The Life Science Track champion "Niu Dao Small Test" team is composed of Wang Taifu, bioinformatics major from the University of Chinese Academy of Sciences, and Liu Zhijian, molecular biology major from South China Agricultural University.
For the "Biological age Assessment and Risk prediction of aging disease" challenge, they innovatively used two models to predict biological age and Alzheimer's disease, respectively, while taking the output of age prediction as the input for Alzheimer's disease prediction.
Faced with a large amount of data, they carried out detailed exploratory analysis of the data and targeted screening to get the final characteristics. The optimization of the sample data played a key role in their achievement of the first prediction.
Such a topic is a difficult task for the two master's students. However, Liu Zhijian believes that it is thanks to the good feedback of the competition system and the healthy competition between the teams, so that they can complete the task that one or two people could not complete.
"We started with the simplest model at the beginning, based on the feedback of the a and b lists of the competition, and iterated the parameters and models through one jupyter notebook after another, trying different strategies to achieve good prediction results."
"By learning such a large and widely distributed data set, members of the ai background can learn how to better understand the data from members of the chemistry background, and members of the chemistry background have an opportunity to get up close and personal with the intersection of ai and chemistry."
Wentao Guo, a PhD student at the University of California, Davis, is a member of gpt-4 auto agent. This is an international "industry-university collaboration" team in the quantum chemistry circuit, consisting of two algorithm researchers from Deep Potential Technology,
a PhD student from the Hong Kong University of Science and Technology, and one. According to the characteristics of the data, the members proposed a "molecular attribute prediction" scheme based on directional message passing neural network.
Since the preliminary competition in June, the team has continued to improve the efficiency of the model, from the beginning of the selection of gemnet-dt with general accuracy but simpler model,
which only considers three-body interaction, to the semifinal adjustment to the introduction of four-body interaction under the same family gemnet-oc model.
Inspired by the machine learning model engineering task solving pipeline, they also experimented with strategies such as model fusion, hyper-parametric fine-tuning, and introducing more molecular information such as bond information.
"An increase in accuracy is always accompanied by a decrease in model speed, so we are constantly making adjustments to better balance accuracy and speed." Liu Siyuan explained.
