Salient Experiences are Represented by Unique Transcriptional Signatures in the Brain

Inducible transcription is essential for consolidation of salient experiences into long-term memory. However, the question of whether inducible transcription relays information representing the identity of the experience being encoded, has not been explored. To this end, we have analyzed transcription across multiple brain regions, induced by a variety of rewarding and aversive experiences. Our results define robust transcriptional signatures uniquely characterizing individual salient experiences. A subset of these induced transcriptional markers suffice for near-perfect decoding of the identity of recent experiences at the level of individual mice. Furthermore, experiences with shared attributes display commonalities in their transcriptional representation, exemplified in the representation of valence, habituation and reinforcement. Taken together, our results demonstrate the existence of a neural transcriptional code that represents the encoding of experiences in the mouse brain. This code is comprised of distinct transcriptional signatures that correlate to the affective attributes of the experiences that are being encoded.


Introduction:
Neuronal plasticity enables cognitive and behavioral flexibility underlying the development of adaptive behaviors 1,2 . This neuroplasticity, induced by salient experiences, has been shown to depend on the induction of temporally-defined waves of transcription 1-5 . The earliest of these waves consists of the expression of immediate-early genes (IEGs). IEGs have been conventionally treated as molecular markers for labelling neuronal populations that undergo plastic changes, underlying the formation of long-term memory 6,7 . However, recent literature indicates a much more significant contribution of IEGs in synaptic plasticity and memory formation 8,9 . It has been proposed that IEG transcription could represent the molecular signatures of long-term plastic changes underlying the formation of memory 1 . This implies that IEG expression induced by an experience could represent a neural transcriptional code for long-term storage of information that is specific to an experience. The existence of a neural code embedded in transcription implies that it should be possible to decode the identity of recent experiences, and potentially derive information regarding the nature of the experience, from its transcriptional representation 10 . This proposition forms the basis of our investigation.
To address the existence of a neural transcriptional code, we performed detailed analysis of IEG transcription for 14 different experiences (see methods for details), induced by cocaine (acute, repeated and challenge), volitional sucrose drinking (acute and repeated), reinstatement of feeding following food deprivation, lithium chloride administration (LiCl; acute and repeated), saline (acute injection without habituation, acute injection after habituation and following repeated administrations), acute administration of a mild foot shock, and exposure to a novel chamber with no foot shock.
The experiences analyzed were selected to enable identification of the transcriptional representations of affective attributes characterizing an experience, such as valence and salience 11,12 . Analysis of repeated exposure to experiences also enabled identification of common attributes in the transcriptional representation of habituation and positive reinforcement. 3 Assuming that encoding of complex behaviors involves the coordinated transcriptional activation of multiple brain regions, we analyzed transcription across multiple structures of the reward circuitry 13 . The brain structures that were analyzed include limbic cortex (LCtx; including medial prefrontal cortex and anterior cingulate cortex), nucleus accumbens (NAc), dorsal striatum (DS), amygdala (Amy), lateral hypothalamus (LH), dorsal hippocampus (Hipp) and ventral tegmental area (VTA).
Our results demonstrate that the transcriptional representations of each experience are robust, reliable and consistent, enabling the decoding of the recent salient experience of mice with high levels of accuracy from a minimal transcriptional signature. We identify transcriptional correlates for affective attributes of experience, prominently demonstrated in the encoding of valence. Moreover, we report opposing patterns of transcriptional modulation underlying the development of habituation to experiences of neutral or negative valence, in comparison to reinforcement for repeated rewarding experiences. We finally discuss the potential implications of the identification of a neural transcriptional code for salient experiences.

Unique transcriptional signatures represent the history of cocaine experience
Drugs of abuse are known to hijack endogenous mechanisms of neural plasticity, inducing long-lasting modifications of neural circuits in the mesolimbic dopamine system 14 through transcription-dependent mechanisms 15 . Cocaine sensitization is one of the most widely applied paradigms for studying mechanisms of neural plasticity, due to the robustness of the behavioral model and the detailed insight provided into the underlying mechanisms 3,[15][16][17][18][19][20] . We initiated our study with an investigation of the gene expression programs induced during the development of behavioral sensitization to cocaine.
