{"id":66281,"date":"2026-10-23T09:05:45","date_gmt":"2026-10-23T01:05:45","guid":{"rendered":"https:\/\/www.biocloudservice.com\/wordpress\/?p=66281"},"modified":"2026-10-23T09:05:45","modified_gmt":"2026-10-23T01:05:45","slug":"cellbender-gpu-droplet-memory-planning","status":"publish","type":"post","link":"https:\/\/www.biocloudservice.com\/wordpress\/cellbender-gpu-droplet-memory-planning\/","title":{"rendered":"CellBender \u8dd1\u5230\u6700\u540e\u624d\u7206\u5185\u5b58\uff1f\u771f\u6b63\u653e\u5927\u8d44\u6e90\u7684\u53ef\u80fd\u4e0d\u662f\u7ec6\u80de\u6570\uff0c\u800c\u662f\u6db2\u6ef4\u6570"},"content":{"rendered":"
\n

\u5355\u7ec6\u80de RNA-seq \u505a\u5b8c Cell Ranger \u540e\uff0c\u5e38\u4f1a\u770b\u5230\u67d0\u4e9b\u201c\u672c\u4e0d\u8be5\u8868\u8fbe\u201d\u7684\u57fa\u56e0\u6563\u843d\u5728\u591a\u79cd\u7ec6\u80de\u91cc\u3002\u73af\u5883 RNA\u3001\u7a7a\u6db2\u6ef4\u548c\u6761\u5f62\u7801\u4ea4\u6362\u53ef\u80fd\u628a\u6280\u672f\u566a\u58f0\u5e26\u8fdb\u5dee\u5f02\u5206\u6790\uff0cCellBender \u7684 remove-background<\/code> \u56e0\u6b64\u7ecf\u5e38\u88ab\u52a0\u5165\u9884\u5904\u7406\u6d41\u7a0b\u3002<\/p>\n

\u4f46\u5b83\u5e76\u4e0d\u662f\u628a\u4e00\u4e2a\u8fc7\u6ee4\u9608\u503c\u5957\u5728\u77e9\u9635\u4e0a\uff0c\u800c\u662f\u7528\u6df1\u5ea6\u751f\u6210\u6a21\u578b\u4f30\u8ba1\u7ec6\u80de\u6982\u7387\u3001\u80cc\u666f\u8868\u8fbe\u548c\u6c61\u67d3\u6bd4\u4f8b\u3002\u5f88\u591a\u4efb\u52a1\u8bad\u7ec3\u9636\u6bb5\u770b\u4f3c\u6b63\u5e38\uff0c\u5374\u5728\u751f\u6210 posterior \u6216\u8f93\u51fa\u77e9\u9635\u65f6\u7a81\u7136\u7206\u663e\u5b58\u3001\u7206\u5185\u5b58\u3002\u8d44\u6e90\u89c4\u5212\u82e5\u53ea\u62a5\u201c\u7ea6 1 \u4e07\u4e2a\u7ec6\u80de\u201d\uff0c\u5f80\u5f80\u5c11\u7b97\u4e86\u771f\u6b63\u51b3\u5b9a\u8ba1\u7b97\u91cf\u7684\u6db2\u6ef4\u8303\u56f4\u3002<\/p>\n

\"CellBender
CellBender \u8d44\u6e90\u89c4\u5212\uff1a\u7ec6\u80de\u6570\u3001\u6db2\u6ef4\u6570\u3001posterior \u5185\u5b58\u3001GPU \u548c NVMe \u5fc5\u987b\u4e00\u8d77\u8ba1\u7b97\u3002<\/figcaption><\/figure>\n

