(C) PLOS One
This story was originally published by PLOS One and is unaltered.
 . . . . . . . . . .



The human hypothalamus coordinates switching between different survival actions [1]

['Jaejoong Kim', 'Department Of Humanities', 'Social Sciences', 'Computation', 'California Institute Of Technology', 'Pasadena', 'California', 'United States Of America', 'Sarah M. Tashjian', 'Dean Mobbs']

Date: 2024-07 

Comparative research suggests that the hypothalamus is critical in switching between survival behaviors, yet it is unclear if this is the case in humans. Here, we investigate the role of the human hypothalamus in survival switching by introducing a paradigm where volunteers switch between hunting and escape in response to encounters with a virtual predator or prey. Given the small size and low tissue contrast of the hypothalamus, we used deep learning-based segmentation to identify the individual-specific hypothalamus and its subnuclei as well as an imaging sequence optimized for hypothalamic signal acquisition. Across 2 experiments, we employed computational models with identical structures to explain internal movement generation processes associated with hunting and escaping. Despite the shared structure, the models exhibited significantly different parameter values where escaping or hunting were accurately decodable just by computing the parameters of internal movement generation processes. In experiment 2, multi-voxel pattern analyses (MVPA) showed that the hypothalamus, hippocampus, and periaqueductal gray encode switching of survival behaviors while not encoding simple motor switching outside of the survival context. Furthermore, multi-voxel connectivity analyses revealed a network including the hypothalamus as encoding survival switching and how the hypothalamus is connected to other regions in this network. Finally, model-based fMRI analyses showed that a strong hypothalamic multi-voxel pattern of switching is predictive of optimal behavioral coordination after switching, especially when this signal was synchronized with the multi-voxel pattern of switching in the amygdala. Our study is the first to identify the role of the human hypothalamus in switching between survival behaviors and action organization after switching.

Similar to the animal studies, we found that the human hypothalamus encodes the switching between hunting and escaping behavior in the MVPA analysis. Further, we show that the hypothalamus-PAG circuit and the hypothalamus-amygdala circuit are involved in the encoding of survival switching. However, human survival switching also involves network-scale interaction between regions known for “cognitive” task switching and survival behavior, including prefrontal cortical regions and the hippocampus. Model-based fMRI analyses showed that a strong hypothalamic multi-voxel pattern of switching, especially when it is synchronized with the multi-voxel pattern of switching of the amygdala is predictive of an optimal behavioral coordination, suggesting the involvement of hypothalamus–amygdala circuit in the motor coordination after the switching. Finally, in a control task, we did not observe ta role of the hypothalamus in switching between continuous behaviors, nor did it not interact with other regions to encode the switching, suggesting that the hypothalamus is likely to be involved in a switching between survival behaviors rather than being involved in a switching between general continuous behaviors.

We aimed to test several questions: (i) If the hypothalamus is involved in survival-related behavioral switches, we should be able to decode transitions between hunting and escaping using MVPA approaches; (ii) further we hypothesized that the hypothalamic connections with other regions, including the PAG, are involved in the switching between hunting and escaping; (iii) we next predicted that the hypothalamus is involved in the coordination of the behavior after the switching (e.g., coordinating escape behavior after the transition from the hunt). This would predict a stronger neural pattern indicating the switching would predict better movement coordination after the switching; (iv) finally, if the null hypothesis is correct, the hypothalamus is generally involved in switching between continuous movement behaviors, we expect to see the same in the control task.

(A) Experimental paradigm; participants controlled an avatar (player; yellow circle) and either hunted (virtual prey) or escaped (virtual predator) from the cyan circle. At the start of every trial, participants are shown a 2D virtual arena with either a pink (in this example, signaling “Hunt” condition) or orange (in this example, signaling “Escape” condition) boundary for 1 s. In the “Hunt condition,” the participants should chase and attempt to capture a virtual prey. Conversely, in the “Escape condition,” the participants should escape from the virtual predator. Before starting the experimental task, participants occasionally entered a pre-hunt or pre-escape condition where participants prepared for the predator or prey before they entered the arena, allowing us to examine preparatory defensive or predatory movements. Players used an MRI-compatible joystick to complete all tasks. A successful hunt was defined as the capture of the prey before the termination of each trial. Successful escape was defined as not being captured. (B) Visualization of the visiting frequency in the pre-hunt and pre-escape period (yellow: high visiting frequency; blue low visiting frequency; normalized with a maximum visiting frequency). Players tend to move closer to the boundary (thigmotaxis) before escaping compared to hunting in both online and fMRI studies since both predator and prey tend to appear close to the center of the field. (C) Bar graphs show significant differences in thigmotaxis (a proportion of the time spent closer to the boundary; ranged between 0 and 1; ideally, the thigmotaxis would be close to 1 before the escape while it would be close to 0 before the hunt considering the expected location of the predator/prey appearance) before Escape compared to Hunt Conditions. (D, E) Movement difference between Hunt and Escape conditions. Linearity of the movement was defined as the movement that minimizes distance from the current position of the prey in the Hunt task and the movement that maximizes distance from the current position of the predator in the Escape task, which ranged between −1 and 1 where 1 means the movement toward ideal direction that minimize (maximize) distance from the current position of the prey (predator) and −1 means the movement toward the opposite direction of an ideal direction. (D) The movement was significantly more linear in the Hunt task than in the Escape task (Fig 1E; left-online study; right-fMRI study). (F) Movement variability in the Hunt and Escape task. Movement variability was defined as an average number of changes in direction per unit time (second; ranged between 0 and 59). Movement variability was significantly higher in the Escape task than in the Hunt task. ***p < 0.001, **p < 0.01, *p < 0.05. The data underlying this figure is available from https://osf.io/k5p3m/ .

