康 · 学术 | Pharmaron Launches the 69th Pharmaron Virtual Lecture

康龙化成举办第六十九期“合成与药物化学前沿”名师线上讲座

2026年10月29日，北京——美国耶鲁大学的Timothy Newhouse教授做客康龙化成第六十九期“合成与药物化学前沿”名师线上讲座，报告主题为“计算方法驱动的合成化学”。报告重点探讨了四个研究方向：（1）DFT计算辅助的逆合成分析与反应预测；（2）机器学习驱动的小数据集配体优化；（3）生成式AI在配体设计与逆合成分析中的应用；（4）反应可行性评估与文献智能检索工具。传统的合成设计主要依赖化学家的经验直觉，而Newhouse教授则通过将理论计算深度嵌入合成规划过程，搭建从可行性预判、配体筛选到路线设计的多角度、全方位支持。

第一部分，Newhouse教授介绍了DFT计算辅助的逆合成分析与反应预测。通过对过渡态能量的考量，对合成路线上的关键步骤进行热力学与动力学评估。他的团队成功预测了非常规环化反应的可行性与选择性，并将canangaterpine II等天然产物的合成步骤大幅简化。Newhouse教授还展示了利用DFT计算分析钯催化反应过渡态、指导配体进化的案例，证明了理论计算在关键步骤优化中的强大功能。

第二部分，Newhouse教授介绍了机器学习驱动的小数据集配体优化。针对合成方法学论文中数据集规模普遍有限的挑战，其团队开发了一种基于神经网络的建模方法，通过提取几何特征（如空间位阻体积）和电子特征（如C-H-π色散作用）构建参数化模型。通过增量验证，该模型成功指导了多个反应中的配体选择，证明了有限数据条件下机器学习依然能够生成可靠的合成指导。

第三部分，Newhouse教授聚焦于生成式AI在配体设计与逆向合成中的应用。他的团队利用自编码器对两千个膦配合物的几何与电子参数进行编码，于潜在空间中采样解码，生成结构新颖的膦配体。尽管部分AI生成结构看似不合理，团队仍成功合成其中一种烷基膦并高效应用于催化反应。此外，Newhouse教授展示了将生成模型用于逆合成分析的探索，令该系统可直接从目标分子输出全新合成模板，开辟更高效的路线。

第四部分，Newhouse教授介绍了新颖的反应可行性评估与文献智能检索工具。他的团队开发了“MOSAIC评分”系统，基于文献大数据训练模型，为化学反应的可行性和预期产率提供量化评分。实验验证表明，MOSAIC评分与实验结果高度相关，且优于常用的商业大语言模型，其结论可追溯至具体文献实例，具备可解释性。同时，他的团队还评估了多种描述底物复杂度的指标，提出基于结构相似性的自动化评估方案，为计算辅助逆向合成提供了可靠判据。在文献检索方面，Newhouse教授探讨了大语言模型在反应条件查询中的应用与局限，并展示了MOSAIC在化学文献智能读取中的实用性。

讲座结束后，Newhouse教授与听众在问答环节中，围绕数据集偏差（阴性结果缺失）、开源代码可及性、下一代化学家的AI技能培养等方向展开了热烈的讨论。

Frontiers in Synthetic and Medicinal Chemistry

--The 69^th Pharmaron Virtual Lecture

Beijing, China, Oct. 29, 2026 - Professor Timothy Newhouse from Yale University delivered a lecture at the 69^th session of Pharmaron’s Frontiers in Synthetic and Medicinal Chemistry virtual lecture series. His presentation was entitled "Computationally Augmented Total Synthesis". The lecture highlighted four research topics: (1) DFT-Enabled Retrosynthetic Analysis and Reaction Prediction; (2) Machine Learning-Driven Ligand Optimization on Small Datasets; (3) Generative AI in Ligand Design and Retrosynthetic Analysis; and (4) Reaction Feasibility Assessment and Intelligent Literature Mining Tools. While traditional synthetic design primarily relies on chemists' empirical intuition, Professor Newhouse has deeply integrated theoretical computations into the synthesis planning process, building multi-perspective, comprehensive support spanning feasibility prediction, ligand screening, and route design.

In the first part, Professor Newhouse introduced DFT-enabled retro-synthetic analysis and reaction outcome prediction. By evaluating transition state energies, his team performed thermodynamic and kinetic assessments of key steps along synthetic routes. The team successfully predicted the feasibility and selectivity of unconventional cyclization reactions, anddrastically simplified the synthesis of natural products such as canangaterpine II. Professor Newhouse also demonstrated how DFT calculations were used to analyze transition states in palladium-catalyzed reactions and guide ligand evolution, proving the powerful capabilities of theoretical computations in optimizing key synthetic steps. 

In the second part, Professor Newhouse presented machine learning-driven ligand optimization on small datasets. Addressing the common challenge of limited dataset sizes in synthetic methodology papers, his team developed a neural network-based modeling approach that extracts geometric features (such as steric hinderance) and electronic features (such as C–H–πdispersion interactions) to construct a parameterized model. Through incremental validation, the model successfully guided ligand selection in multiple transformations, demonstrating that machine learning can still generate reliable synthetic guidance under data-limited conditions.

In the third part, Professor Newhouse focused on generative AI in ligand design and retrosynthetic analysis. His team employed an autoencoder to encode geometric and electronic parameters of more than 2,000 phosphine complexes, sampling the latent space to decode structurally novel phosphine ligands. Although some AI-generated structures appeared unreasonable, the team successfully synthesized one such alkyl phosphine and efficiently applied it in catalysis. Furthermore, Professor Newhouse showcased explorations in applying generative models to retro-synthetic analysis, enabling the system to directly output entirely new synthetic templates from a target molecule, opening up more efficient routes.

In the fourth part, Professor Newhouse introduced novel reaction feasibility assessment and intelligent literature mining tools. His team developed the "MOSAIC" system, a model trained on large-scale literature data that provides quantitative feasibility and yield predictions for chemical reactions. Experimental validation confirmed a strong correlation between MOSAIC scores and experimental outcomes, outperforming commonly used commercial large language models. Its conclusions are traceable to specific literature examples, ensuring interpretability. Concurrently, his team evaluated multiple metrics for describing substrate complexity and proposed a structural similarity-based automated assessment scheme, providing a reliable criterion for computation-assisted retrosynthesis. Regarding literature retrieval, Professor Newhouse discussed both the applications and limitations of large language models in querying reaction conditions, anddemonstrated the practical utility of MOSAIC in intelligent chemical literature reading.

Following the lecture, Professor Newhouse engaged in a lively Q&A session with the audience, discussing topics such as dataset bias (the absence of negative results), open-source code accessibility, and AI skill development for the next generation of chemists.