Improving QSAR Models for the Prediction of the Activities and Toxicity of Small Molecule Candidates

本项目旨在显著改善小分子药物活性和毒性的定量构效关系（QSAR）模型预测。本研究的背景是迫切需要提高药物开发中QSAR模型的可靠性，因为准确的预测可以加快潜在候选药物的识别，从而减少与传统实验方法相关的成本和时间。该项目的主要目标涉及实施创新方法，例如使用自我训练、标签平滑和类原型等技术来管理噪声数据集。该过程旨在提高训练数据的质量，从而增强QSAR模型的鲁棒性。此外，该项目利用深度学习研究的最新进展，探索不同的分子表示作为QSAR模型的输入。最近在二维和三维图形神经网络方面取得了许多进展，但尚未应用于分子的图形表示。对不同分子表征的探索有可能揭示分子结构和活动之间更复杂的关系，从而得出更准确的预测。此外，该研究将活动悬崖（AC）预测纳入QSAR建模，这方面的研究还很有限。活性悬崖是指一对结构相似的分子，对给定目标的结合亲和力表现出显著差异。通过将AC预测集成到QSAR模型中，该项目旨在捕捉这些细微的区别，进一步提高模型的预测能力。根据EPSRC的战略，该项目属于EPSRC计算和理论化学研究领域。EPSRC的目标是促进计算化学的前沿研究，以推动药物发现和其他科学领域的进步。值得注意的是，这项研究是与拉萨有限公司合作进行的，该公司利用了一个联邦学习数据集。联邦学习允许多个组织在不共享敏感数据的情况下进行协作，从而实现跨分布式私有数据集的机器学习算法训练。这种协作扩大了项目的范围和影响，并为进一步研究本地模型预测和在更广泛的联邦模型中使用本地模型提供了潜在的途径，以生成统一的预测活动标签，而不会暴露任何敏感数据。总之，该研究项目旨在通过多种创新途径增强QSAR建模，包括数据管理、不同分子表示、活动悬崖预测和联邦学习。通过与EPSRC的研究领域保持一致，并与行业领导者拉萨有限公司合作，该项目有可能彻底改变药物发现和计算化学领域。

The project aims to significantly improve Quantitative Structure-Activity Relationship (QSAR) model prediction for small molecule drug activities and toxicity. The context of this research lies in the critical need to improve the reliability of QSAR models in drug development, as accurate predictions can expedite the identification of potential drug candidates, thus reducing costs and time associated with traditional experimental methods. The project's primary objectives involve the implementation of innovative methodologies such as curating noisy datasets using techniques such as Self-Training, Label Smoothing, and Class Prototyping. This process aims to enhance the quality of training data, thereby enhancing the robustness of QSAR models. Furthermore, the project leverages recent advancements in deep learning research to explore diverse molecular representations as inputs for QSAR models.There have been many recent advancements in graph neural networks, both 2 and 3-dimensional, which have yet to be applied to graphical representations of molecules. The exploration of different molecular representations has the potential to uncover more intricate relationships between molecular structures and activities, leading to more accurate predictions. Additionally, the research incorporates activity cliffs (AC) prediction into QSAR modelling, in which only limited research has been done. Activity cliffs refer to pairs of structurally similar molecules that exhibit significant differences in binding affinity to a given target. By integrating AC prediction into QSAR models, the project aims to capture these subtle distinctions, further improving the predictive capabilities of the models.In alignment with EPSRC's strategies, this project falls within the EPSRC Computational and Theoretical Chemistry research area. The endeavour resonates with the EPSRC's goal of promoting cutting-edge research in computational chemistry to drive advancements in drug discovery and other scientific domains. Notably, this research is conducted in collaboration with Lhasa Limited, which utilises a federated learning data set. Federated learning allows multiple organisations to collaborate without sharing sensitive data, enabling the training of machine learning algorithms across distributed, private datasets. This collaboration broadens the project's scope and impact, as well as providing potential paths for further investigation into both local model prediction and the use of local models within a broader federated model to generate consolidated predicted activity labels, without exposing any sensitive data.In conclusion, the research project aims to enhance QSAR modelling through multiple innovative avenues, including data curation, diverse molecular representations, activity cliffs prediction, and federated learning. By aligning with EPSRC's research area and partnering with industry leader Lhasa Limited, this project has the potential to revolutionise the field of drug discovery and computational chemistry.