{"id":52490,"date":"2024-09-04T21:41:12","date_gmt":"2024-09-04T13:41:12","guid":{"rendered":"https:\/\/swarma.org\/?p=52490"},"modified":"2024-09-04T21:41:12","modified_gmt":"2024-09-04T13:41:12","slug":"%e5%87%a0%e4%bd%95%e5%9b%be%e7%9a%84%e6%b7%b1%e5%ba%a6%e7%94%9f%e6%88%90%e6%a8%a1%e5%9e%8b%e7%a0%94%e7%a9%b6%e4%b8%a8%e5%87%a0%e4%bd%95%e6%b7%b1%e5%ba%a6%e5%ad%a6%e4%b9%a0%e8%af%bb%e4%b9%a6%e4%bc%9a","status":"publish","type":"post","link":"https:\/\/swarma.org\/?p=52490","title":{"rendered":"\u51e0\u4f55\u56fe\u7684\u6df1\u5ea6\u751f\u6210\u6a21\u578b\u7814\u7a76\u4e28\u51e0\u4f55\u6df1\u5ea6\u5b66\u4e60\u8bfb\u4e66\u4f1a\u00b7\u5468\u4e94\u76f4\u64ad"},"content":{"rendered":"
\n
\n
\n
\n

<\/p>\n<\/section>\n

\n
\n
\n
\n
\n
\n

<\/section>\n<\/section>\n
\n
\n

\u5bfc\u8bed<\/strong><\/p>\n<\/section>\n<\/section>\n

\n

<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
\n
\u5206\u5b50\u76843D\u7ed3\u6784\u89e3\u91ca\u4e86\u5206\u5b50\u5982\u4f55\u4e0e\u751f\u7269\u9776\u70b9\u76f8\u4e92\u4f5c\u7528\uff0c\u8fd9\u5bf9\u4e8e\u7406\u89e3\u836f\u7269\u7684\u6709\u6548\u6027\u548c\u9009\u62e9\u6027\u81f3\u5173\u91cd\u8981\u3002\u76ee\u524d\uff0c\u867d\u7136\u4ee5\u6269\u6563\u6a21\u578b\uff08Diffusion Model\uff0cDMs\uff09\u4e3a\u57fa\u7840\u7684Stable Diffusion\u3001DALL\u00b7E\u3001Sora\u7b49\u6587\u751f\u56fe\u3001\u6587\u751f\u89c6\u9891\u5de5\u5177\uff0c\u5df2\u662f\u751f\u6210\u5f0f\u4eba\u5de5\u667a\u80fd\u4e2d\u4e0eChatGPT\u6bd4\u80a9\u7684\u6740\u624b\u950f\u7ea7\u5e94\u7528\u3002\u4f46\u662f\uff0c\u6269\u6563\u6a21\u578b\u57283D\u5206\u5b50\u751f\u6210\u4e0a\u5374\u4ecd\u6548\u679c\u4e0d\u5c3d\u4eba\u610f\uff0c\u96be\u4ee5\u7a33\u5b9a\u5730\u62df\u5408\u771f\u5b9e\u5206\u5b50\u7ed3\u6784\u7684\u6982\u7387\u5206\u5e03\u3002\u5982\u4f55\u8ba9\u751f\u6210\u5f0fAI\u66f4\u8fdb\u4e00\u6b65\uff0c\u62df\u54083D\u7ed3\u6784\uff0c\u8bbe\u8ba1\u836f\u7269\u5206\u5b50\uff0c\u52a0\u901f\u5065\u5eb7\u533b\u7597\u53d1\u5c55\uff1f<\/span><\/strong><\/section>\n