\u5148\u770b\u4e09\u4e2a\u53ef\u6838\u5bf9\u7684\u6570\u5b57<\/h2>\n

CellBender \u8bba\u6587\u62a5\u544a\uff1a\u5bf9\u5178\u578b\u7684 10x \u6570\u636e\uff0c\u7ea6 5,000\u201430,000 \u4e2a\u7ec6\u80de<\/strong>\uff0c\u4f7f\u7528 NVIDIA T4 \u6216 K80 GPU \u65f6\uff0c\u603b\u8fd0\u884c\u65f6\u95f4\u901a\u5e38\u7ea6 20 \u5206\u949f\u5230 1 \u5c0f\u65f6<\/strong>\uff0c\u5177\u4f53\u53d6\u51b3\u4e8e\u6570\u636e\u89c4\u6a21\u548c\u53c2\u6570\u3002\u5b98\u65b9 v0.3.2 \u6559\u7a0b\u4e2d\uff0c\u5b8c\u6574\u672a\u88c1\u526a\u7684 heart10k \u6570\u636e\u5728 T4 \u4e0a\u7ea6\u9700 30 \u5206\u949f<\/strong>\uff1b\u8bad\u7ec3\u901a\u5e38\u4e3a 150 epochs<\/strong>\u3002<\/p>\n

\u8fd9\u4e9b\u6570\u5b57\u9002\u5408\u505a\u91cf\u7ea7\u53c2\u8003\uff0c\u4e0d\u662f\u4ea4\u4ed8\u627f\u8bfa\u3002\u540c\u6837 1 \u4e07\u4e2a\u7ec6\u80de\uff0c\u5982\u679c\u4e00\u4e2a\u6837\u672c\u7eb3\u5165 20,000 \u4e2a\u6db2\u6ef4\uff0c\u53e6\u4e00\u4e2a\u7eb3\u5165 70,000 \u4e2a\u6db2\u6ef4\uff0c\u540e\u8005\u9700\u8981\u8bc4\u4f30\u7684\u7a7a\u6db2\u6ef4\u548c posterior \u8303\u56f4\u660e\u663e\u66f4\u5927\u3002\u5b98\u65b9 v0.3.2 \u7684\u66f4\u65b0\u91cd\u70b9\u4e4b\u4e00\uff0c\u6b63\u662f\u964d\u4f4e --total-droplets-included<\/code> \u5f88\u5927\u65f6\u7684 CPU \u5cf0\u503c\u5185\u5b58\u3002<\/p>\n

\n\n\n\n\n\n\n\n\n
\u73af\u8282<\/th>\n\u8d44\u6e90\u4e3b\u8981\u82b1\u5728\u54ea\u91cc<\/th>\n\u53ef\u89c2\u5bdf\u4fe1\u53f7<\/th>\n\u5148\u68c0\u67e5\u4ec0\u4e48<\/th>\n<\/tr>\n<\/thead>\n
\u8bfb\u53d6\u539f\u59cb\u77e9\u9635\u4e0e\u5148\u9a8c\u4f30\u8ba1<\/td>\nCPU \u5185\u5b58\u3001\u78c1\u76d8\u8bfb\u53d6<\/td>\n\u65e5\u5fd7\u505c\u5728\u52a0\u8f7d\u6216\u5148\u9a8c\u9636\u6bb5<\/td>\n\u8f93\u5165\u662f\u5426\u4e3a raw \u77e9\u9635\u3001\u6587\u4ef6\u662f\u5426\u5b8c\u6574<\/td>\n<\/tr>\n
150 \u8f6e\u5de6\u53f3\u6a21\u578b\u8bad\u7ec3<\/td>\nGPU \u8ba1\u7b97\u4e0e\u663e\u5b58<\/td>\nloss \u66f2\u7ebf\u3001GPU \u5229\u7528\u7387\u3001epoch \u901f\u5ea6<\/td>\nCUDA\/PyTorch\u3001batch size\u3001\u662f\u5426\u65ad\u70b9\u7eed\u8dd1<\/td>\n<\/tr>\n
posterior \u751f\u6210<\/td>\nGPU\/CPU \u5cf0\u503c\u5185\u5b58<\/td>\n\u8bad\u7ec3\u7ed3\u675f\u540e\u624d OOM<\/td>\n\u7eb3\u5165\u6db2\u6ef4\u6570\u3001posterior batch \u53c2\u6570\u3001\u7248\u672c<\/td>\n<\/tr>\n
\u8f93\u51fa\u4e0e\u62a5\u544a<\/td>\nCPU \u5185\u5b58\u3001\u4e34\u65f6\u76d8<\/td>\nh5\u3001PDF\u3001HTML \u62a5\u544a\u5199\u5165\u5931\u8d25<\/td>\n\u4e34\u65f6\u7a7a\u95f4\u3001\u5199\u6743\u9650\u3001\u662f\u5426\u4fdd\u7559 checkpoint<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n