We developed an experimental paradigm that could examine the role of the hypothalamus in switching between hunting and escape in humans ( Fig 1A ). Using a virtual arena that mimics experimental environments used in rodents, subjects were asked to either hunt (virtual prey; cyan circle in Fig 1A ) or escape from a computer agent (virtual predator). Participants also performed a control task that similarly involves switching between continuous motor movement as in the experimental task and does not involve survival behaviors such as escaping or hunting. This was done to test whether the hypothalamus is generally involved in switching between continuous movement or it is more specific to switching between survival behaviors such as escaping and hunting. Further, our paradigm allowed us to measure movements on 2 fronts: to determine if anxiety-like behaviors (i.e., thigmotaxis), are more evident for the pre-escape compared to the pre-hunt condition. It also allowed us to investigate the movement trajectories involved in optimal hunting and escape.

One reason for the lack of research on the human hypothalamus is that the hypothalamus contains several small nuclei that are below the resolution of fMRI; consequently, there is no experimental human analog for investigating this region [ 18 ]. To overcome these concerns, we take several approaches: (i) individual-specific segmentation of the hypothalamus was utilized to define hypothalamic ROIs by using a recent deep-learning-based convolutional neural network (CNN) algorithm [ 19 ]; (ii) since the hypothalamic signal is influenced by cerebral spinal fluid (CSF) pulsatility of the ventricle, we regressed out CSF signal [ 20 ]. (iii) We used 2-mm isotropic voxels, which have a better signal-to-noise ratio than smaller (1 to 1.5 mm) voxel imaging [ 21 ]; and (iv) we used multi-voxel pattern analyses (MVPA) to investigate the changing patterns of voxels in the hypothalamus when switching between survival behaviors. Compared to univariate analyses, MVPA is known to have a higher sensitivity to detect neural information by utilizing distributed multi-voxel patterns rather than using mean activation [ 22 ]. We also extended our multi-voxel approach to investigate network-level encoding of switching between survival behaviors by testing multi-voxel pattern synchronization between the hypothalamus and other regions using the informational connectivity (IC) [ 23 – 25 ].

Mammalian studies have emphasized the role of the amygdala, periaqueductal gray (PAG), and prefrontal cortex in threat detection. These regions, however, seem to focus on the salience and scalability of the threat [ 5 – 9 ] rather than determining which survival action to pursue. Experiments on rodents have discovered that the hypothalamus controls both flight and freeze [ 10 , 11 ]. More recently, optogenetic and calcium imaging studies show that the ventromedial hypothalamus elicits the switch between aggression and mounting [ 12 ]. Furthermore, the lateral hypothalamus can induce switching between escaping and hunting behavior [ 13 , 14 ]. These survival behaviors, however, reflect the link between the most basic of motor reactions and cognitive heuristics that involve goal-directed survival behavior such as avoidance of future danger [ 6 ] and predictive hunting behavior [ 15 , 16 ]. Further, the hypothalamus is connected to both the frontal cortex and PAG [ 17 ], making it an intriguing interface between cognitive and reactive strategizing.

A vital determinant of an organism’s longevity is its ability to efficiently select and switch among survival states. This suggests that it would be highly advantageous to evolve a specialized neural circuit that coordinates the transient switching between survival states (e.g., hunting versus escape). The hypothalamus, a region involved in many of life’s basic functions, is one candidate region that is perfectly situated for such purposes. This conserved structure, present across all vertebrate species, has also been linked to many survival behaviors including escape, aggression, and hunting behaviors [ 1 ]. While the hypothalamus controls these basic survival states, it is also a key player in a larger network of brain regions that are involved in interoceptive states, defensive motor output, and learning and memory [ 2 – 4 ]. Given the role of the hypothalamus in self-preserving behaviors in mammals, one key question remains: Does the hypothalamus perform the same function in humans?

Additionally, given that different amygdala subnuclei play different roles, we tested which nuclei play a more active role in conjunction with the hypothalamus (see the S3 Text for details; exploratory analysis) and we found that the hypothalamus-centromedial amygdala (CMA) synchronization (beta = 0.02, t = 1.99, p = 0.047), but not the hypothalamus-basolateral amygdala (BLA) synchronization (p = 0.341) predicts optimal movement coordination after the switching. We also tested whether each of the amygdala subnuclei encodes switching itself or if they directly predict optimal movement coordination after the switching without synchronization with the hypothalamus (please see the S3 Text for details).

To answer this question, we classified the hypothalamic switching signal into 4 cases. The first case is a hypothalamic switching signal without synchronization with other regions (1 if hypothalamic multi-voxel pattern indicates switching and other connected regions’ multi-voxel patterns (that is, PAG, ACC, and amygdala) are non-switching pattern; otherwise 0), and the other 3 cases are hypothalamic switching signal synchronized with switching signal of connected regions (e.g., for the PAG case, 1 if both hypothalamic and PAG multi-voxel pattern indicates switching while patterns of ACC, and amygdala is non-switching; 0 otherwise). We ran 4 regressions to test which case is associated with optimal behavioral coordination after switching (exploratory analysis). Results showed that optimal coordination was enabled only when there is a hypothalamic switching signal synchronized with the switching signal of the amygdala (beta = 0.06, t = 3.18, p = 0.0015; Fig 5B and 5C ) while the hypothalamic switching signal without synchronization nor synchronization with ACC/PAG was not associated with optimal coordination (p > 0.15; Fig 5B ).

Previous studies consistently showed that coordination of survival behavior requires interactions between multiple regions such as an interaction between the hypothalamus, PAG [ 13 , 14 ], and amygdala [ 37 ] and we showed that these regions (hypothalamus, PAG, and amygdala) show synchronization related to switching of survival behaviors. Then, the next question is, if the hypothalamic switching signal is involved in the coordination of optimal behavior after switching, does this coordination depend solely on the hypothalamus or does it require interaction between the hypothalamus and connected regions?

