"id":"5efa4837-b9bf-4abd-a6db-bab379ef4911","metadata":{},"outputs":[],"source":"import torch\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom matplotlib.patches import Patch\n\nwith torch.no_grad():\n z_rna = model.cite_rna_encoder(rna)\n\n z_protein = z_rna.clone()\n\n protein_noise = torch.randn_like(z_protein[..., :64])\n\n z_protein = z_protein[..., :64] + protein_noise\n\nrna_values = z_rna.cpu().numpy().ravel()\nprotein_noise_values = protein_noise.cpu().numpy().ravel()\n\nplt.figure(figsize=(5, 5))\n\nplt.hist(\n rna_values,\n bins=50,\n density=True,\n alpha=0.5,\n color=\"yellow\"\n)\n\nplt.hist(\n protein_noise_values,\n bins=50,\n density=True,\n alpha=0.5,\n color=\"red\"\n)\n\nlegend_elements = [\n Patch(facecolor=\"yellow\", alpha=0.5, label=\"ATAC → RNA\"),\n Patch(facecolor=\"red\", alpha=0.5, label=\"RNA → Protein\")\n]\n\nplt.xlabel(\"Latent / Noise Value\")\nplt.ylabel(\"Density\")\nplt.title(\"Diffusion Noise latent comparison\")\nplt.legend(handles=legend_elements)\n\nplt.tight_layout()\nplt.show()"},{"cell_type":"code","execution_count":56,"id":"c6113232-a1bc-40c3-8be7-616d06845e8f","metadata":{},"outputs":[],"source":"@torch.no_grad()\ndef sample_protein_latent(model, z_rna):\n model.eval()\n batch_size_current = z_rna.shape[0]\n\n z = torch.randn(\n batch_size_current,\n protein_latent_dim,\n device=device\n )\n\n for step in reversed(range(num_diffusion_steps)):\n t = torch.full(\n (batch_size_current,),\n step,\n device=device,\n dtype=torch.long\n )\n\n beta_t = schedule.beta[step]\n alpha_t = schedule.alpha[step]\n alpha_bar_t = schedule.alpha_bar[step]\n\n noise_pred = model.rna_to_protein_diffusion(\n z,\n z_rna,\n t\n )\n\n z = (\n 1.0 / torch.sqrt(alpha_t)\n ) * (\n z -\n (\n beta_t /\n torch.sqrt(1.0 - alpha_bar_t)\n ) * noise_pred\n )\n\n if step > 0:\n noise = torch.randn_like(z)\n z = z + torch.sqrt(beta_t) * noise\n\n return z"},{"cell_type":"code","execution_count":66,"id":"45ec55c2-9fd7-4ca3-bb4c-dec5f0019167","metadata":{},"outputs":[],"source":"import torch\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom matplotlib.patches import Patch\n\nwith torch.no_grad():\n if isinstance(peak, torch.Tensor):\n peak_tensor = peak.to(device)\n else:\n if hasattr(peak, \"toarray\"):\n peak = peak.toarray()\n peak_tensor = torch.tensor(\n np.asarray(peak, dtype=np.float32),\n device=device\n )\n\n if isinstance(rna, torch.Tensor):\n rna_tensor = rna.to(device)\n else:\n if hasattr(rna, \"toarray\"):\n rna = rna.toarray()\n rna_tensor = torch.tensor(\n np.asarray(rna, dtype=np.float32),\n device=device\n )\n\n z_atac = model.encode_atac(peak_tensor)\n z_rna_denoised = sample_rna_latent(model, z_atac)\n\n z_rna = model.cite_rna_encoder(rna_tensor)\n z_protein_denoised = sample_protein_latent(model, z_rna)\n\nrna_denoised_values = z_rna_denoised.cpu().numpy().ravel()\nprotein_denoised_values = z_protein_denoised.cpu().numpy().ravel()\n\nplt.figure(figsize=(5, 5))\n\nplt.hist(\n rna_denoised_values,\n bins=50,\n