<\/span><\/strong><\/section>\n
\u300c\u51e0\u4f55\u6df1\u5ea6\u5b66\u4e60\u300d\u8bfb\u4e66\u4f1a<\/span><\/strong><\/a>\u7b2c\u516b\u671f\u5c06\u7531\u6e05\u534e\u5927\u5b66\u535a\u58eb\u751f\u5b8b\u5b87\u8f69\uff0c\u5c06\u5e26\u9886\u793e\u533a\u6210\u5458\u5168\u9762\u5730\u8ba4\u8bc6\u51e0\u4f55\u56fe\u7684\u6df1\u5ea6\u751f\u6210\u6a21\u578b\u7684\u73b0\u6709\u7814\u7a76\uff0c\u4ecb\u7ecd\u5728\u8be5\u9886\u57df\u4e2d\u6240\u53d6\u5f97\u7684\u4e00\u7cfb\u5217\u53d1\u8868\u5728ICML\u3001NeurIPS\u3001ICLR\u4e0a\u7684\u521b\u65b0\u6210\u679cHierDiff\u3001EquiFM\u4ee5\u53caMolCRAFT\uff0c\u6700\u540e\u5c06\u4e0e\u5927\u5bb6\u4e00\u8d77\u63a2\u7d22\u672a\u6765\u7684\u7814\u7a76\u8ba1\u5212\u3002\u8bfb\u4e66\u4f1a\u5c06\u4e8e9\u67086\u65e5\uff08\u672c\u5468\u4e94\uff0919:00-21:00\u8fdb\u884c\uff0c\u6b22\u8fce\u611f\u5174\u8da3\u7684\u670b\u53cb\u53c2\u4e0e\u8ba8\u8bba\u4ea4\u6d41\uff01<\/span><\/strong><\/section>\n<\/section>\n
\n
<\/strong><\/section>\n
<\/mp-common-videosnap><\/section>\n<\/section>\n
\n
<\/strong><\/section>\n<\/section>\n
\n
<\/strong><\/section>\n<\/section>\n
\n
<\/strong><\/section>\n<\/section>\n
\n
<\/strong><\/strong><\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
\n
\n
<\/svg><\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
<\/section>\n<\/section>\n
\n

<\/section>\n<\/section>\n

<\/p>\n


<\/p>\n

<\/h2>\n

\n
\n
\n
\n
\n

<\/section>\n

<\/section>\n
\n
<\/svg><\/section>\n<\/section>\n<\/section>\n
\n

\u5206\u4eab\u5185\u5bb9\u7b80\u4ecb<\/strong><\/span><\/strong><\/strong><\/p>\n<\/section>\n

\n

<\/section>\n

<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n


<\/p>\n

3D \u5206\u5b50\u751f\u6210\u5bf9\u836f\u7269\u53d1\u73b0\u81f3\u5173\u91cd\u8981\uff0c\u9700\u8981\u540c\u65f6\u786e\u5b9a\u5206\u7c7b\u7279\u5f81\uff08\u539f\u5b50\u7c7b\u578b\uff09\u548c\u8fde\u7eed\u7279\u5f81\uff08\u539f\u5b50\u5750\u6807\uff09\u3002\u4f20\u7edf\u7684\u6df1\u5ea6\u751f\u6210\u6a21\u578b\uff0c\u5305\u62ec\u6269\u6563\u6a21\u578b\uff08DMs\uff09\uff0c\u5c3d\u7ba1\u663e\u793a\u51fa\u524d\u666f\uff0c\u4f46\u9762\u4e34\u8bf8\u5982\u4e0d\u7a33\u5b9a\u7684\u6982\u7387\u52a8\u6001\u548c\u4f4e\u6548\u91c7\u6837\u8fc7\u7a0b\u7b49\u91cd\u5927\u6311\u6218\u3002\u4e3a\u4e86\u7a81\u7834 3D \u5206\u5b50\u751f\u6210\u7684\u754c\u9650\u5e76\u89e3\u51b3\u8fd9\u4e9b\u9650\u5236\uff0c\u6211\u4eec\u5c06\u4ecb\u7ecd\u4e00\u5957\u521b\u65b0\u65b9\u6cd5\uff0c\u8d85\u8d8a\u6269\u6563\u6a21\u578b\uff1a<\/span><\/section>\n