We next tested whether the hypothalamic influence on coordination of the initial behavior after switching is associated with the suppression of the previous behavioral coordination or is associated with initiating a new condition-specific behavior. First, we showed that the degree of condition-specific behavioral coordination in the previous trial (last C p of the previous trial) does not predict the MVPSS of the hypothalamus, supporting the claim that the hypothalamic switching signal is not associated with the suppression of the previous task coordination (p = 0.8864). Second, in an additional regression model to predict initial C p by using both MVPSS of the hypothalamus and the last C p of the previous trial, we found that MVPSS of the hypothalamus significantly predicted the initial C p (beta = 0.17, t = 3.38, p = 0.0007) after regressing out the effect of the last C p of the previous trial, which affected initial C p negatively with marginal significance (beta = −0.03, t = −1.85, p = 0.064). These results suggest that the hypothalamic switching signal affects initial condition-specific behavior coordination after switching that is independent of suppressing previous task coordination.

(A) The regional multi-voxel pattern of switching (MVPSS) predicts condition-specific movement coordination after switching. In the regression analyses, only hypothalamic MVPSS significantly predicted optimal condition-specific movement coordination after switching (pFDR < 0.01). (B) Regions connected to the hypothalamus. (C, D) Switching pattern synchronization between the hypothalamus and connected regions predicts optimal behavior coordination. The hypothalamic multi-voxel pattern of switching (hypothalamic MVPSS) predicted optimal task coordination only when there is a simultaneous multi-voxel pattern of switching in the amygdala (p = 0.003; H-AMY Sync in Fig 5D). Hypothalamic MVPSS was irrelevant with optimal behavior coordination when there was synchronization with connected regions other than the amygdala (H-ACC Sync or H-PAG Sync in Fig 5D) or there was no synchronization with any of the connected regions (H-NoSync in Fig 5D). See Fig 5C for the schematic explanation. **p < 0.01. The data underlying this figure is available from https://osf.io/k5p3m/ . ACC, anterior cingulate cortex; MVPSS, multi-voxel pattern strength of switching; PAG, periaqueductal gray; VMPFC, ventromedial prefrontal cortex.

We found that among the nine ROIs, only the hypothalamic MVPSS predicts the initial C p of the trial after switching (mixed-effect regression to predict initial C p after switch by using MVPSS of the hypothalamus; beta = 0.18, t = 3.76, p = 0.0002; Fig 5A ), while all other 8 ROIs including the PAG and hippocampus that encoded Switch/Stay in MVPA analyses were not associated with task behavior coordination (all p > 0.12; Fig 5A ), meaning that only the hypothalamic switching signal, but not the switching signals from other regions, is associated with the coordination of the condition-specific behavior after switching. However, hypothalamic MVPSS did not predict initial C p after stay behavior (p = 0.6243), showing that the MVPSS is not associated with a continuation of coordination of ongoing task behavior, but is instead associated with coordination of the new task behavior after switching. Furthermore, a strong hypothalamic MVPSS also predicted success in each trial (beta = 0.11, t = 2.06, p = 0.039; a mixed-effect logistic regression to predict trial-by-trial success in the experimental task) while it did not predict success in the control task (beta = −0.06, t = −1.24, p = 0.215), supporting that the hypothalamus is more actively involved in survival behavior switching along with the MVPA results. Lastly, the hypothalamic MVPSS predicted a shortening of the time to achieve stable task coordination (time to achieve C p = 60 after switching; beta = −0.12, t = −3.23, p = 0.001), which also supports that the hypothalamic switching signal is associated with the facilitation of the condition-specific behavior coordination after switching.

We showed that the hypothalamus encodes Switch/Stay information in MVPA analyses. Then, what is the role of the hypothalamus in the switching process? Animal studies show the role of the hypothalamus and PAG in switching between predation and escape [ 13 ], as well as the coordination of those behaviors after survival switching [ 14 ]. Thus, we expected that the hypothalamus and PAG would be involved in the coordination of survival behavior after switching in humans. We expected that the strength of neural signaling indicating switching of survival behaviors measured by a multi-voxel pattern strength of switching (MVPSS; defined using a decoder output strength in MVPA analyses with a range between 0 and 1; 1: 100% switch; 0: 0% switch, same as 100% stay) in the hypothalamus and PAG would facilitate the condition-specific internal movement generation of the following task after switching (Exploratory analyses).

Then, to find hubs connecting regions within the survival behavior switching network, we computed the betweenness centralities of each region of the survival behavior switching network which represent the fraction of all shortest paths that contain a specific node. The top 3 hubs based on BC were ACC, PAG, and amygdala, and the BC of VMPFC was 0 which means this region does not contain any shortest path that connects other regions. Interestingly, the hypothalamus was connected with the top 3 hubs ( Fig 4C and 4D ), which is consistent with previous animal studies showing hypothalamic interaction with the PAG and amygdala. These results also suggest a possibility that the hypothalamus might have communication with regions of the survival behavior switching network through these hubs. Additionally, in a whole brain seed-based IC analysis using hypothalamus as a seed, we found one very small cluster in the right middle temporal lobe (pTFCE < 0.05).

IC was measured between every pair of the 9 ROIs and the network-based statistics (NBS) revealed a network whose multi-voxel pattern synchronization changes with switching between escaping and hunting (survival behavior switching network; Fig 4C and 4D ). In other words, information content (voxel patterns) in regions of the survival behavior switching network were connected. A total of 8 regions except right dorsolateral PFC (DLPFC_R) were included in the survival behavior switching network and they were densely connected, showing a network-level encoding of the switching between survival behaviors. Note that although the amygdala, thalamus, vmPFC, and ACC did not encode Switch/Stay in an MVPA analysis, information in these regions (multi-voxel pattern) was connected to ROIs encoding Switch/Stay at a regional level.

