Avoidance Learning, a crucial survival mechanism, allows organisms to preemptively evade unpleasant or harmful situations. This active process is more than just reacting to fear; it’s about learning to predict and prevent negative experiences. Researchers are continually seeking to understand the neural pathways and brain regions that orchestrate this complex behavior. A recent study sheds new light on the role of the medial prefrontal cortex (mPFC) in avoidance learning, utilizing an innovative two-dimensional active avoidance paradigm in mice. This article delves into this groundbreaking research, exploring how the brain learns to avoid, and what this implies for our understanding of learning and behavior.
I. A Novel Paradigm for Studying Avoidance Learning
Traditional studies of avoidance learning often involve simple, one-dimensional tasks. Recognizing the need for a more nuanced approach, scientists developed a novel 11-day instrumental conditioning paradigm for mice, termed “two-dimensional active avoidance.” This paradigm, detailed in Figure 1a, is structured into three phases: habituation (Day 1), active avoidance training (Days 2–9), and extinction (Days 10–11). Each session consists of 50 trials, initiated by a tone (8 kHz, 80 dB, up to 10 seconds). During active avoidance training, this tone is followed by a mild foot shock (0.2 mA, up to 5 seconds).
Figure 1: The two-dimensional active avoidance paradigm and recording of prefrontal population activity. a, Task schematic and time course of the 11-day learning paradigm. Tasks 1 and 2 are defined by shuttling along the x and y axes of the shuttle box, respectively.
The innovative aspect of this paradigm lies in the “safe zone.” In each trial, half of the experimental chamber is designated as safe, its location dynamically determined by the mouse’s starting position and the specific task. Mice learn to avoid the shock by moving into this safe zone during the tone, immediately ending the trial (Figure 1b).
Figure 1: The two-dimensional active avoidance paradigm and recording of prefrontal population activity. b, Trial structure and illustration of the different trial types (avoid and error).
Days 2–4 (Task 1) require mice to shuttle along the x-axis to reach safety, while days 5–9 (Task 2) necessitate shuttling along the y-axis, testing behavioral flexibility. Failure to reach the safe zone during the tone results in a foot shock, after which platforms rise to allow escape (Methods section in the original article provides further detail). Habituation and extinction phases involve tone presentation without shocks, with safe zone logic randomly following Task 1 or 2. Trials successfully avoided by shuttling during the tone are termed “avoid trials,” while those with shock delivery are “error trials” (Figure 1b).
II. Behavioral Adaptation and Learning Curves
During Task 1, the proportion of avoid trials significantly increased, from an initial 40% to an impressive 84% (Figure 1c). This demonstrates rapid learning of the avoidance behavior. Upon switching to Task 2 on day 5, performance initially dropped to 44%, indicating the task shift’s impact. However, by the end of Task 2, mice recovered, reaching an 81% avoid trial rate. This recovery highlights the mice’s ability to adapt and relearn avoidance behavior in a new spatial dimension.
Figure 1: The two-dimensional active avoidance paradigm and recording of prefrontal population activity. c, Percentage of successful avoid trials per active avoidance session (n = 12 mice, mean ± s.e.m.). d, Shuttle rates for X shuttle (solid line) and Y shuttle (dashed line) across 11 days of learning (n = 12 mice, mean ± s.e.m.).
This adaptation was further evidenced by analyzing shuttle direction (Figure 1d). As mice transitioned from Task 1 to Task 2, Y-shuttling rates increased dramatically from 19% to 81%, while X-shuttling rates correspondingly decreased from 84% to 27%. This clear shift in behavior underscores the successful learning and adaptation to the two-dimensional active avoidance paradigm.
III. Neural Correlates in the Medial Prefrontal Cortex (mPFC)
To investigate the neural basis of this avoidance learning, researchers focused on the mPFC, a brain region known for its role in complex behaviors and learning. They employed miniaturized fluorescence microscopy to record neuronal activity in freely behaving mice. Genetically encoded calcium indicator GCaMP6m was expressed in excitatory neurons of the prelimbic area of the mPFC (Figure 1e). This allowed for tracking the activity of thousands of mPFC neurons (3,333 neurons across 12 mice) throughout the 11-day paradigm (Figure 1f,g).