<\/span><\/section>\n
\u9996\u5148\uff0c\u5c42\u6b21\u6269\u6563\u6a21\u578b (HierDiff)<\/strong> \u901a\u8fc7\u91c7\u7528\u7c97\u5230\u7ec6\u7684\u7b56\u7565\u89e3\u51b3\u751f\u6210\u5927\u5206\u5b50\u7684\u8d28\u91cf\u95ee\u9898\u3002HierDiff \u901a\u8fc7\u7b49\u53d8\u7684\u6269\u6563\u8fc7\u7a0b\u751f\u6210\u7c97\u7c92\u5ea6\u7684\u51e0\u4f55\u7ed3\u6784\uff0c\u7136\u540e\u901a\u8fc7\u6d88\u606f\u4f20\u9012\u8fc7\u7a0b\u5c06\u5176\u7ec6\u5316\u4e3a\u7ec6\u7c92\u5ea6\u7684\u7247\u6bb5\uff0c\u4fdd\u5b58\u5c40\u90e8\u6bb5\u7684\u6709\u6548\u6027\u800c\u65e0\u9700\u4f9d\u8d56\u81ea\u56de\u5f52\u5efa\u6a21\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u6b64\u5916\uff0c\u51e0\u4f55\u6d41\u5339\u914d(EquiFM)<\/strong> \u901a\u8fc7\u6df7\u5408\u6982\u7387\u8def\u5f84\uff0c\u5229\u7528\u7b49\u53d8\u5efa\u6a21\u6765\u7a33\u5b9a\u6982\u7387\u52a8\u6001\uff0c\u901a\u8fc7\u7b49\u53d8\u6700\u4f18\u4f20\u8f93\u5bf9\u4e0d\u540c\u6a21\u6001\u7684\u4fe1\u606f\u8fdb\u884c\u5bf9\u9f50\u3002\u8fd9\u79cd\u65b9\u6cd5\u5728\u751f\u6210\u9ad8\u8d28\u91cf\u5206\u5b50\u7ed3\u6784\u65b9\u9762\u4e00\u81f4\u6027\u4f18\u4e8e\u73b0\u6709\u6a21\u578b\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u4e3a\u4e86\u5904\u7406\u5206\u5b50\u51e0\u4f55\u751f\u6210\u7684\u591a\u6a21\u6001\u548c\u566a\u58f0\u654f\u611f\u6027\uff0c\u51e0\u4f55\u8d1d\u53f6\u65af\u6d41\u7f51\u7edc (GeoBFN)<\/strong> \u5f15\u5165\u4e86\u5728\u53ef\u5fae\u53c2\u6570\u7a7a\u95f4\u4e2d\u7684\u5efa\u6a21\u3002GeoBFN\u64c5\u957f\u4e8e\u4fdd\u6301\u4e0d\u53d8\u7684\u5bc6\u5ea6\u5efa\u6a21\uff0c\u5e76\u63d0\u4f9b\u53ef\u9002\u5e94\u7684\u91c7\u6837\u7b56\u7565\u4ee5\u5e73\u8861\u6548\u7387\u548c\u8d28\u91cf\uff0c\u5728\u591a\u4e2a\u57fa\u51c6\u4e0a\u5b9e\u73b0\u4e86\u6700\u5148\u8fdb\u7684\u6027\u80fd\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u6700\u540e\uff0c\u6211\u4eec\u4ecb\u7ecd MolCRAFT<\/strong>\uff0c\u8fd9\u662f GeoBFN<\/strong> \u9488\u5bf9\u57fa\u4e8e\u7ed3\u6784\u7684\u836f\u7269\u8bbe\u8ba1 (SBDD) \u7684\u6761\u4ef6\u7248\u672c\u3002MolCRAFT \u5728\u8fde\u7eed\u53c2\u6570\u7a7a\u95f4\u4e2d\u8fd0\u884c\uff0c\u5e76\u91c7\u7528\u51cf\u566a\u91c7\u6837\u7b56\u7565\u3002\u5b83\u89e3\u51b3\u4e86\u81ea\u56de\u5f52\u548c\u6269\u6563\u65b9\u6cd5\u56fa\u6709\u7684\u6a21\u5f0f\u5d29\u6e83\u548c\u6df7\u5408\u7a7a\u95f4\u95ee\u9898\u3002MolCRAFT\u5728\u7ed3\u5408\u4eb2\u548c\u529b\u548c 3D \u7ed3\u6784\u7a33\u5b9a\u6027\u65b9\u9762\u8868\u73b0\u51fa\u8272\uff0c\u521b\u4e0b\u4e86 Vina Scores \u7684\u65b0\u7eaa\u5f55\uff0c\u5e76\u663e\u8457\u4f18\u4e8e\u73b0\u6709\u57fa\u51c6\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u8fd9\u4e9b\u8fdb\u6b65\u4e0d\u4ec5\u63d0\u9ad8\u4e86 3D \u5206\u5b50\u751f\u6210\u7684\u6548\u7387\u548c\u8d28\u91cf\uff0c\u8fd8\u4e3a\u66f4\u51c6\u786e\u548c\u53ef\u9760\u7684\u836f\u7269\u53d1\u73b0\u8fc7\u7a0b\u94fa\u5e73\u4e86\u9053\u8def\u3002<\/span><\/section>\n