The coordination of survival behavior requires interaction between multiple regions. For example, animal studies showed that lateral hypothalamic projection to PAG underlies the switching between predation and evasion [ 13 , 14 ] as well as the role of interaction between hypothalamus and amygdala in predation [ 33 ] and evasion [ 11 ]. Similarly, we expected that the switching of survival behaviors in humans should be associated with a network-level interaction including a hypothalamus-PAG interaction and a hypothalamus-amygdala interaction. To identify such networks, we measured inter-regional multi-voxel pattern synchronization associated with switching between escaping and hunting using Informational Connectivity [ 23 , 25 , 34 , 35 ] (Exploratory analyses). Like MVPA, an advantage of IC over univariate functional connectivity is that IC utilizes all patterns of responses within regions to encode information that is lost by averaging, which identifies functional connections that cannot be found in univariate functional connectivity analyses [ 23 , 24 ], including connections with small brainstem nuclei [ 36 ]. Furthermore, IC allows us to test regional interactions in terms of specific experimental conditions in [ 24 ] such as switching between escaping and hunting in our study.

(A) ROIs used in this study. Those ROIs included the Hypothalamus (H), Thalamus (T), Hippocampus (HC), Amygdala (AMY), ACC, PAG, left and right dorsolateral prefrontal cortices (DLPFC_L and DLPFC_R), and vmPFC. (B) Decoding accuracy of switch/stay in ROI-based MVPA analyses. The hypothalamus, hippocampus, and PAG significantly encoded switch/stay information (pFDR < 0.05). (C) Multi-voxel functional connectivity (informational connectivity) between a priori selected ROIs. Informational connectivity was computed by using covariation trial-by-trial decoding accuracy between a pair of regions. This resulted in a connectivity matrix between ROIs. NBS revealed a dense network of regions encoding switch information. DLPFC_R was not included in this network (survival-related behavioral switching network). The hypothalamus was connected to ACC, PAG, and amygdala. (D) We computed BC within this network to find hubs connecting regions within the survival behavior switching network. The top three hubs based on BC were ACC, PAG, and amygdala which were connected to the hypothalamus. *p < 0.05. The data underlying this figure is available from https://osf.io/k5p3m/ . ACC, anterior cingulate cortex; BC, betweenness centrality; MVPA, multi-voxel pattern analyses; NBS, network-based statistics; PAG, periaqueductal gray; ROI, region of interest; VMPFC, ventromedial prefrontal cortex.

Our main question was whether the hypothalamus encodes information regarding switching between hunting and escaping behaviors. MVPA analyses were performed on 9 ROIs (preregistered; except amygdala) that were associated with task switching or survival behavior in previous literature [ 9 , 13 , 27 , 29 – 31 ] ( Fig 4A ) using the COSMOMVPA toolbox pipeline [ 32 ]. We found that Switch/Stay information is significantly decodable in the hypothalamus (decoding accuracy: 51.5%; t(20) = 2.49, pFDR = 0.036, one-tailed, in a t test against chance level (50%), p-values were FDR corrected for 9 ROIs; Fig 4B ). In addition to the hypothalamus, we found that PAG and hippocampus (HC) also encoded Switch/Stay information (decoding accuracy: 51.6% and 51.0% for PAG and HC, respectively; all pFDR < 0.05, one-tailed; Fig 4B ). We also asked whether this encoding of switching information in the hypothalamus was specific to switching between survival behaviors (hunting versus escaping) or whether it generalized to other types of motor switching. Using the same MVPA analyses with the control task data, we found that the hypothalamus did not encode switch information in the control task (decoding accuracy: 50.0%; t(14) = −0.05, pFDR = 0.518, one-tailed; Fig 3C ). The searchlight analysis of the control task showed a cluster in the DLPFC ( S2A Fig ; also see S2B Fig for the searchlight results of the experimental task). Note that we tested the possibility that this difference between the experimental task and the control task could be due to a difference in the sample size. We re-conducted the MVPA on the experimental task only using 15 participants’ data who completed the control task which showed similar results ( S2 Text ).

We found that the high C p of the previous trial (just before the switch cue), which would increase the suppression load, negatively affects the initial C p of the current trial (beta = −0.02 and −0.03; p = 0.015 in Expt. 1; marginally significant with p = 0.092 in Expt. 2, respectively, Fig 2G ). Furthermore, the high C p of the previous trial also increased the time to achieve a Condition-specific state (beta = 0.03 and 0.02; p < 0.001 and p = 0.043 in Expts. 1 and 2, respectively, Fig 2G ). These results support that the switching process not only involves the coordination of new tasks but also involves the suppression of the previous condition-specific movement generation process.

Previous studies showed that the task-switching process involves both the initiation of the new task and/or the suppression of ongoing task representation [ 27 , 28 ]. Therefore, we next asked (i) whether the switching process involves both suppression of the previous movement generation process (e.g., escaping) and the coordination of the new task; or (ii) whether the switching process only involves the coordination of the new task. If the former is true (suppression and coordination), switching to the new task would be more difficult if the previous task requires stronger suppression. Therefore, we defined the suppression load of the switching as the C p of the previous trial before switching, since high C p before switching means that the internal movement generation process was more biased into the previous task, which makes it harder to suppress.

We then explored the characteristics of the condition-specific internal movement generation process. C p at the start of each trial was significantly condition-specific, but was still close to the undifferentiated state (0.61 and 0.59 in Expts. 1 and 2; all p < 0.001 in t tests against the undifferentiated state; Fig 2D ) and increased to achieve the Condition-specific state in the later times (median C p of 1 trial = 0.74 and 0.74 in experiments 1 and 2; Fig 2D ). Also, we observed the change of condition-specific movement generation state after switch by testing the change of C p toward the direction of the current task from the previous task (e.g., from escaping-specific state to hunting-specific state; change of C p toward current task = 0.3 and 0.32 in experiments 1 and 2, respectively, all p < 0.001 in one-sample t test, Fig 2F ).

