从大量蛋白-配体复合物中准确预测其亲和力，一直是计算机辅助药物设计和生物信息研究中的难点问题。准确的蛋白-配体亲和力打分函数对虚拟筛选和药物发现有着重要意义。尽管在打分函数上投入了大量研究，但是目前打分函数仍然存在准确率较差等缺陷，从而制约了新药的研发。同时，大量蛋白-配体复合物结构及其亲和力的测定，也为深度学习研究更准确的打分函数提供了可能。因此，为了提高预测蛋白-配体亲和力的能力，本项目围绕使用深度学习算法建立更准确的蛋白-配体亲和力打分函数展开研究。主要工作包括整合和开发用于打分函数研究的数据集；研究将蛋白-配体复合物结构转换为数值特征的方法；研究和构建半监督多任务深度学习模型进行打分函数训练，并对结果在打分函数的对接能力、排序能力、打分能力和筛选能力上进行评测。通过这些研究，预期得出在四种能力上表现更佳的蛋白-配体亲和力打分函数，从而提高新药研发成功率，最终为公众健康做出贡献。

Accurately predicting the binding affinities of large sets of protein-ligand complexes is an extremely challenge in bioinformatics and computer aided drug design, with a significant bearing on virtual screening and drug discovery. Despite intense efforts in developing scoring functions, there are still some drawbacks on them, such as poor accuracy and others, which have been the major roadblock toward cost-effective drug discovery. Meanwhile, large amounts of protein-ligand complex structures and their affinities have been measured, making it possible to study more accurate scoring function by means of deep learning. Therefore, this project focus on establishing more accurate scoring function to improve predicting protein-ligand affinity based on deep learning. The main works include integrating and developing the database for the study of scoring function, researching the methods of converting protein-ligand complex structures into numerical features, researching and building the semi-supervised multi-task deep learning model to train scoring function, and evaluating the results on docking power, ranking power, scoring power and screening power of scoring function. The expected results of these researches will improve the accuracy of scoring function and the success rate of drug discovery, ultimately contributing to public health.