\u4e3a\u4ec0\u4e48\u6db2\u6ef4\u6570\u6bd4\u201c\u6700\u540e\u4fdd\u7559\u591a\u5c11\u7ec6\u80de\u201d\u66f4\u5173\u952e<\/h2>\n

CellBender \u9700\u8981\u540c\u65f6\u89c2\u5bdf\u542b\u7ec6\u80de\u6db2\u6ef4\u548c\u7a7a\u6db2\u6ef4\uff0c\u624d\u80fd\u5b66\u4e60\u73af\u5883\u80cc\u666f\u3002--total-droplets-included<\/code> \u63a7\u5236\u7eb3\u5165\u6a21\u578b\u7684\u6db2\u6ef4\u603b\u6570\uff0c\u800c\u4e0d\u662f\u6700\u7ec8\u7ec6\u80de\u6570\u3002\u628a\u8fd9\u4e2a\u53c2\u6570\u8bbe\u5f97\u8fc7\u5c0f\uff0c\u53ef\u80fd\u622a\u65ad\u7a7a\u6db2\u6ef4\u80cc\u666f\uff1b\u8bbe\u5f97\u8fc7\u5927\uff0c\u5219\u4f1a\u589e\u52a0\u8bad\u7ec3\u548c posterior \u751f\u6210\u7684\u5185\u5b58\u538b\u529b\u3002<\/p>\n

v0.3.0 \u8d77\uff0c\u5b98\u65b9\u8bf4\u660e --expected-cells<\/code> \u4e0e --total-droplets-included<\/code> \u901a\u5e38\u4e0d\u5fc5\u624b\u52a8\u8bbe\u7f6e\uff0c\u81ea\u52a8\u884c\u4e3a\u5df2\u6709\u6539\u8fdb\u3002\u5de5\u7a0b\u4e0a\u66f4\u7a33\u59a5\u7684\u505a\u6cd5\uff0c\u662f\u5148\u68c0\u67e5 UMI rank \u56fe\u548c\u81ea\u52a8\u5148\u9a8c\uff0c\u518d\u51b3\u5b9a\u662f\u5426\u8986\u76d6\u9ed8\u8ba4\u503c\uff0c\u800c\u4e0d\u662f\u7167\u6284\u53e6\u4e00\u6279\u6570\u636e\u7684 5,000\/20,000 \u914d\u7f6e\u3002<\/p>\n

\u5f53\u524d\u4ee3\u7801\u4e2d\u7684\u9ed8\u8ba4\u7eb3\u5165\u6db2\u6ef4\u6570\u4e3a 25,000<\/strong>\uff0c\u9ed8\u8ba4\u8bad\u7ec3 batch size \u4e3a 128<\/strong>\u3002\u8fd9\u4e9b\u9ed8\u8ba4\u503c\u5e76\u4e0d\u610f\u5473\u7740\u6240\u6709\u6837\u672c\u90fd\u5e94\u76f8\u540c\uff1b\u8d85\u8f7d\u5b9e\u9a8c\u3001\u7ec6\u80de\u6570\u7279\u522b\u591a\u3001\u7a7a\u6db2\u6ef4\u5e73\u53f0\u4e0d\u6e05\u6670\u65f6\uff0c\u4ecd\u8981\u7ed3\u5408\u62a5\u544a\u548c\u751f\u7269\u5b66\u80cc\u666f\u5224\u65ad\u3002<\/p>\n

