Kanat Chanthongdee ^1,2 Yerko Fuentealba ¹ Thor Wahlestedt ¹ Lou Foulhac ¹ Tetiana Kardash ¹ Andrea Coppola ¹ Markus Heilig ¹ Estelle Barbier ^1*

• ¹ Center for Social and Affective Neuroscience, Linköping University, Linkoping, Sweden
• ² Department of physiology, Faculty of Medicine, Siriraj Hospital, Mahidol University, Bangkok, Bangkok, Thailand

Defensive responses to threat-associated cues are commonly evaluated using conditioned freezing or suppression of operant responding. However, rats display a broad range of behaviors and shift their defensive behaviors based on immediacy of threats and context. This study aimed to identify the defensive behaviors that are triggered in response to threat-associated cues and assess whether they can accurately be identified using DeepLabCut in conjunction with SimBA. We evaluated behavioral responses to fear using the auditory fear conditioning paradigm. Observable behaviors triggered by threat-associated cues were manually scored using Ethovision XT. Subsequently, we investigated the effects of diazepam (0, 0.3, or 1mg/kg), administered intraperitoneally before fear memory testing, to assess its anxiolytic impact on these behaviors. We then developed a DeepLabCut + SimBA workflow for ethological analysis employing a series of machine learning models. The accuracy of behavior classifications generated by this pipeline was evaluated by comparing its output scores to the manually annotated scores. Our findings show that, besides conditioned suppression and freezing, rats exhibit heightened risk assessment behaviors, including sniffing, rearing, free-air whisking, and head scanning. We observed that diazepam dose-dependently mitigates these risk-assessment behaviors in both sexes, suggesting a good predictive validity of our readouts. With adequate amount of training data (approximately >30,000 frames containing such behavior), DeepLabCut + SimBA workflow yields high accuracy with a reasonable transferability to classify well-represented behaviors in a different experimental condition. We also found that maintaining the same condition between training and evaluation data sets is recommended while developing DeepLabCut + SimBA workflow to achieve the highest accuracy. Our findings suggest that an ethological analysis can be used to assess fear learning. With the application of DeepLabCut and SimBA, this approach provides an alternative method to decode ongoing defensive behaviors in both male and female rats for further investigation of fear-related neurobiological underpinnings.