Figure 1: The two-dimensional active avoidance paradigm and recording of prefrontal population activity. e, Miniaturized (single photon) population calcium imaging in freely behaving mice. GCaMP6m was genetically expressed in pyramidal neurons, and a GRIN lens was implanted above the PL. Scale bar: 1 mm. f, Cell map of an example animal. Scale bar: 100 μm. g, Calcium fluorescence traces of ten example neurons on days 1, 6 and 11.
Analysis of neuronal activity during avoid trials, aligned to both tone start and shuttle start (Figure 1h), revealed that a significant portion of mPFC neurons (54%) showed altered activity during the trial compared to baseline (Figure 1j). This proportion was lower during habituation (15%) and extinction (26%), suggesting mPFC involvement is specific to active avoidance learning. The overlap in active neuron subsets was high across avoidance sessions (60%) but low during extinction (28%; Figure 1k), further supporting mPFC’s role in encoding learned avoidance behaviors.
Figure 1: The two-dimensional active avoidance paradigm and recording of prefrontal population activity. h, Top, mouse speed for five exemplary avoid trials including markers for three reference time points (tone start, shuttle start and tone end). Bottom, distributions of latencies from tone start to shuttle start and shuttle start to tone end over all avoid trials (days 2–9, 12 mice). i, Top, calcium fluorescence traces of one example neuron aligned to tone start (left) or shuttle start (right). Trials are sorted according to trial length. Bottom, trial-averaged neuronal activity of the same neuron. j, Percentage of trial-responsive neurons across 11 days of learning (n = 12 mice, mean ± s.e.m.). k, Overlap of trial-responsive subpopulations across 11 days, where the overlap between days i and j is defined as *ni and j/((ni + n*j)/2). l, Trial-averaged response of four example neurons aligned to tone start (left) or shuttle start (right). OFC, orbitofrontal cortex; IL, infralimbic cortex; PL, prelimbic cortex; D1, day 1; D2–D4, days 2–4; D5–D9, days 5–9; D10–D11, days 10–11.
Individual neuron responses were diverse (Figure 1l), some aligning with the tone, others with avoidance action, and some exhibiting complex temporal dynamics. This heterogeneity prompted researchers to employ population-level decoding approaches to disentangle the neural signals.
IV. Decoding Avoidance Actions from mPFC Activity
To overcome the challenge of limited trials per subject in neuroscience research, the study employed a novel approach to align neural recordings from different mice into a joint subspace (Figure 2a). This allowed for the joint analysis of trials across subjects, enhancing statistical power for decoding analyses. Dimensionality reduction confirmed that neural activity could be effectively represented in a low-dimensional space, and task-related dynamics were consistent across subjects.
Figure 2: Subject alignment and prediction of avoidance actions. a, Illustration of the neuronal subspace alignment procedure across animals. b, Schematic representation of the decoding approach to predict avoidance behavior from mPFC neuronal activity.
Support vector machine (SVM) decoders were trained to discriminate between neural activity during avoid and error trials (Figure 2b). Decoding accuracy increased leading up to the shuttle action, exceeding chance levels before shuttle start (Figure 2c). This indicates that mPFC population activity contains predictive information about impending avoidance actions.
Figure 2: Subject alignment and prediction of avoidance actions. c, Decoding accuracies across time for decoding of avoid versus error trials (AV, black) and ITI shuttles versus random ITI periods (ITI, magenta; mean and 95% CIs for 80 repetitions of the analysis using different samples of trials). Black bar indicates significant differences between the AV and ITI settings based on nonoverlapping CIs. d, Same as c, but for decoders trained using the animals’ speed extracted from video tracking data.