Using the cocaine sensitization paradigm, we studied the transcriptional programs induced in mice following acute exposure to cocaine (intraperitoneal (i.p.) injections, 20 mg/kg), or repeated exposure to the drug, characterized by robust locomotor 4 sensitization. The response to re-exposure to cocaine after a period of abstinence from repeated drug exposures ('cocaine challenge'), is a classic metric of the potent experience-dependent maladaptive plasticity induced by cocaine exposure 15 (Figure   1A,B). We analyzed transcriptional dynamics at 0, 1, 2, 4 hrs following each of these cocaine experiences across 7 brain structures (LCtx, NAc, DS, Amy, LH, Hipp and VTA, applying a comprehensive set of qPCR probes developed against putative IEGs (Table   S1). We observed that the transcriptional representation of distinct cocaine experiences (acute, repeated, challenge) was characterized by the robust induction of a handful of genes across brain structures ( Figure 1C, Figure 1 -Figure Supplement 1). A subset of these genes (Arc, Egr2, Egr4, Fos and Fosb) displayed consistently high induction and low variance, as well as clear temporal dynamics of induction, with expression peaking at 1 hour following cocaine experience (Figure 1 -Figure Supplement 2). Importantly, these five genes were robustly induced across all salient experiences investigated in this work, and were therefore chosen as the representative markers of the recent experience of individual mice for the rest of the study.
As a preliminary test of the hypothesis that experiences can be decoded from patterns of induced transcription, we performed supervised classification of animals that experienced distinct saline and cocaine experiences. Mice were classified based on induction of the five most prominently induced genes (Arc, Egr2, Egr4, Fos and Fosb) across three brain structures (LCtx, NAc and DS). Each gene-structure combination was identified as a "feature" and classification was performed using the k-Nearest Neighbor classifier. Linear projection following dimension reduction of these 15 features into three principal components, resulted in unique clustering of the acute, repeated and challenge experiences of cocaine ( Figure 1D). More importantly, the information contained in these 15 features, allowed precise allocation/classification of individual animals based on the identity of the recent cocaine experiences ( Figure 1E). However, while saline experiences could easily be segregated from cocaine experiences, using this approach and the given set of features, we could only partially segregate between the neutral experiences induced by exposure to acute or repeated saline. We attribute this inability to accurately decode the saline experiences to the effect of habituation of the mice to a 5 neutral stimulus, rendering the experience of repeated saline injections non-salient and therefore not represented by inducible transcription (Figure 1 -Figure Supplement 3, 4).
Taken together, these results suggest that transcriptional signatures comprising of a minimal subset of marker genes reliably represent the comprehensive gene expression programs induced by experience, and suffice to decode the recent salient experience at the resolution of individual mice.

Distinct experiences are represented by unique transcriptional signatures.
To test this fundamental concept, we expanded our study, including experiences that were characterized by negative valence such as LiCl and foot shock, along with naturalistic volitional experiences of positive valence -sucrose consumption, and reinstatement of feeding following food withdrawal. We represent the experience-specific IEG expression patterns induced 1 hour following an experience across multiple brain nuclei using radar plots (Figure 2). This representation provides a birds-eye view of the transcriptional landscape, and enables immediate identification of the major brain regions and transcripts recruited by each experience. Four genes (Arc, Egr2, Egr4 and Fos) were shown for simplicity of presentation; for the data from individual mice, see This presentation further highlights the unique nature of the transcriptional code characterizing each experience, and the dynamic changes in IEG induction following repeated exposures to an experience. Furthermore, commonalities in the transcriptional representation of experiences with shared affective attributes were also visually accessible in this presentation.