Similar to the behavioral analyses, in both Expts. 1 and 2, we found that internal movement generation process of escaping and hunting is different in terms of model parameter settings ( Materials and methods ) and we were able to decode whether the current task state is hunting or escaping from model parameters with high accuracy (88.7% in Expt. 1; 90.0% in Expt. 2; all p < 0.001 in t test against chance level; all decoding analyses of parameters are exploratory) which means that movement generation states are highly specific to each survival behavior. Then, we quantified the Condition-specific state (C p ), representing how the current movement generation state is specific to the current survival behavior using the decoder output (e.g., if the decoder output is 80% hunt/20% escape and the current trial is Hunt condition, C p = 0.8). C p = 1 means that the current state is specialized for the current task (e.g., hunting in the Hunt condition) while C p = 0 means that the state is specialized for the inappropriate task (e.g., escaping in the Hunt condition). C p = 0.5 means the state is undifferentiated for neither of the tasks. We expected that this condition-specific internal movement generation process would enable appropriate task behavior, which was reflected in increased success rate by the average C p of the trial (beta = 0.42 and 0.2 in Expts 1 and 2, respectively, all p < 0.001; Fig 2E ).

Note that hunting and escaping behavior was controlled by the same sets of parameters (θ, τ, μ). The unit time step parameter θ (possible ranges between 0 and 30; 30 is the length of the model fitting time window; S1 Table and S1B Fig ) controls the degree of the prediction (higher θ will cause more predictive movement as an agent will predict more distant future position) as well as the frequency of movement decision. In an M2 model, θ was fixed to 1 meaning that the degree of prediction and computation is invariant though an agent utilizes prediction. The τ (possible ranges between −1,000 and 1,000; S1 Table and S1B Fig ) controls the deterministicity of the movement decision toward the optimal direction that minimizes the cost function (high τ: high probability of movement decision to the optimal direction), and the μ (possible ranges between 0 and 1; S1 Table and S1B Fig ), a momentum parameter, controls the consistency of the movement between each movement decision (whether subjects consistently move toward the direction decided in time step k step before making new computation in the next time step k+1). This enables us to directly compare the movement generation process between hunting and escaping behavior.

We tested 3 computational models of the hunting and escaping behavior (M1, M2, and M3; preregistered) where in M2 and M3, an agent utilizes their internal predictive model of virtual prey/predator’s movement to guide their escaping of hunting behavior while in the M1, an agent only utilizes current location information of the virtual prey/predator to make movement decision. In the winning model (M3 in Fig 2A ) according to the Bayesian model comparison (Protected Exceedance Probability = 1 in both Expts. 1 and 2; mean BIC for M1, M2, M3: 58.8, 59.7, 21.3 in Expt. 1, and 140.8, 140.9, 73.7 in Expt 2; S1A Fig ), subjects first made a prediction about the future position of the computer agent (virtual predator or prey) based on the computer agents’ previous location, velocity, and acceleration information (see below equation and Methods section for details). Then, subjects made a movement decision toward the direction of the predicted location of the prey (y pred in the below equation) during hunting behavior or made a movement decision away from the direction of the predator during an escaping behavior ( Fig 2A ; simulation results for this model, Fig 2B ).

The average success rate over all the conditions (Escape and Hunt) was 54.8% in Expt. 2 (53.6% in Expt. 1) and the success rate was higher in the Escape condition than in the Hunt condition (59.0% versus 50.7% in Expt. 2, 56.1% versus 51.0% in Expt. 1; p < 0.001 in paired t test of success rate difference between the 2 conditions; preregistered), despite that we applied the same algorithms to allow subjects to achieve a success rate of approximately 50%. Pre-encounter behavioral analysis showed that participants made more thigmotaxic movements in the pre-escape condition compared to the pre-hunt condition (0.52 versus 0.23, t[276] = 27.06, p < 0.001 in Expt. 1; 0.41 versus 0.18, t[ 20 ] = 5.85, p < 0.001 in Expt. 2; preregistered), showing that participants exhibited more anxiety-like behaviors and understood the experimental context (meaning of the boundary colors). There was a significant difference in the average degree of direction change per unit of time, with more variability in the Escape condition (paired t test for movement variability difference between escaping versus hunting; 19° versus 9°, t[276] = 23.77, p < 0.001 in Expt. 1; 18° versus 9°, t[ 20 ] = 7.44, p < 0.001 in Expt. 2; exploratory). Movement direction followed a less predictable direction in the Escape condition compared to the Hunt condition (linearity: 0.55 versus 0.35, t[276] = 21.90, p < 0.001 in Expt. 1; 0.58 versus 0.45, t[ 20 ] = 13.19, p < 0.001 in Expt. 2; preregistered), which is consistent with previous studies showing that movement direction of evasion behavior is less predictable than pursuit behaviors [ 26 ]. We tested whether behavioral differences between hunting and escaping were affected by “boosting” moments by comparing movement linearity and variability between escape and hunt conditions, excluding time points when the predator “boosted.” Results were consistent after excluding those time points: the hunt condition showed higher movement linearity (t[276] = 21.89, p < 0.001 in Expt. 1; t[ 20 ] = 10.24, p < 0.001 in Expt. 2), while the escape condition had higher movement variability (t[276] = 23.27, p < 0.001 in Expt. 1; t[ 20 ] = 6.88, p < 0.001 in Expt. 2). However, there is a possibility that the introduction of boosting exaggerated or decreased the behavioral difference between the 2 conditions. Success rates in the control task were 58.2% (averaged over all conditions), 60.1% (descending order task), and 56.3% (ascending order task).