To determine if this predictive information was specific to avoidance or simply reflected general motion, decoders were also trained to distinguish spontaneous shuttles during intertrial intervals (ITI) from random ITI periods. While ITI shuttle decoding accuracy was also above chance, it was significantly lower than for avoid trials (Figure 2c). Crucially, decoding based on motion tracking data showed no difference between avoid and ITI shuttles (Figure 2d). This demonstrates that mPFC activity encodes avoidance-specific information beyond general motion signals.
V. Disentangling Motion, Avoidance, and Sensory Signals
To further dissect mPFC activity, researchers used principal component analysis (PCA) to identify dimensions representing motion-related activity, focusing on variance during ITI shuttles. Removing these motion dimensions from the joint subspace significantly reduced ITI decoding accuracy, but had a lesser impact on avoid versus error decoding (Figure 3a,b). This suggests that motion-related activity resides in a low-dimensional subspace distinct from avoidance-specific activity.
Figure 3: Decomposition of mPFC population activity into dimensions related to motion, avoidance actions and tone stimuli. a, Mean accuracy of neural decoders for ITI (left) or avoid shuttles (right) after the progressive removal of up to four motion dimensions (n = 80 repetitions). b, Drop in time-averaged accuracy (−3 s to 1 s) of ITI and avoid decoders from a with respect to the baseline setting (0 dimensions removed).
An iterative decoding approach was then used to identify avoidance-specific coding dimensions. After removing motion dimensions, SVM decoders were trained to discriminate between avoid and error trials. Removing the identified “avoidance dimensions” strongly reduced decoding performance for avoid versus error trials, but not for ITI shuttles (Figure 3c,d). This confirmed the existence of a low-dimensional subspace dedicated to avoidance-specific activity, orthogonal to motion-related dimensions.
Figure 3: Decomposition of mPFC population activity into dimensions related to motion, avoidance actions and tone stimuli. c, Mean accuracy of neural decoders for ITI (left) or avoid shuttles (right) after the progressive removal of up to four avoid dimensions (n = 80 repetitions). d, Drop in time-averaged accuracy (−3 s to 1 s) of ITI and avoid decoders from c.
Finally, tone-related activity was investigated. Decoders trained to discriminate between tone and non-tone periods showed high accuracy (Figure 3g). Removing “tone dimensions” significantly reduced tone decoding accuracy (Figure 3g,h). This decomposition revealed five orthogonal dimensions in mPFC activity: two for motion, two for avoidance, and one for tone, providing a compact representation of task-related neural activity (Figure 3e).
Figure 3: Decomposition of mPFC population activity into dimensions related to motion, avoidance actions and tone stimuli. e, Schematic representation showing the progressive decomposition of the joint subspace into five coding dimensions and a residual space. f, Schematic representation illustrating tone versus BSL decoding. g, Tone decoding accuracies after progressive removal of up to two-tone dimensions. h, Drop in time-averaged tone decoding accuracy (0–4 s) of the decoders from g and drop in VE for the respective decoding dimensions. In b, d and h, lines and shaded areas correspond to mean and 95% CIs for 80 repetitions, and vertical dotted lines correspond to the number of dimensions chosen for the subspace decomposition. Black bars in b and d indicate significant differences between the AV and ITI settings based on nonoverlapping CIs. BSL, baseline.
VI. Dynamic Activity in Coding Dimensions and Task-Related Changes
Projecting neural activity onto these five dimensions revealed distinct temporal dynamics during avoid and error trials (Figure 4a). Motion and avoidance dimensions showed similar trajectories, increasing towards shuttle start, with stronger activation during avoid trials. The tone dimension responded strongly to tone onset. These five dimensions captured a significant portion of the variance (91.9%) in avoid and error trials (Figure 4b). During ITI shuttles, motion dimensions were dominant, while avoidance dimensions showed minimal activity (Figure 4a,c). This further distinguishes motion-related activity from avoidance-specific signals.
Figure 4: Characterization of low-dimensional task-related population activity. a, Mean projections (n = 80 repetitions) of neural data onto the five coding dimensions for avoid and error trials (top row) and ITI shuttles and random ITI periods (bottom row). b, VE by individual dimensions for avoid and error trials (distributions over 80 repetitions, mean and 95% CIs). c, Same as b for ITI shuttles.
