I. Future predictions based on Anticipated Learning Machine and short-term data
Chen Luonan's research group from HIAS, UCAS, together with the research groups from Sun Yat-sen University, Soochow University, and the University of Tokyo, published online their latest research result entitled "Predicting future dynamics from short-term time series by anticipated learning machine" in the international academic journalNational Science Review in 2020. The study proposes the Anticipated Learning Machine (ALM) for small-sample time series predictions based on the delay embedding theory, thus providing a new solution to this sort of problem. At the same time, ALM can be regarded as a large-sample data method where predicted variables are constructed by high-dimensional small-sample data.
The research group established a small sample-based ALM neural network using a new spatial-temporal information transformation (STI) method, and applied the neural network to learn non-linear STI mapping in a comprehensive and robust manner. Different from the linear proximal mapping of traditional learning, the excellent non-linear function learning ability of the neural network enables it to better simulate STI mapping. In addition, the Dropout scheme of the ALM neural network enables effective simulation of the process of random sampling. The constructed ALM neural network can integrate the dynamics information in multiple sub-sampling systems to achieve multistep-ahead predictions. Experiments show that the ALM method enables accurate multistep-ahead predictions on multiple real-world datasets, including the Lorentz system, gene expression, wind speed, stock index, traffic flow, and the route of the typhoon center.
The study proposes the new ALM method for small-sample time series data analysis, which can not only be used for multistep-ahead predictions of time series, but also be applied to small-sample data construction and new learning establishment in AI and brain science.
II. Building a new algorithm for identifying causal networks
Recently, Chen Luonan's research group from HIAS, UCAS, together with the research groups from Fudan University, Soochow University, and the University of Tokyo, proposed a new algorithm for data-driven causal network identification. This method can be used to reproduce the intrinsic causal networks of large-scale complex dynamical systems, which helps to analyze the essential mechanisms and laws of real-world system evolution. The research result was published online in the comprehensive academic journalNature Communications on May 26, with the title "Partial cross mapping eliminates indirect causal influences".
Causal relationships are the most pervasive and essential links between natural phenomena. It is of great scientific significance to discover the intrinsic causal relationships and causal networks that reflect the core interaction mechanism of system evolution in natural sciences such as physics, life sciences, and geography, as well as in social sciences such as philosophy and economics. Therefore, how to accurately identify the causal relationships and causal networks between system variables based on large-scale data and on the premise of lacking accurate models of complex systems has become a focus in scientific research, including AI research, and has attracted wide attention from scholars.
This study further improves the theoretical system of causal analysis of existing complex systems and provides effective mathematical methods for common scientific problems of multiple disciplines, with a wide application prospect in data-driven research, reflecting the value of applied mathematical research. Research Fellow Chen Luonan, Prof. Lin Wei of Fudan University, and Prof. Aihara Kazuyuki of the University of Tokyo are the co-corresponding authors, Dr. Leng Siyang is the first author of the paper, and Prof. Ma Huanfei of Soochow University is the co-author of the paper. The study was supported by the National Natural Science Foundation of China, the Major Research Program of the Ministry of Science and Technology of the People's Republic of China, and the major research program of the People's Republic of China, and Science and Technology Commission of Shanghai Municipality.
III. Disease characterization using a partial correlation-based sample-specific network
Chen Luonan's research group from HIAS, UCAS proposed a new method of disease characterization using a partial correlation-based sample-specific network.
A single-sample network (SSN) is a biological molecular network constructed from single-sample data given a reference dataset and can provide insights into the mechanisms of individual diseases and aid in the development of personalized medicine. In this study, we proposed a computational method, a partial correlation-based single-sample network (P-SSN), which not only infers a network from each single-sample data given a reference dataset but also retains the direct interactions by excluding indirect interactions. By applying P-SSN to analyze tumor data from the Cancer Genome Atlas and single cell data, we validated the effectiveness of P-SSN in predicting driver mutation genes (DMGs), producing network distance, identifying subtypes and further classifying single cells. In particular, P-SSN is highly effective in predicting DMGs based on single-sample data. P-SSN is also efficient for subtyping complex diseases and for clustering single cells by introducing network distance between any two samples.
Thus, our method manifests that these nondifferential genes may have rich information concerning diseases and are the "dark genes" in terms of expression. It should be declared that the P-SSN is a theoretical network constructed using a single sample that quantifies the changing degree of correlations of a single sample against the reference samples. The P-SSN reflects the differences between normal and cancer samples from a regulation, interaction or network perspective. Traditional molecule networks are aggregated networks based on multiple samples, while a P-SSN is a specific network based on a single sample. Therefore, compared to traditional molecular biomarkers, it can be applied to construct NBs or dynamic network biomarkers for individual diagnosis and prediction, and be a potential tool in personalized medicine.
