A head-fixed rodent virtual reality (VR) system allows imaging of neuronal activity from the brain during behavioral tasks. To accurately measure the complex two-dimensional behavior of animals, it is necessary to detect the motion of the spherical treadmill of the VR system with multiple sensors. However, commercial VR systems for rodents are not widely used; they are expensive and the algorithms used to calculate the motion of the spherical treadmill from sensor signals are not publicly available. We developed a system to detect the motion of a spherical treadmill at high-speed using two optical mice. The system can be used on a general-purpose VR platform. A novel algorithm to robustly calculate the rotation axis of the spherical treadmill using spherical geometry was also devised. This system enables accurate reconstruction of complex two-dimensional behaviors of animals. The system is open source, which should encourage the use of VR-based approaches for imaging brain activity in rodents during behavioral tasks.

Identifying factors that contribute to microbial growth is important for realizing efficient production of useful substrates. Our objective was to predict unknown metabolic pathways from experimental time-series data in model organisms such as Escherichia coli. We focused on a previous method that replaces the computation of auto-regression in the Granger causality test with non-parametric multiplicative regression (NPMR) to allow inferences on noisy and nonlinear data. We then proposed a new causal inference method that creates a multi-dimensional space based on the error between the time series predicted by NPMR and the original time series. We confirmed that the inference accuracy of the proposed method outperforms that of NPMR by 50% using short time series generated by coupled logistic equations, which allows for adjustment of the strength of the causal relationship. The proposed method was applied to simulation data obtained from a kinetic model for glycolysis in E. coli and achieved 61% accuracy.

Eukaryotic genomes contain exons and introns, and it is necessary to accurately identify exon-intron boundaries, i.e., splice sites, to annotate genomes. To address this problem, many previous works have proposed annotation methods/tools based on RNA-seq evidence. Many recent works exploit neural networks (NNs) as their prediction models, but only a few can be used to generate new genome annotation in practice. In this study, we propose AtLASS, a fully automated method for predicting splice sites from genomic and RNA-seq data using attention-based Bi-LSTM (Bidirectional Long Short-Term Memory). We exploit two-stage training on RNA-seq data to address the problem of biased label problem, thereby reducing the false positives. The experiments on the genomes of three species show that the performance of the proposed method itself is comparable to that of existing methods, but we can achieve better performance by combining the outputs of the proposed method and the existing method. The proposed method is the first program specialized in end-to-end splice site prediction using NNs.