The transcriptional representation of negative valence
Investigating the transcriptional representation of negative valence, we focused on the aversive experiences induced either pharmacologically by LiCl administration, or by acute administration of mild foot shock. While animals that received LiCl display malaise, nausea and reduced locomotion 21 ; administration of mild foot shock induces 6 acute pain, fear and consequent freezing 22 . To facilitate a direct comparison of the transcriptional patterns induced by experiences of opposing valence, LiCl administration was performed using the same experimental setup as cocaine sensitization. The experiences of cocaine and acute LiCl exposure drove robust induction of a small but similar subset of IEGs, which most prominently included Arc, Egr2, Egr4, Fos (Figure 2 - Figure Supplement 2). However, in comparison to the cocaine experiences, where transcriptional induction was observed to be pronounced in the LCtx, NAc, DS, and VTA, the LiCl experiences displayed induction predominantly in the Amy (Figure 2).
We next compared the experience of LiCl shared with a different aversive experience, induced by foot shock. It is worth noting that while LiCl and foot shock are both characterized by negative valence, they are otherwise distinct in multiple aspects -LiCl causes prolonged visceral discomfort, while foot shock associates acute peripheral pain with a novel environment (a metal grid located in a small enclosure). Interestingly, exposure to the context in which the foot shock was performed, alone, induced robust IEG transcription in a number of limbic structures, (Figure 2). Mice that received a foot shock within this context displayed an overall indistinguishable pattern of transcriptional induction compared to their no-shock controls, with the sole distinction being a robust induction of transcription (predominantly of Egr2 and Egr4) in the Amy ( Taken together, these results support the notion that recent salient experiences are characterized by robust induction of specialized transcription programs in relevant brain structures. From a broader perspective, our results demonstrate that experiences of opposing valence induce distinct IEG expression patterns in different brain structures

Transcriptional representation of habituation and reinforcement
Repeated experience can either cause habituation or reinforcement, depending on the valence of the experience -habituation is observed to experiences of neutral and negative valence, while reinforcement is expressed in response to repeated experiences of positive valence. We therefore expected to observe a reflection of these distinct behavioral shuffling of the association of mice to experiences demonstrated the reliability of the classifier, and the potential for our results to generalize beyond the given dataset ( Figure   3D). An intuitive representation of the differentiation of experiences based on particular features is provided by a decision tree (one of a number of possible trees), in which mice are manually assigned to appropriate branches according to the level of induction of a particular gene in a given structure ( Figure 3E).
Taken together, these results establish that a minimal set of transcriptional markers form representative signatures of the recent experience of individual mice, and enable precise decoding of recent salient experiences at the resolution of individual mice. 9

Discussion
The brain creates representations of the world, encoding salient information for long-term storage to support the development of adaptive behaviors. In real time, the representation of information has been shown to be correlated with neural activity in distinct brain structures 25 . Powerful demonstrations of the potential to decode sensory experiences and correlates of emotional state have been made in both rodents and humans from neural activation patterns using in-vivo electrophysiology, fMRI, and other imaging techniques [26][27][28][29][30] . In this study, we demonstrate that IEG expression data from multiple regions of the mouse brain enables the decoding of recent salient experiences with high precision. We report that beyond mere 'activity markers' for labelling neurons activated during an experience, IEG expression provides a quantitative and scalable metric, representing a neural transcriptional code of recent experience. Interestingly, this neural transcriptional code is comprised of distinct transcriptional signatures that correlate to the affective attributes of the experiences that are being encoded, such as salience and valence. Moreover, these IEG expression patterns are modulated following repeated administration of a stimulus of a given value (positive, negative or neutral), suggesting a role for inducible transcription in sustaining long-term adaptations underlying the development of adaptive behavior. As this code is comprised of molecular components, it also provides a rich resource for biological insight into the processes underlying the longterm encoding of experience-dependent plasticity.
Transcriptional markers have been successfully utilized for the classification of developmental stages 31 , diseases 32,33 , and many other aspects of contemporary biomedical science 34 . Here we describe the utility of transcriptional markers for classification of salient experiences characterized by diverse affective properties. While the information embedded in the expression pattern of a single gene is not sufficient, a minimal subset of transcriptional markers enables the decoding of recent experience with high accuracy.
Importantly, the principles we identify likely generalize to a broad set of experiences.