Analysis of dimension weight vectors revealed that most neurons exhibited mixed selectivity, contributing to multiple coding dimensions (Figure 4e,f). This suggests that the identified coding dimensions are encoded by distributed activity across a population of neurons, rather than distinct subpopulations.
Figure 4: Characterization of low-dimensional task-related population activity. d, Pearson correlation coefficient between pairs of coding dimension projections (avoid and error projections concatenated; mean over 80 repetitions). e, Weight distributions for four example cells. We calculated how individual cells contributed to the five coding dimensions and normalized the weight values such that the sum of their absolute values was equal to 1. The following four examples show different types of distributions: (1) selective, (2 and 3) mixed-selective and (4) nonselective. f, Distribution of weight entropy values over all recorded cells. Black dots indicate the four example cells from e, and the dashed line indicates the maximum possible entropy. Ent., entropy.
Examining activity across the 11-day paradigm showed that motion-related activity was dominant during habituation and extinction, but its relative contribution decreased during active avoidance learning (Figure 5a–c). Conversely, tone and avoidance-specific activity emerged during active avoidance sessions, demonstrating learning-related changes in mPFC activity.
Figure 5: Emergence of low-dimensional, task-related neuronal signals in mPFC. a, Mean projections (n = 80 repetitions) of neural data onto the five coding dimensions for trials with and without shuttling during habituation (day 1), task 1 (days 2–4), task 2 (days 5–9) and extinction (days 10–11). b, Absolute variance within each coding dimension. c, Relative VE by the five coding dimensions across the 11-day learning paradigm. Mean over 80 repetitions.
VII. Task Switching and Avoidance-Specific Activity
The study further investigated how the task switch (from x-axis to y-axis shuttling) impacted avoidance-specific activity. Avoidance dimensions appeared to be differentially engaged in Task 1 and Task 2 (Figure 5a–c). Decoders trained to discriminate avoid and error trials were tested on individual sessions. The “avoid 1” decoder performed well in Task 2 and also showed above-chance performance in Task 1 (Figure 5d). However, the “avoid 2” decoder was effective in Task 2 but not Task 1 (Figure 5e). This suggests that “avoid 1” activity generalizes across tasks, while “avoid 2” activity is task-specific, emerging to accommodate the altered avoidance behavior in Task 2. The task switch appears to modify mPFC coding of avoidance actions by adding a layer of task-specific avoidance activity.
Figure 5: Emergence of low-dimensional, task-related neuronal signals in mPFC. d, Avoid versus error decoding accuracy (time-averaged for the 2 s preceding shuttle start) for the decoder that was used to define the avoid 1 dimension (mean and 95% CIs for 80 repetitions). Decoders were trained with data from tasks 1 and 2 and separately evaluated with test data from individual days. The black bar indicates performance that is significantly above chance based on 50% not being included in the CI. e, Same as d, but for the decoder that was used to define the avoid 2 dimension (the avoid 1 dimension was already removed). T1, task 1; T2, task 2.
Task decoders trained to distinguish between Task 1 and Task 2 trials based on the five coding dimensions showed higher accuracy for avoid trials compared to error trials or ITI shuttles (Figure 6b,c). Task decoding accuracy increased towards avoidance actions during avoid trials, but less so during error trials, indicating that the task switch specifically altered neural dynamics related to successful avoidance. The “avoid 2” dimension showed the highest task-decoding accuracy in avoid trials (Figure 6d), further emphasizing its role in task-specific avoidance coding.
Figure 6: The task switch affects avoidance population coding. a, Schematic representation of task decoding, where we trained decoders to distinguish data from tasks 1 and 2 trials. b, Task-decoding accuracy for time-dependent decoders trained to discriminate between task 1 and task 2 data separately for avoid trials, error trials or ITI shuttles (mean over 80 repetitions). c, Temporal average (−3 s to 1 s) for data from b (mean and 95% CIs for 80 repetitions). Significant differences are reported based on nonoverlapping CIs. d, Time-averaged task-decoding accuracy (−3 s to 1 s) for decoding performed with individual dimensions for avoid trials, error trials and ITI shuttles (mean and 95% CIs for 80 repetitions).