\u4e00\u5957\u53ef\u4ee5\u6267\u884c\u7684\u6392\u9519\u987a\u5e8f<\/h2>\n

\u7b2c\u4e00\u6b65\uff0c\u786e\u8ba4\u8f93\u5165\u4f7f\u7528\u672a\u9884\u8fc7\u6ee4\u7684 raw UMI count matrix\u3002\u5b98\u65b9\u6559\u7a0b\u660e\u786e\u63d0\u9192\uff0c\u771f\u5b9e\u6570\u636e\u8fd0\u884c\u524d\u4e0d\u8981\u5148\u505a\u7279\u5f81\u9009\u62e9\u6216\u6761\u5f62\u7801\u7b5b\u9009\uff1b\u5426\u5219\u6a21\u578b\u770b\u4e0d\u5230\u5b8c\u6574\u7684\u7a7a\u6db2\u6ef4\u80cc\u666f\u3002<\/p>\n

\u7b2c\u4e8c\u6b65\uff0c\u56fa\u5b9a\u7248\u672c\u548c\u73af\u5883\u3002CellBender \u5b98\u65b9\u76ee\u524d\u5efa\u8bae\u65b0\u5206\u6790\u4f7f\u7528 v0.4.0 \u6216\u66f4\u9ad8\u7248\u672c\uff1bv0.4.0 \u5c06 Python \u517c\u5bb9\u8303\u56f4\u66f4\u65b0\u5230 3.10\u20143.12\uff0c\u7b97\u6cd5\u8f93\u51fa\u4e0e v0.3.2\u3001v0.3.0 \u4fdd\u6301\u4e00\u81f4\u3002\u4e0d\u8981\u4f7f\u7528\u66fe\u77ed\u6682\u53d1\u5e03\u7684 v0.3.1\uff0c\u5b98\u65b9\u8bf4\u660e\u8be5\u7248\u672c\u5b58\u5728\u4f1a\u7834\u574f\u53bb\u566a\u77e9\u9635\u7684\u91cd\u5927\u9519\u8bef\u3002<\/p>\n

\u7b2c\u4e09\u6b65\uff0c\u7528\u4e00\u4e2a\u4ee3\u8868\u6837\u672c\u8dd1\u57fa\u7ebf\uff0c\u540c\u65f6\u8bb0\u5f55 GPU \u5229\u7528\u7387\u3001\u5cf0\u503c\u663e\u5b58\u3001\u7cfb\u7edf\u5185\u5b58\u3001\u6db2\u6ef4\u6570\u3001\u6bcf\u8f6e\u8017\u65f6\u548c\u4e34\u65f6\u76d8\u5360\u7528\u3002\u4e0b\u9762\u7684\u547d\u4ee4\u4fdd\u7559 checkpoint\uff0c\u4efb\u52a1\u88ab\u62a2\u5360\u6216\u4e2d\u65ad\u540e\u66f4\u5bb9\u6613\u6062\u590d\uff1a<\/p>\n