density=True,\n alpha=0.5,\n color=\"yellow\"\n)\n\nplt.hist(\n protein_denoised_values,\n bins=50,\n density=True,\n alpha=0.5,\n color=\"red\"\n)\n\nlegend_elements = [\n Patch(facecolor=\"yellow\", alpha=0.5, label=\"ATAC → RNA\"),\n Patch(facecolor=\"red\", alpha=0.5, label=\"RNA → Protein\")\n]\n\nplt.xlabel(\"Denoised Latent Value\")\nplt.ylabel(\"Density\")\nplt.title(\"Diffusion Denoise latent comparison\")\nplt.legend(handles=legend_elements)\n\nplt.tight_layout()\nplt.show()"},{"cell_type":"code","execution_count":61,"id":"70d38616-9f62-4b49-a27a-7bdc2f55ed21","metadata":{},"outputs":[],"source":"import torch\nimport numpy as np\nimport matplotlib.pyplot as plt\n\n@torch.no_grad()\ndef trace_rna_diffusion(model, z_atac):\n model.eval()\n\n batch_size_current = z_atac.shape[0]\n\n z = torch.randn(\n batch_size_current,\n latent_dim,\n device=device\n )\n\n steps = []\n noise_variances = []\n denoise_variances = []\n\n for step in reversed(range(num_diffusion_steps)):\n t = torch.full(\n (batch_size_current,),\n step,\n device=device,\n dtype=torch.long\n )\n\n beta_t = schedule.beta[step]\n alpha_t = schedule.alpha[step]\n alpha_bar_t = schedule.alpha_bar[step]\n\n noise_pred = model.atac_to_rna_diffusion(\n z,\n z_atac,\n t\n )\n\n z_denoised = (\n 1.0 / torch.sqrt(alpha_t)\n ) * (\n z -\n (\n beta_t /\n torch.sqrt(1.0 - alpha_bar_t)\n ) * noise_pred\n )\n\n noise_variances.append(\n z.var(dim=0).mean().item()\n )\n\n denoise_variances.append(\n z_denoised.var(dim=0).mean().item()\n )\n\n steps.append(step)\n\n z = z_denoised\n\n if step > 0:\n noise = torch.randn_like(z)\n z = z + torch.sqrt(beta_t) * noise\n\n return (\n np.array(steps),\n np.array(noise_variances),\n np.array(denoise_variances)\n )\n\n\n@torch.no_grad()\ndef trace_protein_diffusion(model, z_rna):\n model.eval()\n\n batch_size_current = z_rna.shape[0]\n\n z = torch.randn(\n batch_size_current,\n protein_latent_dim,\n device=device\n )\n\n steps = []\n noise_variances = []\n denoise_variances = []\n\n for step in reversed(range(num_diffusion_steps)):\n t = torch.full(\n (batch_size_current,),\n step,\n device=device,\n dtype=torch.long\n )\n\n beta_t = schedule.beta[step]\n alpha_t = schedule.alpha[step]\n alpha_bar_t = schedule.alpha_bar[step]\n\n noise_pred = model.rna_to_protein_diffusion(\n z,\n z_rna,\n t\n )\n\n z_denoised = (\n 1.0 / torch.sqrt(alpha_t)\n ) * (\n z -\n (\n beta_t /\n torch.sqrt(1.0 - alpha_bar_t)\n ) * noise_pred\n )\n\n noise_variances.append(\n z.var(dim=0).mean().item()\n )\n\n denoise_variances.append(\n z_denoised.var(dim=0).mean().item()\n )\n\n steps.append(step)\n\n z = z_denoised\n\n if step > 0:\n noise = torch.randn_like(z)\n z = z + torch.sqrt(beta_t) * noise\n\n return (\n np.array(steps),\n np.array(noise_variances),\n np.array(denoise_variances)\n )\n\n\nwith torch.no_grad():\n if isinstance(peak, torch.Tensor):\n peak_tensor = peak.to(device)\n else:\n if hasattr(peak, \"toarray\"):\n peak = peak.toarray()\n peak_tensor = torch.tensor(\n