Individual-specific segmentation of the hypothalamus and its subunits was done using the CNN-based segmentation algorithm of Billot and colleagues, which showed superior accuracy to atlas-based segmentation and manual segmentation. The input was an individual T1 image and the output was an image containing the segmentation of 5 subunits of the hypothalamus. (B) Hypothalamic ROI in our study. Among 5 subunits, hypothalamus ROI was defined by concatenating Posterior, Inferior tubular, and Superior tubular subunits which contain lateral and ventromedial hypothalamus that is known to encode hunting and escape [ 12 , 13 ]. (C) Decoding accuracy of switch/stay in the hypothalamus . The hypothalamus only encoded switching between survival tasks (escape vs. hunting) but not in switching between the control tasks. *p < 0.05. The data underlying this figure is available from https://osf.io/k5p3m/ . CNN, convolutional neural network; ROI, region of interest.

(A) In the winning model, the subject first made a prediction about the future position of the computer agent (virtual predator or prey) based on the computer agent’s previous location, velocity, and acceleration. The subject then made a movement decision toward the direction of the predicted location of the prey during hunting behavior or made a movement decision away from the direction of the predator during escaping behavior. Unit time step θ controls how frequently the player makes computation to adjust their movement τ controls the deterministicity of the movement decision toward direction that minimizes the cost function, and the μ, a momentum parameter, controls the consistency of the movement between each movement decision. (B) Simulation results of player’s movement trajectory using a computational model. The trajectory of the simulated player’s movement was similar to the original player’s movement in both online (left; movement using the keyboard) and fMRI study (right; movement using the joystick). (C) Task-specific internal movement generation process. The current task state was decoded with high accuracy (>88% in both online and fMRI studies by using model parameters representing different components of movement generation processes, showing that the internal movement generation process is highly specific (red dashed line represents the SVM hyperplane which separates Escape (pink dots) and Hunt (orange dots) conditions). Condition-specific state (C p ) was defined to represent a degree of the current condition-specific internal movement generation process. This was computed using the SVM decoder output, such that a farther distance from the hyperplane to the appropriate direction of the current task means higher C p . For example, in the hunting task, C p becomes higher as the internal state goes farther in the direction of the orange arrow. (D) Change of the condition-specific movement generation state in one trial. C p is low at the start of the experimental task. (E) Relationship between condition-specific internal generation state and success in each trial. In the logistic regression, higher C p significantly predicted success in both online and fMRI studies. (F) Transition effect after switching. Condition-specific states significantly changed after switching. (G) Effect of suppression load on switching. High Condition-specific state (high C p of the previous trial) interfered with initial task movement coordination after switching (left) and increased time to achieve stable movement coordination (right). ***p < 0.001, *p < 0.05, #p < 0.1. The data underlying this figure is available from https://osf.io/k5p3m/ . SVM, support vector machine.

We first examined the characteristics of escaping and hunting behavior in terms of the difference in participants’ movement patterns and preparation behavior. Then, we fit a computational model that explained the generation process of both escaping and hunting behavior ( Fig 2A ) and tested the computational mechanism of transitions in behavioral generation after switching by using this model. At the neural level, we first asked whether the hypothalamus encodes Switch/Stay information and how hypothalamic interaction with other regions is involved in Switch/Stay information. We performed MVPA on nine ROIs including the hypothalamus (see ROI selection and segmentation for MVPA analyses in Methods section; Fig 3A and 3B ) and then performed Informational Connectivity analysis [ 23 ] between the 9 ROIs to show how the hypothalamus interacts with other regions to encode switching of survival behaviors. MVPA analyses were also conducted for the control task to compare survival switching with general attentional-motor switching. Finally, by using a generative model of behavior, we investigated whether and how the hypothalamus initiates the coordination of the survival behavior after switching in conjunction with other regions. Experiments 1 and 2 were preregistered ( https://osf.io/ge8c3 and https://osf.io/kx5af ).

In Expt. 1, we behaviorally tested 277 Prolific participants (details in S1 Text ). In Expt. 2, we scanned 21 participants with fMRI as they performed 4 runs of a novel hunting-escape switching task inside the scanner. Participants were scanned for approximately 4 h each over 2 days (a total of 484 trials). Subjects controlled their avatar (player; yellow circle in Fig 1A ) that was placed in a 2D circular field ( Fig 1A ). Subjects either hunted the computer agent (virtual prey; cyan circle in Fig 1A ) or escaped from the computer agent (virtual predator). Subjects were shown a virtual arena with either a pink or orange boundary for 1 s, the color of which signaled the type of the next task (escape or hunt). The boundary color screen constitutes the Switch/Stay screen. One boundary color signaled the “Hunt Condition,” where the participants should chase and attempt to capture a virtual prey that was programmed to run away from them. The other boundary color signaled the “Escape Condition,” where the participants should prepare to escape a virtual predator which was programmed to attack them. Additionally, in Expt. 2 participants completed the control task to test whether hypothalamus involvement is specific to the survival behavior switching or is related to a general switching of the continuous behaviors (see Experimental paradigm for the control task in Methods section).

Discussion

The role of the hypothalamus in survival behavior has been extensively investigated in non-human animals [11–14] while the evidence supporting the role of the human hypothalamus in survival behavior is very rare. In this preregistered study, we showed the human hypothalamus coordinates switching between survival behaviors by introducing a novel ecologically valid experimental paradigm and by optimizing the acquisition and analyses of hypothalamic BOLD signal. MVPA results showed that the hypothalamus encodes switching between hunting and escaping behaviors as in other animal studies and this hypothalamus is connected to PAG, amygdala, and ACC within the survival behavior switching network. Importantly, by adapting computational modeling of continuous escaping and hunting behavior, we found that the hypothalamus is the only region in the survival behavior switching network that is associated with facilitating the coordination of survival behavior after switching and this function of the hypothalamus requires synchronization with the amygdala.