However, it is also likely that the markers we utilize in our study could be (at least in part) substituted by other markers, depending on the choice of classifier, providing comparable results.
According to the Russell circumplex model 11,12 , affect can be defined in two dimensions -valence and salience. Valence has been suggested to be encoded in the Amy, PFC, NAc and VTA 35 . Our results demonstrate that experiences of negative valence are represented by a distinct transcriptional induction in the Amy. In contrast, experiences with positive valence induce transcription in the LCtx, NAc, DS and VTA.
Moreover, we report that upon repetition, the transcriptional representation within these structures is dynamically modulated, potentially underlying long term adaptations following positive and negative reinforcement. Taken together our results proposes that inducible transcription is a rich resource for the identification of the involvement of specific brain regions in encoding affective properties of an experience, providing biological insight into the molecular processes underlying experience-dependent plasticity.
To explain how changes in transcription could affect future behavior, we introduce the concept of 'predictive transcriptional coding'. Predictive transcriptional coding frames inducible transcription not as a reporter of a recent event, but rather as encoding the valuation of the experience. This experience-dependent plasticity, mediated by transcription, sets the state of the network in the context of a particular experience, priming it for prospective network plasticity, and adjusts the future response of the individual to the occurrence of a similar event. This notion is conceptually similar to the 'reward prediction error' 36 , but is established on prolonged time scales. In this context, transcription also serves as a 'salience filter' -defining whether an experience is significant enough to induce plasticity and be encoded for long-term storage. Our results are consistent with transcription serving as a salience filter, whereas the valuation of a recent experience is encoded by the identity of the neural circuits recruited by the experience and the magnitude of transcription induced within them. A crucial question arising from this concept is: how do neurons or neural networks determine the threshold to commit to induction of transcription? One possibility could be that the threshold for commitment to transcription depends on coincidence of glutamatergic and neuromodulatory inputs. It should be noted that our study revitalizes concepts and questions, which were raised previously in a landmark treatise defining inducible transcription as a 'genomic action potential' for encoding experience 37 .
Our work provides a numerical definition of the imprint of recent experience, demonstrating a quantitative and predictive approach for analysis of neuronal plasticity underlying adaptive behavior. Quantitative definitions of interoceptive states are expected to have implications for drug development -providing objective metrics for comprehensive characterization of the individual perception and valuation ascribed to an experience by individual subjects. For example, in the context of abuse liability, an objective quantitative interoceptive metric of the hedonic potential of a compound could increase standardization, reducing the reliance on variable behavioral outcomes.
While there is substantial investment being made in the development of methodologies for transcriptional profiling with deeper coverage and increasing spatial resolution, our study demonstrates that fundamental phenomena can be identified by applying simple methods with low spatial resolution and coverage. Future work, applying tools of higher resolution, could build on our observations to address additional questions -such as the spatial distribution of neuronal ensembles recruited by experience and the identity of cell types recruited by distinct experiences.
Approaches for non-invasive quantitative measurement of the encoding of experience can be envisioned, utilizing fluorescent markers of inducible transcription in combination with whole-brain imaging 38 . New technologies are rapidly emerging for whole-brain analyses of transcription [39][40][41] , as are strategies for comprehensive profiling of single neurons 42,43 . These technological developments, together with the novel concept we develop here, are expected to provide the foundation for a new area of neuroscience research. This discipline, of "Behavioral Transcriptomics", will apply transcriptional analysis for investigation of intricate mechanisms of neural circuit plasticity underlying cognition. We propose that the approach of behavioral transcriptomics will provide a systems-level view of the encoding of experiences to long-term memory. Potentially, different elements of an experience may be mediated by activation of defined receptor sub-types (e.g. NMDA, dopamine, serotonin, etc.') at specific dendritic locations, each inducing a component of the transcriptional program. If so, taken to its extreme, deciphering this transcriptional code will enable precise decoding of synapse-specific plasticity from quantitative analysis of inducible transcriptional markers. Behavioral Assays. Mice were acclimated to the animal facility for at least 2-5 days, followed by 3-4 days of experimenter handling, before the start of an experiment.