<\/span><\/section>\n
Deep Generative Models for Geometric Graphs Generation: Beyond Diffusion Models<\/strong><\/span><\/section>\n

<\/strong><\/span><\/section>\n
The generation of 3D molecules is crucial for drug discovery, requiring both categorical (atom types) and continuous (atom coordinates) feature determination. Traditional deep generative models, including Diffusion Models (DMs), have shown promise but face significant challenges such as unstable probability dynamics and inefficient sampling processes. To push the boundaries of 3D molecular generation and address these limitations, we will introduce a suite of innovative methods beyond diffusion models:<\/span><\/section>\n

<\/span><\/section>\n
Firstly, the Hierarchical Diffusion-based Model (HierDiff)<\/strong> addresses poor quality in generated large molecules by employing a coarse-to-fine strategy. HierDiff generates coarse-grained geometries via an equivariant diffusion process and refines them into fine-grained fragments through a message-passing procedure, preserving local segment validity without relying on autoregressive modeling.<\/span><\/section>\n

<\/span><\/section>\n
Moreover, Equivariant Flow Matching<\/strong> leverages equivariant modeling to stabilize probability dynamics via a hybrid probability path, aligning information across different modalities through equivariant optimal transport. This approach consistently outperforms existing models in generating high-quality molecular structures.<\/span><\/section>\n

<\/span><\/section>\n
To handle the multi-modality and noise sensitivity of molecule geometry generation, Geometric Bayesian Flow Networks (GeoBFN)<\/strong> introduce modeling in a differentiable parameter space. GeoBFN excels in maintaining invariant density modeling and offers adaptable sampling strategies to balance efficiency and quality, achieving state-of-the-art performance across multiple benchmarks.<\/span><\/section>\n

<\/span><\/section>\n
Finally, we present MolCRAFT<\/strong>, a conditional version of GeoBFN tailored for structure-based drug design (SBDD). MolCRAFT operates in continuous parameter space and employs a noise-reduced sampling strategy. It addresses key challenges such as mode collapse and hybrid space issues inherent in autoregressive and diffusion methods. MolCRAFT achieves superior binding affinity and 3D structural stability, setting new records in Vina Scores and significantly outperforming existing baselines.<\/span><\/section>\n

<\/span><\/section>\n
Together, these advancements not only enhance the efficiency and quality of 3D molecule generation but also pave the way for more accurate and reliable drug discovery processes.<\/span><\/section>\n


<\/span><\/h2>\n

<\/span><\/section>\n

\n
\n
\n
\n
\n

<\/section>\n

<\/section>\n
\n
<\/svg><\/section>\n<\/section>\n<\/section>\n
\n

\u5206\u4eab\u5185\u5bb9\u5927\u7eb2<\/strong><\/span><\/strong><\/strong><\/p>\n<\/section>\n

\n

<\/section>\n

<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n


<\/p>\n

1. \u80cc\u666f\u4e0e\u7814\u7a76\u76ee\u6807\uff1a<\/span><\/strong><\/p>\n