Animals modify their behavior based on experiences as learning, although identifying component(s) of behavior modulated by learning has been difficult. In contrast to neural activities, which can be monitored in large numbers of cells simultaneously recently, behavior in general is still analyzed in classic ways and insufficiently studied using simple measures, such as velocity, migratory distance, and/or the probability of selecting a particular goal. Comprehensive classifications of animals' behavior by using machine vision and machine learning methods have been achieved (Gomez-Marin et al., Nat Neurosci, 2014; Brown et al., PNAS, 2013). However, these methods have not been applied to animals' sensory behavior because of technical limitations in measuring sensory signals that animals receive during the behavior. To overcome this problem and effectively identify behavioral components modulated by learning, we used machine learning aiming to detect changes in navigation of worms in a measured odor gradient. We have previously reported that, after experiencing the repulsive odor 2-nonanone for 1 h, worm's odor avoidance behavior is enhanced, and that they move away from the odor source more efficiently (Kimura et al., J Neurosci, 2010). We have also quantified the dynamic changes in the odor concentration during odor avoidance behavior (Yamazoe-Umemoto et al., Neurosci Res, 2015). In the present study, we used decision tree, a machine learning algorithm, to extract features of the animal's sensorimotor response during navigation modified by odor learning. During the migration down the odor gradient, naive worms responded to slight increases in the repulsive odor concentration by stopping forward movements and initiating turns. In contrast, the probability of response was lowered after learning, suggesting that the learned worms ignore "a yellow light". Consistently, by calcium imaging of ASH neurons, whose activation causes turns under a virtual odor gradient (Tanimoto et al., this meeting), we found that the ASH response to a small increase in the odor concentration was reduced after learning. Furthermore, by applying the decision tree analysis, multiple mutant strains were categorized into several groups based on behavioral features. Thus, the integrative machine learning analysis of sensory information and behavioral response is a powerful tool to obtain comprehensive understanding of dynamic activities of neural circuits and its modulation by learning.

Authors: Yamazaki, S., Ikejiri, Y., Hiramatsu, F., Yamazoe-Umemoto, A., Fujita, K., Tanimoto, Y., Hashimoto, K., Maekawa, T., Kimura, K.

Affiliations:
- Graduate School of Information Sciences, Tohoku University, Miyagi, Japan
- Graduate School of Information Science and Technology, Osaka University, Osaka, Japan
- Graduate School of Science, Osaka University, Osaka, Japan