在生命科学的前沿领域，一项由诺华生物医学研究（Novartis Biomedical Research）团队进行的最新研究揭示了机器学习（ML）模型在预测靶向蛋白质降解剂（TPD）特性的潜力。这项研究是Nature子刊的最新成果，旨在评估ML模型在理解并预测TPD的关键性质上的能力，包括被动渗透性、代谢清除率、细胞色素P450抑制、血浆蛋白结合和亲脂性等。
在过去的几十年里，定量结构-性质关系（QSPR）模型一直被广泛用于预测新分子的物理化学性质。然而，随着TPD等新型药物设计模式的兴起，对QSPR模型的有效性提出了挑战。诺华生物医学研究的团队选择了ML模型作为探索的工具，以期在TPD项目中实现更准确的性质预测。
研究发现，ML模型在预测TPD的性质时，性能与其他预测模型相当。在具体的性质预测中，对“glues”和“异双功能蛋白”的预测出现了不同的结果，其中“glues”的风险类别预测相对较低，而“异双功能蛋白”的预测则相对较高。对于被动渗透性、细胞色素P450抑制以及人类和大鼠微粒体清除率的预测，模型的错误分类错误率在4%以下，而对于“glues”的高风险和低风险类别的预测错误率则更低，范围在0.8%至8.1%之间。
这项研究的独特之处在于，它对ML预测ADME（吸收、分布、代谢和排泄）性质的全面评估，不仅涵盖了TPD分子，还特别关注了异双功能分子和分子胶子模态。这是首次对ML在这一领域的应用进行全面、深入的评估，为未来基于ML的药物发现和设计提供了重要的参考。
该研究结果不仅对药物开发领域具有重要意义，也展示了ML技术在生物医学研究中的潜在应用价值。通过改进预测模型的准确性，研究人员可以更有效地设计和优化新型药物，加速新药的开发过程，最终为患者带来更有效的治疗方案。
英语如下：
News Title: “Novartis Study: ML Models Show Exceptional Performance in Predicting Characteristics of Targeted Protein Degradants”
Keywords: Novartis Study, ML Prediction, Targeted Protein Degradants
News Content: In the cutting-edge field of life sciences, a recent study conducted by the team at Novartis Biomedical Research has unveiled the potential of machine learning (ML) models in predicting the characteristics of targeted protein degradation agents (TPDs). This study, published in a prestigious subsidiary of Nature, aims to evaluate the capability of ML models in understanding and predicting the key properties of TPDs, including passive permeability, metabolic clearance, inhibition of cytochrome P450, plasma protein binding, and lipophilicity.
Over the past few decades, quantitative structure-property relationship (QSPR) models have been widely used for predicting the physical and chemical properties of new molecules. However, with the emergence of novel drug design paradigms such as TPDs, the validity of QSPR models has been challenged. The team at Novartis Biomedical Research opted for ML models as a tool to explore and improve the accuracy of property predictions in TPD projects.
The study found that ML models performed comparably to other prediction models in predicting the properties of TPDs. In specific property predictions, there were divergent outcomes for “glues” and “heterobifunctional proteins,” with a lower risk category prediction for “glues” and a higher prediction for “heterobifunctional proteins.” For passive permeability, cytochrome P450 inhibition, and human and rat microsomal clearance rates, the model’s misclassification error rates were below 4%, while the range for “glues” between high and low risk categories was from 0.8% to 8.1%.
What sets this study apart is its comprehensive assessment of ML predictions for ADME (absorption, distribution, metabolism, and excretion) properties, not only for TPD molecules but also with a particular emphasis on heterobifunctional and molecular glue modes. This is the first comprehensive and in-depth evaluation of ML’s application in this domain, providing valuable insights for future drug discovery and design based on ML.
The study’s findings hold significant importance for the pharmaceutical development field, showcasing the potential application value of ML technologies in biomedical research. By enhancing the accuracy of prediction models, researchers can more efficiently design and optimize new drugs, accelerating the drug development process and ultimately delivering more effective treatment options to patients.
【来源】https://www.jiqizhixin.com/articles/2024-07-20-11
Views: 1