cellbender remove-background \\\r\n  --cuda \\\r\n  --input raw_feature_bc_matrix.h5 \\\r\n  --output sample_cellbender.h5 \\\r\n  --checkpoint sample_ckpt.tar.gz\r\n<\/code><\/pre>\n
\u5982\u679c\u8bad\u7ec3\u5df2\u7ecf\u7ed3\u675f\uff0c\u5374\u5728 posterior sampling \u9636\u6bb5\u51fa\u73b0 GPU OOM\uff0c\u5b98\u65b9 troubleshooting \u5efa\u8bae\u53ef\u628a --posterior-batch-size<\/code> \u4ece\u9ed8\u8ba4 128<\/strong> \u964d\u5230 64<\/strong>\u3002\u4ee3\u4ef7\u662f\u540e\u5904\u7406\u901f\u5ea6\u53ef\u80fd\u4e0b\u964d\uff0c\u4f46\u901a\u5e38\u6bd4\u91cd\u590d\u5b8c\u6574\u8bad\u7ec3\u66f4\u53ef\u63a7\u3002\u53ea\u8981 checkpoint \u4e0e\u539f\u547d\u4ee4\u53c2\u6570\u4fdd\u7559\u5b8c\u6574\uff0c\u91cd\u65b0\u8fd0\u884c\u65f6\u53ef\u4ee5\u4ece\u65ad\u70b9\u7ee7\u7eed\u3002\u9700\u8981\u6ce8\u610f\uff0cCellBender \u5f53\u524d\u5355\u4e2a\u4efb\u52a1\u53ea\u4f7f\u7528 1 \u5f20 GPU<\/strong>\uff0c\u7ed9\u540c\u4e00\u6837\u672c\u589e\u52a0\u7b2c\u4e8c\u5f20\u5361\u5e76\u4e0d\u80fd\u89e3\u51b3\u663e\u5b58\u4e0d\u8db3\uff1b\u591a\u5361\u66f4\u9002\u5408\u8ba9\u4e0d\u540c\u6837\u672c\u5e76\u884c\uff0c\u540c\u65f6\u4e3a\u6bcf\u4e2a\u4efb\u52a1\u5355\u72ec\u9650\u5236 CPU\u3001\u5185\u5b58\u548c\u4e34\u65f6\u76d8\u3002<\/p>\n
\u82e5 loss \u66f2\u7ebf\u5728\u540e\u671f\u4ecd\u6709\u660e\u663e\u5c16\u5cf0\u6216\u6ca1\u6709\u6536\u655b\uff0c\u5b98\u65b9\u5efa\u8bae\u4f18\u5148\u68c0\u67e5\u62a5\u544a\uff0c\u5e76\u53ef\u80fd\u628a learning rate \u964d\u4f4e\u4e00\u534a\u540e\u91cd\u8dd1\u3002\u8fd9\u4e2a\u95ee\u9898\u5c5e\u4e8e\u4f18\u5316\u7a33\u5b9a\u6027\uff0c\u4e0d\u5e94\u901a\u8fc7\u76f2\u76ee\u589e\u52a0 epochs \u63a9\u76d6\u3002\u6bcf\u6b21\u8c03\u53c2\u53ea\u6539\u53d8\u4e00\u4e2a\u5173\u952e\u53d8\u91cf\uff0c\u4fdd\u5b58\u547d\u4ee4\u3001\u65e5\u5fd7\u548c\u62a5\u544a\uff0c\u624d\u77e5\u9053\u7ed3\u679c\u53d8\u5316\u6765\u81ea\u53c2\u6570\u8fd8\u662f\u6570\u636e\u3002<\/p>\n
\u7b2c\u56db\u6b65\uff0c\u4e0d\u8981\u53ea\u770b\u547d\u4ee4\u662f\u5426\u7ed3\u675f\u3002\u81f3\u5c11\u6253\u5f00 output_report.html<\/code>\u3001UMI count \u56fe\u3001loss \u66f2\u7ebf\u548c metrics \u6587\u4ef6\uff0c\u6838\u5bf9\u81ea\u52a8\u5148\u9a8c\u3001\u7ec6\u80de\u6982\u7387\u7a81\u964d\u4f4d\u7f6e\u3001\u53bb\u9664\u8ba1\u6570\u7684\u57fa\u56e0\u5206\u5e03\u4ee5\u53ca\u8b66\u544a\u4fe1\u606f\u3002\u8f93\u51fa\u77e9\u9635\u80fd\u88ab Seurat \u6216 Scanpy \u8bfb\u53d6\uff0c\u53ea\u80fd\u8bc1\u660e\u683c\u5f0f\u53ef\u7528\uff0c\u4e0d\u80fd\u8bc1\u660e\u53bb\u566a\u5408\u7406\u3002<\/p>\n