Analyzing behavioral variability in Task 2 (Y shuttles vs. XY shuttles) revealed that decoders were better at distinguishing Task 2 XY shuttles from Task 1 X shuttles than from Task 2 Y shuttles (Figure 6e). The “avoid 2” dimension carried more information for distinguishing Task 1 X shuttles from Task 2 XY shuttles, while motion dimensions were more informative for differentiating Task 2 Y shuttles from Task 2 XY shuttles (Figure 6f). This suggests that “avoid 2” dimension reflects abstract task-related differences in avoidance actions, beyond simple motion variations.
Figure 6: The task switch affects avoidance population coding. e, Decoding of different types of avoid trials. T1 X shuttles and T2 XY shuttles differ in motion and task, and T2 Y shuttles and T2 XY shuttles also differ in motion but follow the same task rule. Time-averaged accuracy (−3 s to 1 s) for decoders using all five coding dimensions (mean and 95% CIs for 80 repetitions). f, Shuttle-type decoding, as in e, for decoders trained on individual dimensions. The single asterisk denotes significance derived from nonoverlapping CIs.
VIII. Modulation of Sensory Processing by Avoidance Behavior
Despite the independent coding dimensions, the study found that tone processing in the mPFC is modulated by avoidance behavior. Tone dimension activity was generally correlated with tone presentation (Figure 7a). However, a notable drop in tone dimension activity occurred shortly after shuttle start (Figure 7b,c), even though the tone remained on. This drop was also observed during incorrect X-shuttles in early Task 2 trials (Figure 7d,e), suggesting that it is linked to the execution of the learned avoidance action, regardless of its immediate outcome in the changed task context. This indicates that mPFC sensory representations are not static, but are dynamically modulated by learned behavioral responses associated with stimulus termination and avoidance of aversive outcomes.
Figure 7: Avoidance behavior affects mPFC tone encoding. a, Tone dimension projection over a 10-min time window from an example session of one subject. Tone presentations are marked in gray. b, Mean projection (n = 80 repetitions) of neural data onto the tone dimension during avoid trials aligned to tone start (left) and shuttle start (right). The maximum drop point (purple star) refers to the time step before the maximum decrease of tone dimension activity between two consecutive time steps (5 Hz). c, Same as b, but aligned to tone start (left) and tone end (right). d, Tone dimension activity for task 2 X-shuttle trials (error trials with an incorrect shuttle). Trials were either aligned to the tone start (left) or shuttle start (right). e, Trial averages for data from d (black line) aligned to tone start (left), shuttle start (middle) and tone end (right). The orange line represents trials without shuttles. For these trials, the shuttle start point was randomly sampled to match the distribution of shuttle starts from the shuttle trials.
IX. Conclusion: Implications for Understanding Avoidance Learning
This research provides a detailed neurophysiological characterization of avoidance learning in the mPFC using a novel two-dimensional active avoidance paradigm. The findings reveal that mPFC activity during avoidance learning is multifaceted, encompassing distinct neural dimensions for motion, avoidance actions, and sensory processing. The study demonstrates that avoidance-specific neural activity is separable from general motion signals and that the mPFC dynamically adapts its coding strategy when avoidance requirements change. Furthermore, sensory processing in the mPFC is not simply a passive reception of information, but is actively modulated by learned avoidance behaviors.
These findings have significant implications for our understanding of avoidance learning and the role of the mPFC in flexible behavioral adaptation. By disentangling the neural components of avoidance behavior, this research paves the way for future investigations into the precise mechanisms of avoidance learning and potential therapeutic interventions for disorders involving maladaptive avoidance, such as anxiety disorders and PTSD. The two-dimensional active avoidance paradigm itself offers a valuable tool for future research, allowing for more complex and ecologically relevant studies of how organisms learn to navigate and avoid threats in their environment.
