正文：
近日，浙江大学侯廷军团队在化学逆合成预测领域取得重要突破，提出了一种基于Transformer的逆合成模型——EditRetro，该模型在Nature子刊《Nature Communications》上发表。研究显示，该模型在标准基准数据集USPTO-50K上取得了60.8%的top-1准确率，显示出其在AI驱动的化学合成规划领域的巨大潜力。
逆合成预测是药物发现和有机合成中的一个关键任务，它要求将复杂的分子逆向分解为更简单的组成部分，以便于合成。传统的AI方法在预测多样性方面存在局限性，且难以适应复杂分子的合成。为此，侯廷军团队将单步逆合成预测重新定义为分子串编辑任务，通过迭代细化目标分子串，实现了高质量和多样化的预测。
EditRetro模型通过三种编辑操作实现分子串的生成：序列重新定位、占位符插入和标记插入。该模型由一个编码器和三个解码器组成，它们协同工作，提高了生成效率。与现有的基于序列和图编辑的方法相比，EditRetro模型在预测准确度方面取得了显著优势。
这项研究不仅推动了AI在化学合成领域的应用，也为生物医学、制药和材料工业等领域提供了有力的技术支持。未来，随着AI技术的不断发展，我们可以期待更多的创新成果，加速化学合成过程，推动相关产业的进步。
英语如下：
News Title: “Zhejiang University Model in Nature Journal: 60.8% Accuracy Breakthrough in Reverse Synthesis Prediction”
Keywords: Zhejiang University model, reverse synthesis prediction, Nature journal
News Content:
Title: Zhejiang University Team Proposes New Model to Enhance Reverse Synthesis Prediction Accuracy
Recent breakthroughs in the field of chemical reverse synthesis prediction by the Hou Tingjun team at Zhejiang University have been published in the Nature journal, Nature Communications. The team has developed a Transformer-based reverse synthesis model named EditRetro, which achieved a top-1 accuracy rate of 60.8% on the standard benchmark dataset USPTO-50K, demonstrating its significant potential in AI-driven chemical synthesis planning.
Reverse synthesis prediction is a critical task in drug discovery and organic synthesis, requiring the complex molecules to be decomposed into simpler components for synthesis. Traditional AI methods have limitations in predicting diversity and are difficult to adapt to the synthesis of complex molecules. To address these challenges, the Hou Tingjun team redefined single-step reverse synthesis prediction as a molecular string editing task, achieving high-quality and diverse predictions through iterative refinement of the target molecular string.
The EditRetro model generates molecular strings through three editing operations: sequence repositioning, placeholder insertion, and tag insertion. The model consists of an encoder and three decoders that work together to improve generation efficiency. Compared to existing methods based on sequence and graph editing, the EditRetro model has achieved significant advantages in prediction accuracy.
This research not only advances the application of AI in the field of chemical synthesis but also provides strong technical support for biomedical, pharmaceutical, and material industries. As AI technology continues to develop, we can look forward to more innovative achievements that will accelerate the chemical synthesis process and promote the progress of related industries.
【来源】https://www.jiqizhixin.com/articles/2024-08-06-3
Views: 2