\u6279\u91cf\u9879\u76ee\u53ef\u4ee5\u628a metrics \u4e2d\u201c\u627e\u5230\u7ec6\u80de\u6570\u4e0e expected cells \u7684\u6bd4\u503c\u201d\u201c\u975e\u7a7a\u6db2\u6ef4\u5360\u6bd4\u201d\u548c\u201c\u6536\u655b\u6307\u6807\u201d\u7eb3\u5165\u81ea\u52a8\u8d28\u63a7\uff0c\u4f46\u9608\u503c\u4ecd\u5e94\u7ed3\u5408\u6837\u672c\u7c7b\u578b\u8bbe\u5b9a\u3002\u53d1\u73b0\u5f02\u5e38\u65f6\u56de\u770b\u5355\u6837\u672c\u62a5\u544a\uff0c\u4e0d\u8981\u8ba9\u4e00\u6761\u7edf\u4e00\u89c4\u5219\u8986\u76d6\u6240\u6709\u7ec4\u7ec7\u3001\u5efa\u5e93\u6279\u6b21\u548c\u7ec6\u80de\u6d3b\u6027\u5dee\u5f02\u3002<\/p>\n
GPU \u670d\u52a1\u5668\u600e\u6837\u914d\u624d\u4e0d\u4f1a\u201c\u53ea\u5feb\u4e00\u534a\u201d<\/h2>\n
\u5355\u6837\u672c\u8fd0\u884c\u65f6\uff0cGPU \u8d1f\u8d23\u6a21\u578b\u8bad\u7ec3\u548c\u63a8\u65ad\uff0c\u4f46\u539f\u59cb\u77e9\u9635\u52a0\u8f7d\u3001posterior \u6574\u7406\u3001HDF5 \u5199\u5165\u548c\u62a5\u544a\u751f\u6210\u4ecd\u4f1a\u6d88\u8017 CPU \u5185\u5b58\u4e0e\u78c1\u76d8\u3002\u5927\u91cf\u6837\u672c\u5e76\u53d1\u65f6\uff0c\u663e\u5361\u6570\u91cf\u589e\u52a0\u8fd8\u4f1a\u540c\u6b65\u653e\u5927 CPU\u3001\u5185\u5b58\u548c NVMe \u538b\u529b\u3002\u82e5\u6240\u6709\u4efb\u52a1\u628a checkpoint\u3001posterior \u548c\u62a5\u544a\u5199\u8fdb\u540c\u4e00\u4e2a\u62e5\u5835\u7684\u7f51\u7edc\u76ee\u5f55\uff0cGPU \u4e5f\u4f1a\u7b49\u5f85 I\/O\u3002<\/p>\n
\u56e0\u6b64\uff0c\u670d\u52a1\u5668\u89c4\u5212\u5e94\u6309\u201c\u5355\u6837\u672c\u5cf0\u503c \u00d7 \u5e76\u53d1\u6570\u201d\u6838\u7b97\uff1a\u6bcf\u4e2a\u4efb\u52a1\u7ed1\u5b9a GPU\uff0c\u9650\u5236\u5e76\u53d1\uff0c\u7ed9\u7cfb\u7edf\u5185\u5b58\u7559\u4f59\u91cf\uff0c\u5e76\u628a\u4e34\u65f6\u6587\u4ef6\u653e\u5230\u8282\u70b9\u672c\u5730\u9ad8\u901f\u76d8\uff1b\u5b8c\u6210\u540e\u518d\u628a h5\u3001\u62a5\u544a\u3001\u65e5\u5fd7\u548c checkpoint \u5f52\u6863\u3002\u5bf9\u4e8e\u51e0\u5341\u4e2a\u6837\u672c\uff0c\u53ef\u5148\u5206\u5c42\u8bd5\u8dd1\u4f4e\u80cc\u666f\u3001\u4e2d\u7b49\u80cc\u666f\u548c\u9ad8\u80cc\u666f\u6837\u672c\uff0c\u518d\u786e\u5b9a\u961f\u5217\u8d44\u6e90\uff0c\u800c\u4e0d\u662f\u4e00\u6b21\u6027\u5168\u90e8\u63d0\u4ea4\u3002<\/p>\n