Maintenance of uniform conditions across experiments and extensive handling were essential for reducing experimental variability, enabling the identification of a robust transcriptional response specifically induced by the experience being tested and minimal contamination from contextual background. Behavioral sensitization to cocaine. Mice were subjected to three days of intraperitoneal (i.p.) saline injections (250 microliter/injection), prior to exposure to cocaine (20 mg/kg freshly dissolved in physiological saline to 2 mg/ml and injected at a volume of 10 ml/kg; cocaine was obtained from the pharmacy at Hadassah Hospital, Jerusalem). The acute cocaine group received a single i.p. dose of cocaine, followed by analysis of locomotor behavior for 15 13 minutes in a video-monitored open-field arena. Animals were finally taken from their home cage and sacrificed at 1, 2 and 4 hrs following the cocaine injection. The repeated cocaine group received five consecutive daily injections of cocaine, and were studied (similar to the acute cocaine group) following the fifth cocaine injection. The challenge cocaine group were treated as the repeated cocaine group, and then made abstinent from cocaine for 21-22 days, following which they were challenged with cocaine and reexposed to the open-field arena. All responses were normalized to baseline controls (time 0), which were interleaved with their peer group, but were not treated on the day of the experiment. Additional reference groups included acute saline without habituation, which were habituated to the open-field arena for three days after a brief period of handling, and were sacrificed 1 hr following a single injection of saline. Responses in this group were normalized to controls (time 0), which were not exposed to any saline were re-exposed to LiCl or saline. Mice were divided into four groups: a) Received saline injections for five days and were not exposed to an injection on the last day (saline-0h), b) Received LiCl injections for five days and were not exposed to an injection on the last day (LiCl-0h), c) Received saline injections for five days and were subjected to saline injection on the last day (repeated saline), d) Received LiCl injections for five days and were exposed to LiCl injection on the last day (repeated LiCl). In all experiments, immediately following administration of LiCl or saline, mice were placed in video-14 monitored open-field arenas for 30 min. Reinstatement of feeding. Mice were food deprived for 18 hrs before the experiment and then re-exposed to food for 1, 2 or 4 hours before they were sacrificed. Control animals (0 hr) were sacrificed immediately after the 18 hr food restriction. An additional reference group was allowed to continuously feed.
Sucrose Consumption. Mice were single-housed for at least seven days before the experiment and habituated to the addition of a second water bottle in the cage for three days before the onset of the experiment. Acute exposure to sucrose was tested by habituating mice to the bottle with 10% sucrose overnight (16 hr), and 48 hrs later, reexposing the mice to a bottle with sucrose or water (control) for 1 hr. Repeated exposure to sucrose was tested by exposing mice to sucrose repeatedly for eight consecutive days, 2 hr each day (12:00-14:00), and after a 48 hr break, re-exposed to sucrose or water (control) for 1 hr. Mice were sacrificed 1 hr following the exposure to sucrose. Sucrose and water intake were measured as a test for sucrose preference over water. Foot Shock.  Schematic of experimental paradigm. Mice were exposed to cocaine (i.p., 20 mg/kg) or saline, either acutely, repeatedly or re-exposed after abstinence (challenge), with transcriptional dynamics studied at 0, 1, 2 or 4 hrs.                Schematic of experimental paradigm. Mice were exposed to cocaine (i.p., 20 mg/kg) or saline, either acutely, repeatedly or re-exposed after abstinence (challenge), with transcriptional dynamics studied at 0, 1, 2 or 4 hrs. (B) Locomotor activity of mice following acute, repeated or challenge cocaine experiences (compared to saline). Sample size: acute saline n = 6; acute cocaine n = 30; repeated saline n = 4; repeated cocaine n = 22; challenge cocaine n = 19 mice. Results indicate mean ± s.e.m.