Behaviorally, we utilized an ecologically valid experimental paradigm that resembles tasks in animal studies of survival behavior [13,16]. In both Expts. 1 and 2, we showed both similar and divergent computational mechanisms underlying human hunting and escaping behavior. The winning model showed that participants made predictive movements by a computation of future prey/predator position both in hunting and escaping behavior, which extends previous studies showing predictive predation behavior in the macaque [16]. On the other hand, model parameter comparison also showed that the behavioral generation process of hunting and escape is highly divergent and specific to each state (decoding accuracy >88%). Compared to hunting behavior, escaping behavior was more stochastic which is consistent with other modeling results in other species [26]. The unit time step of movement decision in escaping behavior was also shorter. These stochastic and short-time step decisions make escaping movements more unpredictable to the predator and enable more flexible coordination of behavior to avoid the predator (e.g., by frequent directional change). We also explored the computational mechanisms of switching between escaping and hunting by defining the degree of the Condition-specific state (C p ). Higher C p was associated with successful hunt/escape and switching also induced appropriate change of Condition-specific state. Furthermore, we showed that successful transition is affected by the suppression load of the previous trial, showing that similar to the classical task switching process, survival switching also requires suppression of the ongoing task process [27,28].

To our knowledge, this is the first human study showing the role of the hypothalamus in switching and coordination of survival behavior. Small size, closeness to ventricles, and low contrast to adjacent tissues were the obstacles to the precise identification of the hypothalamus and its subnuclei. We segmented the hypothalamus at the level of subnuclei by applying state-of-the-art deep learning-based segmentation [19] that outperforms conventional manual segmentation [38] and atlas-based segmentation [39]. The hypothalamus is composed of numerous different nuclei with largely heterogeneous functions such as thermoregulation in the anterior hypothalamus [40], as well as circadian rhythm regulation in the suprachiasmatic nucleus [41]. Thus, the inclusion of hypothalamic subregions that are not related to survival behavior would decrease the sensitivity of detection in our study. Therefore, by applying subnuclei-level segmentation of the hypothalamus, we constructed a hypothalamic ROI focused on the lateral and ventromedial hypothalamus that were previously identified as key hubs of survival behavior in animal studies [12,13]. We excluded other parts of the hypothalamus that have been less relevant to animal survival behavior. In addition, we regressed out CSF and other potential confounders and used a short TR for better statistical modeling of the hypothalamic signal. MVPA after these processing steps showed that the human hypothalamus indeed encodes switching between hunting and escape, as in animal studies [13,14]. Unlike in the survival behavior switching in the experimental task, the hypothalamus was not involved in the control task. Instead, using searchlight MVPA analyses, we found a small cluster in the DLPFC that encodes motor switching, which is consistent with previous studies of non-survival task switching [27,42]. Furthermore, the IC analysis of the control task showed that network encoding the switching in the control task does not contain the hypothalamus (S3 Fig and S2 Text). Lastly, the hypothalamic MVPSS in the control task did not predict success in the trial after the switching. These results support that the hypothalamus is more likely to be involved in the switching of survival behaviors rather than involved in the switching of general continuous behaviors. Lastly, it could be possible that the hypothalamic involvement in switching found in our study might be related to a change of arousal or emotional valence (see S4 Text for the detailed discussion). However, since we concatenated both switching from hunt to escape and switching from escape to hunt it is less likely that the difference in emotional valence or arousal level between hunt and escape influenced the hypothalamic switching signal. Also, we showed that the electric shock that increases arousal was not associated with the hypothalamic switching signal (S4 Text).

We also found that the PAG encodes the switching with the hypothalamus and hypothalamus-PAG multi-voxel activation patterns of switching are functionally coupled. Previous studies consistently showed the importance of hypothalamus-PAG coupling in survival behavior [6,13,14,43]. For example, Li and colleagues showed that an activation of PAG-projecting GABA neurons in the lateral hypothalamus drives predatory attack in mice while an activation of PAG-projecting glutaminergic neurons drives evasion [13]. Recently, Rossier and colleagues further showed that PAG-projecting GABA neurons of the lateral hypothalamus block defensive response encoded in the PAG in mice [14]. The ventromedial hypothalamus also heavily projects to the PAG and coordinates survival behavior [44]. However, unlike previous animal studies, neither the strength of the switching signal in the PAG quantified by MVPSS nor hypothalamus-PAG pattern synchronization were directly associated with optimal movement coordination after switching in our study. Hypothalamic MVPSS and hypothalamus-amygdala pattern synchronization were associated with optimal movement coordination. Unlike previous animal studies, our results showed that the neuronal pattern of switching between hunting and escaping in humans cannot be entirely explained as an interaction between the hypothalamus and PAG, which known as a “reactive” survival circuit that mainly utilizes the PAG as a motor pattern generator for pre-programmed behavior but also involves an interaction between classic the hypothalamus-PAG circuit and the regions involved in “cognitive switching” area that previously known to encode the switching of abstract task representation [27]. These regions include DLPFC, thalamus, hippocampus, and ACC [27,45]. This claim is supported by our computational modeling results suggesting that subjects utilized an internal model of the predator/prey to make predictive escaping/hunting movements. Also, note that experimental context in previous animal experiments usually includes an interaction with close predator/prey which requires a reactive behavior [10,13,44]. Previous animal studies show that the hypothalamus-PAG circuit is required for both simple/reactive and complex/cognitive survival behavior [11]. However, in a “reactive” context, the PAG acts as a direct motor pattern generator such that the direct stimulation of the PAG can produce pre-programmed survival behavior [11]. On the other hand, in a “cognitive” context, the hypothalamus-PAG circuit was necessary but not sufficient such that the direct stimulation of the PAG could not generate the strategic survival behavior under a sophisticated survival situation [11], which is consistent with our results showing significant involvement of the hypothalamus-PAG circuit in encoding switching without direct association between PAG and movement coordination. To our knowledge, the role of PAG in this “cognitive” context is still ambiguous, but one recent study suggested its role in increasing cognitive control [46].