\u53bb\u6389\u66f4\u591a counts\uff0c\u4e0d\u4ee3\u8868\u7ed3\u679c\u4e00\u5b9a\u66f4\u597d<\/h2>\n
CellBender \u9002\u7528\u4e8e\u6db2\u6ef4\u578b scRNA-seq\u3001snRNA-seq \u548c CITE-seq \u7684\u7cfb\u7edf\u80cc\u666f\u5efa\u6a21\uff0c\u4f46\u5b83\u4e0d\u80fd\u4fee\u590d\u4e25\u91cd\u5931\u8d25\u7684\u5b9e\u9a8c\u3002\u5b98\u65b9\u6587\u6863\u6307\u51fa\uff1a\u82e5\u7a7a\u6db2\u6ef4\u51e0\u4e4e\u6ca1\u6709 UMI\uff0c\u5de5\u5177\u53ef\u80fd\u51e0\u4e4e\u4e0d\u53bb\u9664\uff1b\u82e5\u80cc\u666f\u9ad8\u5230\u65e0\u6cd5\u8fa8\u8ba4\u7a7a\u6db2\u6ef4\u5e73\u53f0\uff0c\u7b97\u6cd5\u4e5f\u53ef\u80fd\u56f0\u96be\uff0c\u8fd9\u65f6\u5e94\u8ba4\u771f\u8003\u8651\u5b9e\u9a8c\u8d28\u91cf\u672c\u8eab\u3002<\/p>\n
\u53bb\u566a\u540e\u7684\u5dee\u5f02\u57fa\u56e0\u3001\u7ec6\u80de\u6bd4\u4f8b\u548c\u7a00\u6709\u7fa4\u4f53\u9700\u8981\u4e0e\u539f\u59cb\u77e9\u9635\u3001\u6807\u8bb0\u57fa\u56e0\u548c\u6837\u672c\u80cc\u666f\u5bf9\u7167\u3002FPR \u53d8\u5927\u5f80\u5f80\u4f34\u968f\u66f4\u591a\u4fe1\u53f7\u88ab\u79fb\u9664\uff0c\u4e0d\u80fd\u7528\u201c\u53bb\u5f97\u8d8a\u591a\u8d8a\u5e72\u51c0\u201d\u5224\u65ad\u8d28\u91cf\u3002<\/p>\n
\u5982\u679c\u4f60\u7684 CellBender \u4efb\u52a1\u5361\u5728 CUDA \u73af\u5883\u3001posterior \u9636\u6bb5 OOM\u3001\u591a\u6837\u672c\u6392\u961f\u7f13\u6162\uff0c\u6216\u4e0d\u77e5\u9053\u6db2\u6ef4\u6570\u3001GPU\u3001\u7cfb\u7edf\u5185\u5b58\u548c\u4e34\u65f6\u76d8\u5982\u4f55\u642d\u914d\uff0c\u53ef\u4ee5\u51c6\u5907 CellBender \u7248\u672c\u3001\u8f93\u5165\u683c\u5f0f\u3001\u9884\u8ba1\u7ec6\u80de\u6570\u3001UMI rank \u56fe\u3001\u6db2\u6ef4\u6570\u3001\u62a5\u9519\u548c\u6837\u672c\u6570\u91cf\u3002\u4e91\u751f\u4fe1\u4f1a\u5148\u5224\u65ad\u662f\u53c2\u6570\u3001\u73af\u5883\u8fd8\u662f\u8d44\u6e90\u74f6\u9888\uff0c\u518d\u786e\u8ba4\u79df\u7528\u73af\u5883\u81ea\u884c\u8fd0\u884c\u3001\u73af\u5883\u90e8\u7f72\u4ea4\u63a5\u6216\u59d4\u6258\u5206\u6790\u7684\u8303\u56f4\u3002<\/p>\n
\u8054\u7cfb\u4e91\u751f\u4fe1\uff0c\u54a8\u8be2 CellBender GPU \u73af\u5883\u4e0e\u5206\u6790\u65b9\u6848<\/a><\/p>\n
\u53c2\u8003\u8d44\u6599<\/h2>\n