(C) Expression matrix of IEG induction dynamics following cocaine experiences. Individual animals are represented in columns sorted according to time points of cocaine experiences [sample numbers per time point -LCtx: limbic cortex (n = 5-11), NAc: nucleus accumbens (n = 5-12), DS: dorsal striatum (n = 5-12), Amy: amygdala (n = 3-4), LH: lateral hypothalamus (n = 2-4), Hipp: hippocampus (n = 2-4)]. Fold induction is graded from blue (low) to red (high). Genes represented were induced at least 1.5-fold over control in any one of the brain regions studied. (D) Linear projection of the three principal components after dimension reduction of transcriptional induction of Arc, Egr2, Egr4, Fos and Fosb in the LCtx, NAc and DS segregates cocaine-treated mice into discrete clusters associated with recent experiences. Each dot represents an individual animal, color-coded according to the identity of its recent experience.
(E) Confusion matrix representing the classification accuracy of decoding the recent experience (acute / chronic / challenge cocaine vs saline) of individual mice based on expression of Arc, Egr2, Egr4, Fos and Fosb induction in the LCtx, NAc and DS. Accuracy is scaled from blue to green, with bright green corresponding to 100% accuracy (n = 29 mice). foot shock (acute shock and no-shock controls exposed to the same environment); LiCl (acute and repeated); cocaine (acute, repeated and challenge following abstinence); sucrose (acute and repeated) and reinstatement of feeding (following 18 hrs of deprivation).  Selection of features (expression of IEG in a brain structure) with the highest potential to contribute to the classification of recent experiences ('support'). Support was defined by running 1e 6 classifiers with varying composition of features, and calculating the mean accuracy for each feature over all the iterations to which it contributed, using the Random k-Nearest Neighbors (RKNN) algorithm. The grey zone defines the eight features with the highest support values (Egr2 and Fos in the LCtx, Egr2 and Fos in the NAc, Egr2 and Fosb in the DS, Egr2 and Fos induction in the Amy). (B) Accuracy of classification obtained with an increase in the number of features. For each # (subset) of features, the features ranking highest in support were chosen and a KNN classifier was evaluated using only those features. Peak accuracy was achieved with eight features (93.6%). (C) Confusion matrix representing the classification accuracy (93.6%) of decoding the recent experience of individual mice based on the eight most informative features. Efficiency is scaled from blue to green, with bright green corresponding to 100% efficiency (n = 53 mice). (D) Verification of classification validity. A randomization test was performed, in which the classifier was run on 10 6 random permutations of the association of individual mice to the appropriate experience, and the frequency of classification accuracies is plotted in grey, while the red dotted line represents the classification accuracy obtained for non-randomized data (93.6%). (E) A supervised decision tree enabling the classification of mice according to experience by minimal gene expression (many trees can equivalently segregate the data). Mice are classified manually based on features that enable segregation at each internal node. The decision tree failed in segregating three mice that experienced acute sucrose from the group of mice that experienced repeated sucrose. Mice are color-coded according to experience.    Expression matrix of the transcriptional induction of IEGs following acute or repeated saline experiences. Each column represents the transcriptional profile of an individual mouse. Sample numbers for each time point for acute saline: n = 10-12 in LCtx (limbic cortex); n = 12-13 in NAc (nucleus accumbens) and DS (dorsal striatum); n = 5 in Amy (amygdala), LH (lateral hypothalamus), Hipp (hippocampus). Sample numbers for each time point for repeated saline: n = 3-5 in LCtx, NAc, DS, Amy, LH and Hipp). Transcriptional induction is graded from blue (low) to red (high). Genes represented were induced on average at least 1.5-fold following either cocaine or LiCl experiences in any one of the studied brain nuclei. (C) Average transcriptional induction of 78 genes 1h following acute or repeated saline experience in the LCtx, NAc, DS, Amy, LH and Hipp. Genes are sorted alphabetically. Sample sizes: LCtx: n = 10 for acute saline and n = 5 for repeated saline; NAc: n = 12 for acute saline and n = 5 for repeated saline; DS: n = 12 for acute saline and n = 5 for repeated saline; Amy: n = 5 for acute saline and n = 5 for repeated saline; LH: n = 5 for acute saline and n = 5 for repeated saline; Hipp: n = 6 for acute saline and n = 5 for repeated saline. Results indicate mean ± s.e.m.