Symposium
The dawn of data- and model-driven neuroscience
シンポジウム
データ駆動型/モデル駆動型神経科学研究の幕開け
協賛：新学術領域研究「生物移動情報学」 |
|---|
| 7月26日（金）8:30～9:00 第5会場（朱鷺メッセ 3F 302）
2S05m-1
データ駆動型の科学的発見に向けて：神経科学研究のための機能性タンパク質設計における実例
Ichiro Takeuchi（竹内 一郎）1,2
1名古屋工業大学 情報工学専攻
2理化学研究所 革新知能統合研究センター
Data-driven approach is a new paradigm for scientific discovery coming after experiment, theory, and simulation approaches. In many areas of science and technology, data-driven approach has been shown to be effective. An advantage of data-driven science is that it can systematically generate/select scientific hypothesis, i.e., (semi-)automatic experimental design is possible based on the past experiments and their results. In this talk, we first review recent artificial-intelligence and machine learning approaches for data-driven science in the context of automatic experimental design. Then, we focus more on an approach called Bayesian optimization, which has been shown to be effective for data-driven experimental design. In Bayesian optimization, a statistical model of the research target is developed based on the past experimental data, and the model is used for determining the next experimental setup such as parameters, conditions etc. Finally, we talk about our recent study on data-driven design of functional proteins, such as channel rhodopsin and XFP, using Bayesian optimization. The goal of this work was to develop an artificially modified protein that has desired light absorption or light emission functions. By using data-driven approach, we developed a statistical model of the relationship between amino-acid sequence and light absorption/emission properties without using the information of 3D structures. Then, using the statistical model, we determined how the amino-acid sequence should be modified for making the protein have desired light absorption/emission properties. We demonstrate in this proof-of-concept study that data-driven experimental design is a promising approach for enhancing scientific discovery. | | 7月26日（金）9:00～9:30 第5会場（朱鷺メッセ 3F 302）
2S05m-2
深層学習による動物行動データからの知識発見
Takuya Maekawa（前川 卓也）
大阪大
Animal behavior can be regarded as a final result output of the brain activities. Therefore, computational methods for objective analysis of animal behavior is essential to understand the brain function. Due to the recent advances in sensing technologies such as GPS technologies and computer vision technologies, we can easily observe and record movement data of animals in the wild and/or in laboratories . These animal movement data are translated to a time-series of coordinates, i.e., trajectory, and then are analyzed by the state of the art computational methods such as machine learning techniques. However, conventional machine learning techniques require feature extraction process that are done by prior knowledge of experienced researchers (such as neuroscientists and ecologists ). The manual analysis and manual feature engineering are difficult because the recent sensing technologies have generated big data of animal movements. This talk introduces animal trajectory analysis methods based on deep learning that learns feature extraction processes in a deep neural network. Specifically, we introduce deep learning based methods for comparing behavior of two different groups of animals and for discovering something in common between two different groups. These methods can be applied to comparing 2D trajectory data from any kinds of animals because inputs of the methods are just trajectory data (a series of two dimensional coordinates). Because meaningful features are automatically extracted from the trajectory data in the deep neural networks, the researchers need not to manually extract features based on their prior knowledge. Such unbiased and comprehensive analysis of behavior should provide us clues to understand new aspects of brain function. | | 7月26日（金）9:30～10:00 第5会場（朱鷺メッセ 3F 302）
2S05m-3
Mapping the neural substrates of behavior using machine learning
Alice A. Robie（ロビー エー アリス），Branson Kristin（Kristin Branson）
HHMI Janelia Research Campus
Describing the structure-function relationship of the brain is a necessary first step toward a circuit-level understanding of how behaviors are generated. To do this in a high-throughput and unbiased manner, we quantified the behavioral effects of activating 2,204 genetically targeted populations of neurons in Drosophila melanogaster, using machine-vision and -learning techniques. This resulted in the automated quantification of the locomotor and social behavior of 400,000 flies. We have shared this database of fly behavior in browsable webpages. To discover the neural substrates of these behaviors, we created a novel quantification of fly brain anatomy based on expression pattern correlations. We combined this anatomical data with our behavioral analysis to create 3D brain-behavioral correlation maps. To make these maps accessible to the scientific community, we developed interactive software to browse the brain-wide atlas of brain-behavior maps. Using these maps, we generated hypotheses of brain regions causally related to sensory processing, locomotor control, courtship, aggression, and sleep. These maps directly specify genetic tools to target the neural populations underlying each map, and we used this to identify a small population of neurons with a role in the neuronal control of walking. | | 7月26日（金）10:00～10:30 第5会場（朱鷺メッセ 3F 302）
2S05m-4
（微小な）脳における情報処理のデータ駆動型予測モデル
Chentao Wen（Wen Chentao）1，Kotaro Kimura（木村 幸太郎）1,2
1名古屋市大院自然科学
2理研AIP
The brain perceives and extracts information from external and internal environments, makes decisions based on the information, and regulates behavior. How can the network of multiple neurons in the brain substantiate such sophisticated functions? Monitoring whole brain activity is regarded as an essential first step in understanding the mechanism of dynamic activity in neural networks for sophisticated brain functions. However, the whole brain activity obtained thus far, such as in functional magnetic resonance imaging studies in humans and primates or from fast 3D calcium imaging in zebrafish and the nematode Caenorhabditis elegans, has been studied mostly with dimensionality reduction using principal component analysis or by calculating similarities between neurons using cross correlation analysis, which are not techniques suitable for revealing "information flow" in neural networks in the brain. Here, we propose to model whole brain activity using machine learning to understand information flow in the network. This approach is advantageous in three aspects compared to the methods mentioned above. (1) Using linear or non-linear models, we are able to predict effects of multiple neurons on each neuron. (2) Prediction of neural activities can be quantitatively tested. (3) Because dimensionality reduction is not used, the loss of potentially critical information is avoided. We made and compared such models by using whole brain activity data from ~150 neurons in C. elegans (Wen, Miura et al., bioRxiv 2018). We found tree-based random forest models are more accurate than linear and long short-term memory (LSTM) models. Based on the random forest model, we estimated the potential causal links between neurons. As a next step, we will extract general activity patterns in the worm brain from comparisons of multiple datasets in different conditions and will compare the results with connectome information (White et al., Phil. Trans. R. Soc. Lond. B 1986). Thus, our results suggest that machine learning-based modeling of neural activities can reveal "information flow" in simple brains and may be used in the brains of other animals and humans. | |
|
|