np.asarray(peak, dtype=np.float32),\n device=device\n )\n\n if isinstance(rna, torch.Tensor):\n rna_tensor = rna.to(device)\n else:\n if hasattr(rna, \"toarray\"):\n rna = rna.toarray()\n rna_tensor = torch.tensor(\n np.asarray(rna, dtype=np.float32),\n device=device\n )\n\n z_atac = model.encode_atac(peak_tensor)\n z_rna = model.cite_rna_encoder(rna_tensor)\n\nrna_steps, rna_noise_var, rna_denoise_var = trace_rna_diffusion(\n model,\n z_atac\n)\n\nprotein_steps, protein_noise_var, protein_denoise_var = trace_protein_diffusion(\n model,\n z_rna\n)"},{"cell_type":"code","execution_count":67,"id":"c7116b4f-527f-431b-a485-ed2317f8b573","metadata":{},"outputs":[],"source":"plt.figure(figsize=(5, 5))\n\nplt.plot(\n rna_steps,\n rna_noise_var,\n label=\"ATAC → RNA\"\n)\n\nplt.plot(\n protein_steps,\n protein_noise_var,\n label=\"RNA → Protein\"\n)\n\nplt.xlabel(\"Diffusion Step\")\nplt.ylabel(\"Noise Variance\")\nplt.title(\"Noise Variance Across Diffusion Steps\")\nplt.legend()\nplt.tight_layout()\nplt.show()"},{"cell_type":"code","execution_count":68,"id":"057afafb-cf30-4fa1-b518-dc10b4a845dd","metadata":{},"outputs":[],"source":"plt.figure(figsize=(5, 5))\n\nplt.plot(\n rna_steps,\n rna_denoise_var,\n label=\"ATAC → RNA\"\n)\n\nplt.plot(\n protein_steps,\n protein_denoise_var,\n label=\"RNA → Protein\"\n)\n\nplt.xlabel(\"Diffusion Step\")\nplt.ylabel(\"Denoised Latent Variance\")\nplt.title(\"Denoised Latent Variance Across Diffusion Steps\")\nplt.gca().invert_xaxis()\nplt.legend()\nplt.tight_layout()\nplt.show()"},{"cell_type":"code","execution_count":78,"id":"566ebea3-4eaa-4cc1-8b1e-854f8c76c880","metadata":{},"outputs":[],"source":"import numpy as np\nimport torch\n\nbatch_size_eval = 128\n\nrna_squared_errors = []\nprotein_squared_errors = []\n\nmodel.eval()\n\nwith torch.no_grad():\n\n peak_data = train_multi.layers[\"peak_sum\"]\n rna_gt_data = train_multi.layers[\"rna\"]\n\n for start in range(0, train_multi.n_obs, batch_size_eval):\n end = min(start + batch_size_eval, train_multi.n_obs)\n\n peak_batch = peak_data[start:end]\n rna_gt_batch = rna_gt_data[start:end]\n\n if hasattr(peak_batch, \"toarray\"):\n peak_batch = peak_batch.toarray()\n\n if hasattr(rna_gt_batch, \"toarray\"):\n rna_gt_batch = rna_gt_batch.toarray()\n\n peak_batch = torch.tensor(\n np.asarray(peak_batch, dtype=np.float32),\n device=device\n )\n\n rna_gt_batch = torch.tensor(\n np.asarray(rna_gt_batch, dtype=np.float32),\n device=device\n )\n\n z_atac = model.encode_atac(peak_batch)\n z_rna_pred = sample_rna_latent(model, z_atac)\n rna_pred = model.multi_rna_decoder(z_rna_pred)\n\n mse = (\n (rna_pred - rna_gt_batch) ** 2\n ).mean(dim=1)\n\n rna_squared_errors.append(\n mse.cpu().numpy()\n )\n\n del peak_batch, rna_gt_batch\n del z_atac, z_rna_pred, rna_pred, mse\n\n torch.cuda.empty_cache()\n\n cite_rna_data = train_cite_rna_unique.X\n protein_gt_data = train_cite[\"protein\"].X\n\n for start in range(0, train_cite_rna_unique.n_obs, batch_size_eval):\n end = min(start + batch_size_eval, train_cite_rna_unique.n_obs)\n\n rna_batch = cite_rna_data[start:end]\n protein_gt_batch = protein_gt_data[start:end]\n\n if hasattr(rna_batch, \"toarray\"):\n rna_batch = rna_batch.toarray()\n\n if hasattr(protein_gt_batch, \"toarray\"):\n protein_gt_batch = protein_gt_batch.toarray()\n\n rna_batch = torch.tensor(\n np.asarray(rna_batch, dtype=np.float32),\n device=device\n )\n\n protein_gt_batch = torch.tensor(\n np.asarray(protein_gt_batch, dtype=np.float32),\n device=device\n )\n\n z_rna = model.encode_cite_rna(rna_batch)\n z_protein_pred = sample_protein_latent(model, z_rna)\n protein_pred = model.protein_decoder(z_protein_pred)\n\n mse = (\n (protein_pred - protein_gt_batch) ** 2\n ).mean(dim=1)\n\n protein_squared_errors.append(\n mse.cpu().numpy()\n )\n\n del rna_batch, protein_gt_batch\n del z_rna, z_protein_pred, protein_pred, mse\n\n torch.cuda.empty_cache()\n\nrna_squared_errors = np.concatenate(rna_squared_errors)\nprotein_squared_errors = np.concatenate(protein_squared_errors)\n\nrna_rmse = np.sqrt(np.mean(rna_squared_errors))\nprotein_rmse = np.sqrt(np.mean(protein_squared_errors))\n\nprint(f\"ATAC → RNA RMSE: {rna_rmse:.6f}\")\nprint(f\"RNA → Protein RMSE: {protein_rmse:.6f}\")"},{"cell_type":"code","execution_count":83,"id":"99b96335-3f3d-4a2a-922d-9c56097cbdab","metadata":{},"outputs":[],"source":"import matplotlib.pyplot as plt\n\nmodels = [\n \"ATAC → RNA\",\n \"RNA → Protein\"\n]\n\nrmse_values = [\n rna_rmse,\n protein_rmse\n]\n\nfig, ax = plt.subplots(figsize=(5, 5))\n\nbars = ax.bar(\n models,\n rmse_values\n)\n\nax.set_ylabel(\"RMSE\")\nax.set_title(\"Model RMSE comparison\")\n\nax.tick_params(\n axis=\"both\",\n labelsize=13\n)\nax.set_ylabel(\"RMSE\", fontsize=18)\nfor bar, value in zip(bars, rmse_values):\n ax.text(\n bar.get_x() + bar.get_width() / 2,\n bar.get_height(),\n f\"{value:.4f}\",\n ha=\"center\",\n va=\"bottom\",\n fontsize=12\n )\n\nplt.tight_layout()\nplt.show()"},{"cell_type":"code","execution_count":null,"id":"f862f3f1-dee9-47e5-9472-158a7c62e16e","metadata":{},"outputs":[],"source":""},{"cell_type":"markdown","id":"d0268682-e0de-4805-8bf7-851186c22133","metadata":{},"source":"# Diffusion Predict"},{"cell_type":"code","execution_count":12,"id":"ee3c6a78-0b6c-4c66-b306-5d79c8332655","metadata":{},"outputs":[],"source":"@torch.no_grad()\ndef predict_rna_from_atac(model, peak):\n z_atac = model.atac_encoder(peak)\n\n z_rna = sample_rna_latent(\n model,\n z_atac\n )\n\n rna = model.multi_rna_decoder(z_rna)\n\n return rna\n\n\n@torch.no_grad()\ndef predict_protein_from_rna(model, rna):\n z_rna = model.cite_rna_encoder(rna)\n\n z_protein = sample_protein_latent(\n model,\n z_rna\n )\n\n protein = model.protein_decoder(z_protein)\n\n return protein"},{"cell_type":"code","execution_count":13,"id":"300684cf-11fd-4a42-8749-09041c49158f","metadata":{},"outputs":[],"source":"import os\nimport numpy as np\nimport torch\nimport anndata as ad\n\ndevice = torch.device(\n \"cuda:0\" if torch.cuda.is_available() else \"cpu\"\n)\n\nsave_dir = \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/Diffusion\"\n\nmodel_path = os.path.join(\n save_dir,\n \"unified_diffusion_model.pt\"\n)\n\nmodel.load_state_dict(\n torch.load(\n model_path,\n map_location=device,\n weights_only=True\n )\n)\n\nmodel = model.to(device)\nmodel.eval()\n\nprint(\"Model loaded\")\nprint(\n \"Device:\",\n next(model.parameters()).device\n)\n\n\ncite_keep = ~train_cite[\"rna\"].var_names.duplicated(\n keep=\"first\"\n)\n\ntrain_cite_rna_unique = train_cite[\"rna\"][\n :,\n cite_keep\n].copy()\n\ntest_rna_keep = ~test_rna.var_names.duplicated(\n keep=\"first\"\n)\n\ntest_rna = test_rna[\n :,\n test_rna_keep\n].copy()\n\n\ntrain_multi_genes = train_multi.var_names.to_numpy()\n\ntrain_cite_genes = train_cite_rna_unique.var_names.to_numpy()\n\n\ntest_atac = test_atac[\n :,\n train_multi_genes\n].copy()\n\ntest_rna = test_rna[\n :,\n train_cite_genes\n].copy()\n\n\nassert np.array_equal(\n test_atac.var_names,\n train_multi.var_names\n)\n\nassert np.array_equal(\n test_rna.var_names,\n train_cite_rna_unique.var_names\n)\n\n\nprint(\n \"ATAC feature order:\",\n np.array_equal(\n test_atac.var_names,\n train_multi.var_names\n )\n)\n\nprint(\n \"RNA feature order:\",\n np.array_equal(\n test_rna.var_names,\n train_cite_rna_unique.var_names\n )\n)\n\nprint(\n \"Multiome features:\",\n train_multi.n_vars\n)\n\nprint(\n \"CITE RNA features:\",\n train_cite_rna_unique.n_vars\n)\n\nprint(\n \"Test RNA features:\",\n test_rna.n_vars\n)\n\n\nbatch_size = 128\n\n\npeak = test_atac.layers[\"peak_sum\"]\n\nif hasattr(peak, \"toarray\"):\n peak = peak.toarray()\n\npeak = np.asarray(\n peak,\n dtype=np.float32\n)\n\n\npred_rna_list = []\n\nwith torch.no_grad():\n\n for start in range(\n 0,\n len(test_atac),\n batch_size\n ):\n\n end = min(\n start + batch_size,\n len(test_atac)\n )\n\n peak_batch = torch.tensor(\n peak[start:end],\n dtype=torch.float32,\n device=device\n )\n\n z_atac = model.encode_atac(\n peak_batch\n )\n\n z_rna = sample_rna_latent(\n model,\n z_atac\n )\n\n pred_rna_batch = model.multi_rna_decoder(\n z_rna\n )\n\n pred_rna_list.append(\n pred_rna_batch.cpu().numpy()\n )\n\n\npred_rna = np.concatenate(\n pred_rna_list,\n axis=0\n)\n\n\npred_rna_adata = ad.AnnData(\n X=pred_rna,\n obs=test_atac.obs.copy(),\n var=train_multi.var.copy()\n)\n\nprint(\n \"Predicted RNA:\",\n pred_rna_adata.shape\n)\n\n\nrna = test_rna.X\n\nif hasattr(rna, \"toarray\"):\n rna = rna.toarray()\n\nrna = np.asarray(\n rna,\n dtype=np.float32\n)\n\n\npred_protein_list = []\n\nwith torch.no_grad():\n\n for start in range(\n 0,\n len(test_rna),\n batch_size\n ):\n\n end = min(\n start + batch_size,\n len(test_rna)\n )\n\n rna_batch = torch.tensor(\n rna[start:end],\n dtype=torch.float32,\n device=device\n )\n\n z_rna = model.encode_cite_rna(\n rna_batch\n )\n\n z_protein = sample_protein_latent(\n model,\n z_rna\n )\n\n pred_protein_batch = model.protein_decoder(\n z_protein\n )\n\n pred_protein_list.append(\n