Release of personalized federated learning algorithm framework, empowering AI drug research and development

Abstract: , Chinese Academy of Sciences, Shanghai University of Science and Technology, and Huawei Cloud Medical Agent team published an article entitled "Facing Small and Biased Data Dilemma in Drug Discovery with Enhanced Federated Learning Approaches" in Science China Life Sciences.

This article is shared from the HUAWEI cloud community " Shanghai Institute of Medicine, Chinese Academy of Sciences/Shanghai University of Science and Technology, and HUAWEI CLOUD joint team released a personalized federated learning algorithm framework to AI drug development 1610df02ee8373", author: HUAWEI cloud headline.

Article source: China Science Magazine

Drug R&D is a long process. Traditional drug R&D requires investment of a large number of R&D personnel, and it takes ten to fifteen years and billions of dollars in R&D funding to get a drug to market. In recent years, with the development of technologies such as AI, big data, and cloud computing, more and more pharmaceutical companies and technology giants have set their sights on this field. However, AI drug R&D faces a series of difficulties and challenges. AI models require a large amount of data for modeling. The high barriers, high costs, and high confidentiality of drug R&D data affect the enthusiasm of pharmaceutical companies to contribute to data. At the same time, the phenomenon of data islands is widespread, and many internal data of enterprises are small and highly biased, which brings great challenges to high-quality AI drug research and development models. In recent years, the emerging federated learning can solve this problem well. Federated learning is essentially a distributed machine learning technology, and its goal is to achieve common modeling on the basis of ensuring data privacy, security and compliance. Under the framework of federated learning, multiple pharmaceutical companies do not need to share data, and only by sharing model weights, they can achieve collaborative training between pharmaceutical companies, which can enhance the effects of AI models while ensuring data security.

Recently, Shanghai Institute of Materia Medica, Chinese Academy of Sciences, Shanghai University of Science and Technology, and Huawei Cloud Medical team published an article titled " Facing Small and Biased Data Dilemma in Drug Discovery with Enhanced Federated Learning Approaches " in 1610df02ee83fc Science China Life Sciences The joint team used three tasks to simulate the joint learning process across data islands: predicting drug solubility, kinase inhibitory activity, and hERG cardiotoxicity based on chemical structure. These data cover different medicinal chemistry spaces, experimental measurement methods, experimental conditions, and data sizes, and represent the differences in the data distribution of different pharmaceutical companies in the real world. In this way, I will study the significance of federated learning to break the data island, and found from the analysis results that the effect of federated learning is better than model training with a single data source.

Then, in order to further improve the model effect, the joint team introduced the residual fully connected network (RFCN), by using the AI automatic modeling tool AutoGenome1 to retrain the three tasks to obtain a more accurate model skeleton; in addition, the federated model parameters In the integration strategy, the joint team introduced personalized federated learning (FedAMP)2, which trains personalized models for federated computing participants, and strengthens the collaboration between participants with similar data distribution through the attention message transmission mechanism, so that the data contribution is greater. More and better quality participants will benefit more; in the comparison of the performance of kinase inhibitory activity prediction, we can see that the introduction of RFCN and FedAMP has three AI tasks in terms of drug solubility, kinase inhibitory activity and hERG cardiotoxicity prediction. Above, they are superior to traditional MLP and FedAvg methods.

Recently, Shanghai Institute of Pharmaceutical Sciences/Shanghai University of Science and Technology and Huawei Cloud Medical Intelligence jointly released the drug federation learning service based on HUAWEI CLOUD ModelArts platform to help pharmaceutical companies and research institutions use drug federation learning more conveniently through simple four-step operation , Users who participate in federation learning can easily implement federation training: the first step: the leader creates a federation and defines the federation tasks, such as drug structure prediction water solubility; the second step: the leader invites participants to join the federation, and the participants agree to join; The third step: the federal members deploy agents and configure the federal operating environment; the fourth step: the leader initiates the federal task and starts the federal job training.

Based on the powerful AI capabilities of Huawei Cloud AI Ascension Cluster Service and ModelArts, Huawei Cloud’s one-stop AI development platform, Huawei Cloud Medical Intelligence EIHealth integrates many algorithms, tools, AI models, and automated pipelines in the medical field. The goal is to build a full stack , An open and professional enterprise-level AI research and development platform for the medical industry. more information, please visit: Click to follow and learn about Huawei Cloud's fresh technology for the first time~