In addition to the hypothalamus and PAG, MVPA results showed that the hippocampus encodes switching, which is a critical region of model-based planning [6] and the flexible encoding of the novel context by suppressing the old context [47]. Furthermore, we found a survival behavior switching network which was a network having dense interaction between multiple regions. Unlike in previous animal studies of reactive task switching, this survival behavior switching network not only involved hypothalamic-PAG interactions but also included interaction between (i) regions known for “cognitive” task switching such as DLPFC (left) [42], thalamus [27,48], and ACC [49]; and (ii) cognitive fear circuit regions including hippocampus and ventromedial prefrontal cortex [9] that is associated with strategic and deliberate survival behavior; and (iii) amygdala. The thalamus is known to control task representation in the PFC such that the mediodorsal thalamus suppresses task-irrelevant representation while augmenting task-relevant representation [27,48]. The ACC helps increase cognitive control when there is an increased need for control, as in a situation involving conflict [50] or task switching [49], by computing control demands from the integration of information from multiple regions [50]. In our study, ACC was the hub that connected all regions within the network both in the main task and in the control task. Overall, these findings again support our claim that switching between human survival-related behavior requires recruitment of a “cognitive” task survival behavior switching network and behavioral coordination network in addition to the hypothalamus-PAG circuit that is necessary for survival behavior regardless of the complexity of behavior.

A lingering question is what is the specific role of the hypothalamus in survival switching? Among the regions of the survival behavior switching network, the hypothalamus was the only region that was directly associated with movement coordination after switching such that a strong hypothalamic switching pattern was associated with higher condition-specific movement coordination after switching. However, hypothalamic MVPSS was not affected by the suppression load of the previous trial. This suggests that the hypothalamus might be a region that initiates motor pattern generation after switching while other regions are more involved in non-motor and abstract parts of the switching process, such as integrating sensory information, encoding emotional states, suppressing old task representations, and building a new task representation. Although evidence in human studies supporting the hypothalamus as a survival motor coordination is very rare, many animal studies point to the hypothalamus as a movement initialization region [10,13,14]. The hypothalamus was connected to the PAG, amygdala, and ACC and we found that the hypothalamic switching signal is associated with motor coordination only when there is a switching signal in the amygdala. Especially, among the amygdala subnuclei, only the synchronization with the CMA, but not the synchronization with the BLA was associated with the motor coordination.

Currently, the known role of the amygdala in defensive behavior, especially the central amygdala, is that it integrates sensory cues to represent the current internal state and broadcasts that state to other brain regions including motor pathways, which can enable, state-dependent behavioral responses [31,51]. Furthermore, the amygdala is the upstream of the hypothalamus [52] and the projections from the central amygdala to the hypothalamus coordinate both aggression and avoidance [53,54] as well as controlling approach and avoidance [55]. Based on these prior studies, and our findings showing hypothalamic switching signal synchronized with the amygdala, especially CMA, to predict optimal motor coordination while the switching signal of the amygdala itself did not affect motor coordination, we speculate that the amygdala processes the sensory cue information that informs the switching or staying of ongoing behavioral states and the hypothalamus receives this information to produce appropriate task behavior after switching. However, we note that we cannot determine a causal relationship between the hypothalamus and amygdala in this study and there is also a possibility that the hypothalamus is an input region to the amygdala that coordinates complex survival behavior after switching.

In contrast to the hypothalamus, the hippocampal MVPSS was associated with the suppression load of the previous trial but not with the current C p , suggesting its potential role in suppression rather than motor pattern generation in survival-related behavioral switching. This is consistent with previous literature showing the hippocampus’ role in inhibiting the old context, especially through an interaction with the prefrontal cortex [47,56], which aligns with the hippocampal interactions with DLPFC_L and VMPFC during switching in our IC analyses. However, the role of these interactions in survival-related behavioral switching remains unclear in our study, and this would be an interesting topic for future studies. Although our results provide some indirect evidence of the potentially different roles of the hypothalamus and hippocampus in survival-related behavioral switching, our current task does not clearly distinguish between the suppression-related process and motor coordination independent of the suppression-related process. Therefore, future studies with tasks that clearly differentiate between suppression and coordination within the switching process (for both the survival-related main experimental task and the control motor task) will be needed to further address this question.

Lastly, in this study, we demonstrated that the human hypothalamus plays a role in switching between hunting and escaping, interacting with cognitive task-switching regions. While the hypothalamus was not implicated in the control task, suggesting it may not be part of the classic “domain-general” switching network involving the DLPFC and ACC, which is known to be involved in various domain switches such as reversal learning and set-shifting [45], our findings, when combined with existing animal literature, suggest that the hypothalamus may be involved in switching between general human survival-related states. These states encompass brain functions closely tied to maintaining homeostasis and reproduction [7]. For example, animal studies have shown that the hypothalamus controls switching between mounting and attacking, as well as between feeding and seeking social contact [12,57]. While human survival-related behavior occurs in more complex social contexts requiring intricate, strategic planning compared to innate animal behaviors, we propose that the hypothalamus may interact with both cognitive task-switching areas and brain networks coordinating complex social interactions to facilitate switching between complex survival states. This area presents an intriguing avenue for future human experimental studies.

In conclusion, we found that the human hypothalamus encodes switching between hunting and escaping behavior. Similar to animal studies, the hypothalamus-PAG circuit was involved in the encoding of switching. However, human survival switching also involved network-scale interaction between regions that have been known for “cognitive” task switching and survival behavior, and this was consistent with our behavioral findings showing our human subjects utilized internal predictive models of prey and predator to coordinate their hunting and escaping movement, respectively. Finally, the hypothalamus was the only region within the survival behavior switching network that was associated with optimal survival coordination which required a synchronization with the amygdala which we posit plays a role in integrating sensory information to represent the current state and conveys this information to the hypothalamus. These findings extend our understanding of the human hypothalamus from a region that regulates our internal bodily states to a region that switches survival behaviors and coordinates strategic survival behaviors.

[END]
---
[1] Url: https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3002624

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/