pred_protein_batch.cpu().numpy()\n )\n\n\npred_protein = np.concatenate(\n pred_protein_list,\n axis=0\n)\n\n\npred_protein_adata = ad.AnnData(\n X=pred_protein,\n obs=test_rna.obs.copy(),\n var=train_cite[\"protein\"].var.copy()\n)\n\n\nprint(\n \"Predicted Protein:\",\n pred_protein_adata.shape\n)\n\n\nrna_path = os.path.join(\n save_dir,\n \"test_predicted_rna.h5ad\"\n)\n\nprotein_path = os.path.join(\n save_dir,\n \"test_predicted_protein.h5ad\"\n)\n\n\npred_rna_adata.write_h5ad(\n rna_path\n)\n\npred_protein_adata.write_h5ad(\n protein_path\n)\n\n\nprint(\"Saved:\")\nprint(rna_path)\nprint(protein_path)"},{"cell_type":"code","execution_count":33,"id":"eba83225-4f2f-46a4-91a5-5c6744ac20b1","metadata":{},"outputs":[],"source":"import pandas as pd\n\nevaluation_ids = pd.read_csv(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/evaluation_ids.csv\"\n)\n\nsample_submission = pd.read_csv(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/sample_submission.csv\"\n)"},{"cell_type":"code","execution_count":37,"id":"b2fd25fd-c732-4b59-80f1-2c1460e08525","metadata":{},"outputs":[],"source":"import pandas as pd\nimport requests\nimport time\n\nevaluation_ids = pd.read_csv(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/evaluation_ids.csv\"\n)\n\ngtf_path = \"/home/Data_Drive_8TB/kykim/7. Kaggle/Single-Cell_Perturbations/baseline/gencode.v44.annotation.gtf.gz\"\n\ngtf = pd.read_csv(\n gtf_path,\n sep=\"\\t\",\n comment=\"#\",\n header=None,\n names=[\n \"chr\", \"source\", \"feature\", \"start\", \"end\",\n \"score\", \"strand\", \"frame\", \"attribute\"\n ]\n)\n\ngene_gtf = gtf[gtf[\"feature\"] == \"gene\"].copy()\n\ngene_gtf[\"gene_id\"] = (\n gene_gtf[\"attribute\"]\n .str.extract(r'gene_id \"([^\"]+)\"')[0]\n .str.split(\".\")\n .str[0]\n)\n\ngene_gtf[\"gene_name\"] = (\n gene_gtf[\"attribute\"]\n .str.extract(r'gene_name \"([^\"]+)\"')[0]\n)\n\ngene_map = (\n gene_gtf\n .dropna(subset=[\"gene_id\", \"gene_name\"])\n .drop_duplicates(\"gene_id\")\n .set_index(\"gene_id\")[\"gene_name\"]\n .to_dict()\n)\n\nmask = evaluation_ids[\"gene_id\"].astype(str).str.match(\n r\"^ENSG\\d+\",\n na=False\n)\n\nensg_ids = (\n evaluation_ids.loc[mask, \"gene_id\"]\n .astype(str)\n .str.extract(r\"^(ENSG\\d+)\", expand=False)\n)\n\nmapped_genes = ensg_ids.map(gene_map)\n\nevaluation_ids.loc[mask, \"gene_id\"] = mapped_genes\n\nremaining_mask = (\n evaluation_ids[\"gene_id\"]\n .astype(str)\n .str.match(r\"^ENSG\\d+$\", na=False)\n)\n\nremaining_ids = (\n evaluation_ids.loc[remaining_mask, \"gene_id\"]\n .astype(str)\n .unique()\n)\n\nprint(\"GTF에서 변환되지 않은 ENSG:\", len(remaining_ids))\n\nensembl_map = {}\n\nurl = \"https://rest.ensembl.org/lookup/id\"\n\nfor i in range(0, len(remaining_ids), 1000):\n batch = remaining_ids[i:i + 1000]\n\n response = requests.post(\n url,\n params={\n \"species\": \"homo_sapiens\",\n \"object_type\": \"gene\"\n },\n headers={\n \"Content-Type\": \"application/json\",\n \"Accept\": \"application/json\"\n },\n json={\n \"ids\": batch.tolist()\n }\n )\n\n response.raise_for_status()\n\n result = response.json()\n\n for gene_id, info in result.items():\n if isinstance(info, dict):\n gene_name = info.get(\"display_name\")\n\n if gene_name:\n ensembl_map[gene_id] = gene_name\n\n print(\n f\"{min(i + 1000, len(remaining_ids))} / {len(remaining_ids)}\"\n )\n\n time.sleep(0.1)\n\nevaluation_ids.loc[remaining_mask, \"gene_id\"] = (\n evaluation_ids.loc[remaining_mask, \"gene_id\"]\n .map(ensembl_map)\n .fillna(evaluation_ids.loc[remaining_mask, \"gene_id\"])\n)\n\nunmapped = evaluation_ids.loc[\n evaluation_ids[\"gene_id\"].astype(str).str.match(\n r\"^ENSG\\d+$\",\n na=False\n ),\n \"gene_id\"\n].unique()\n\nprint(\"최종 변환되지 않은 ENSG:\", len(unmapped))\nprint(unmapped[:20])\n\nprint(evaluation_ids.loc[mask, [\"gene_id\"]].tail())"},{"cell_type":"code","execution_count":43,"id":"ff65b1d3-a1f8-4869-9bcd-42dde9629638","metadata":{},"outputs":[],"source":"cell_ids = evaluation_ids[\"cell_id\"].astype(str)\ngene_ids = evaluation_ids[\"gene_id\"].astype(str)\n\nrna_cells = pd.Series(\n np.arange(pred_rna_adata.n_obs),\n index=pred_rna_adata.obs_names.astype(str)\n)\nrna_genes = pd.Series(\n np.arange(pred_rna_adata.n_vars),\n index=pred_rna_adata.var_names.astype(str)\n)\n\nprotein_cells = pd.Series(\n np.arange(pred_protein_adata.n_obs),\n index=pred_protein_adata.obs_names.astype(str)\n)\nprotein_genes = pd.Series(\n np.arange(pred_protein_adata.n_vars),\n index=pred_protein_adata.var_names.astype(str)\n)\n\nrna_r = cell_ids.map(rna_cells)\nrna_c = gene_ids.map(rna_genes)\n\nprotein_r = cell_ids.map(protein_cells)\nprotein_c = gene_ids.map(protein_genes)\n\nrna_valid = rna_r.notna() & rna_c.notna()\nprotein_valid = protein_r.notna() & protein_c.notna()\n\ntarget = np.zeros(len(evaluation_ids), dtype=np.float32)\n\nif rna_valid.any():\n r = rna_r[rna_valid].to_numpy(dtype=np.int64)\n c = rna_c[rna_valid].to_numpy(dtype=np.int64)\n target[rna_valid.to_numpy()] = np.asarray(\n pred_rna_adata.X[r, c]\n ).ravel()\n\nif protein_valid.any():\n protein_only = (~rna_valid) & protein_valid\n r = protein_r[protein_only].to_numpy(dtype=np.int64)\n c = protein_c[protein_only].to_numpy(dtype=np.int64)\n target[protein_only.to_numpy()] = np.asarray(\n pred_protein_adata.X[r, c]\n ).ravel()\n\nsample_submission[\"target\"] = target\n\nmatched = rna_valid | protein_valid\ntotal = len(evaluation_ids)\n\nprint(\"전체:\", total)\nprint(\"RNA 매칭:\", rna_valid.sum(), f\"({rna_valid.sum() / total * 100:.2f}%)\")\nprint(\"Protein 매칭:\", protein_valid.sum(), f\"({protein_valid.sum() / total * 100:.2f}%)\")\nprint(\"매칭 성공:\", matched.sum(), f\"({matched.sum() / total * 100:.2f}%)\")\nprint(\"없는 데이터:\", (~matched).sum(), f\"({(~matched).sum() / total * 100:.2f}%)\")"},{"cell_type":"code","execution_count":47,"id":"8fb6ea18-f06b-4e13-88bc-d6b40bd78139","metadata":{},"outputs":[],"source":"eval_genes = set(evaluation_ids[\"gene_id\"].astype(str))\n\nrna_genes_set = set(pred_rna_adata.var_names.astype(str))\nprotein_genes_set = set(pred_protein_adata.var_names.astype(str))\n\nrna_eval_only = sorted(eval_genes - rna_genes_set)\nrna_object_only = sorted(rna_genes_set - eval_genes)\n\nprotein_eval_only = sorted(eval_genes - protein_genes_set)\nprotein_object_only = sorted(protein_genes_set - eval_genes)\n\nprint(\"===== RNA =====\")\nprint(\"evaluation_ids에만 있는 gene:\", len(rna_eval_only))\nprint(rna_eval_only)\n\nprint(\"\\npred_rna_adata에만 있는 gene:\", len(rna_object_only))\nprint(rna_object_only)\n\nprint(\"\\n===== Protein =====\")\nprint(\"evaluation_ids에만 있는 gene:\", len(protein_eval_only))\nprint(protein_eval_only)\n\nprint(\"\\npred_protein_adata에만 있는 gene:\", len(protein_object_only))\nprint(protein_object_only)"},{"cell_type":"code","execution_count":53,"id":"237e98a3-cf1e-4a68-8db1-dda5cfba7c3e","metadata":{},"outputs":[],"source":"sample_submission"},{"cell_type":"code","execution_count":54,"id":"0b342d31-4b02-47a4-b07b-9838e4663863","metadata":{},"outputs":[],"source":"B = pd.read_csv(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/sample_submission.csv\"\n)\n\nB"},{"cell_type":"code","execution_count":44,"id":"733f3813-e45b-40d8-8902-dfcfac867ac1","metadata":{},"outputs":[],"source":"sample_submission"},{"cell_type":"code","execution_count":45,"id":"78761139-1d50-43bd-9f83-66690a361fc4","metadata":{},"outputs":[],"source":"sample_submission.to_csv(\n \"/home/Data_Drive_8TB/kykim/7. Kaggle/Multimodal_Single-Cell_Integration/submission/sample_submission.csv\",\n index=False\n)"},{"cell_type":"code","execution_count":52,"id":"1c95d288-a321-4767-9679-939f8fe22ed8","metadata":{},"outputs":[],"source":"sample_submission"},{"cell_type":"code","execution_count":null,"id":"b8e68688-6c2e-4b24-bfed-7dbb88da3ace","metadata":{},"outputs":[],"source":""},{"cell_type":"code","execution_count":null,"id":"1124bcd8-8e57-4c23-8232-8ba90057dcee","metadata":{},"outputs":[],"source":""},{"cell_type":"code","execution_count":null,"id":"1e2ff2f0-ebfa-467b-9680-7f5c69f2a552","metadata":{},"outputs":[],"source":""},{"cell_type":"code","execution_count":null,"id":"8f09bd0c-7d7d-4904-9a7d-a2f73266bda4","metadata":{},"outputs":[],"source":""},{"cell_type":"code","execution_count":null,"id":"e6bfe0a4-7e8d-45a9-a4f4-54a33fb96466","metadata":{},"outputs":[],"source":""},{"cell_type":"code","execution_count":null,"id":"7d2af1ae-13c8-4155-98a3-11f850ddcc1f","metadata":{},"outputs":[],"source":""},{"cell_type":"code","execution_count":null,"id":"2477874e-bbca-4bf8-9436-3c65c62ed0b8","metadata":{},"outputs":[],"source":""},{"cell_type":"code","execution_count":null,"id":"0b440088-4f0a-4533-b3ff-60677f488017","metadata":{},"outputs":[],"source":""},{"cell_type":"code","execution_count":null,"id":"b0c93505-32fe-41d3-b01c-059691db2ac3","metadata":{},"outputs":[],"source":""},{"cell_type":"code","execution_count":null,"id":"fa0ba9e7-302e-4fce-8d00-0d068f40a8b8","metadata":{},"outputs":[],"source":""}],"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":"3.11.15"}},"nbformat":4,"nbformat_minor":5}