PREreview of Context rescales a social action code in a hormone-sensitive network

The nature of a social interaction is determined by both the location that it takes place and the identity of the social partner. How an animal behaves in different types of social interactions can be further shaped by its hormonal state. In this technical tour de force, Guthman et al., gives comprehensive insights into the behavior phenotypes and underlying neural code across distinct social contexts (defined by location and partner), and the role of gonadal hormones in generating appropriate social behaviors. Using unsupervised pose-tracking of mice interacting across 6 social contexts (2 territories*3 social partners), the authors report distinct patterns of behaviors for each social context in both male and female. They then probed the underlying neural activity within the social behavior network by performing simultaneous large-scale calcium recording of hormone sensitive neurons (ERɑ+ and ERɑ-) across 11 brain regions. This rich dataset revealed a distributed neural representation of social context that is linearly rescaled by territorial context as well as complex neural-behavioral mapping in both ERɑ+ and ERɑ- populations. Hormonal manipulation via gonadectomy demonstrates that circulating gonadal hormones are required for such rescaling, and that subsequent hormone replacement is sufficient to rescue both behavioral and neural phenotypes. 

Overall, this is an extensive body of work that advances our current understanding of how social behavior phenotypes are encoded across a network of subcortical regions. However, the richness of the dataset also highlights opportunities for deeper mechanistic dissection of network dynamics across subregions, and of how gonadal hormones interact with ERɑ+/- populations to shape social context encoding. Detailed comments are below. 

Major comments:

1. A central claim of the manuscript is that the territorial context of a social encounter rescales partner-specific action code, and that this rescaling operates over the entire behavioral landscape rather than specific behaviors (e.g. attack). This conclusion is supported by the observation that mean population activity across behavior clusters in home versus partner territories is well fit by a linear model with a positive intercept. However, behavior clusters are not independent of one another and there are non-uniform transition probabilities between different behavior clusters. Some clusters follow one another with higher probability (e.g. proactive→mutual, approach→investigate), while others are rarely adjacent (e.g.proactive → asocial, approach→solo nesting). It is therefore possible that territorial rescaling acts primarily on a selection of behavior clusters, and that the apparent global effect emerges from dependencies and structured transitions between clusters. The authors could address this by quantifying and reporting the transition probabilities or covariance matrix between behavior clusters, and then testing whether territorial rescaling remains robust after accounting for these dependencies.

2. The 6 social contexts are not analogous between male and female subject mice. For example, aggressive and non-aggressive males are unlikely to be aggressive towards females, so these two partners represent distinct social contexts for female subjects compared to males. Similarly, females presented to male subjects may represent qualitatively different social contexts depending on their sexual receptivity. In addition, different strains of male mice are used for agg+/agg- contexts, and because mice can differentiate the strain of a conspecific, this may confound how the conspecific is perceived. To more directly compare behavioral and neural codes across social contexts in males and females, it would be helpful to consider how each type of social partner is perceived by the subject. The authors should consider presenting agg+/agg- conspecifics of the same strain and controlling for the sexual receptivity of female mice presented to male subjects, or demonstrate that their results are invariant to strain and sexual receptivity differences. 

3. The claim that ERɑ+ and ERɑ- populations contribute similarly to social context encoding is currently mainly supported by a lack of a significant difference in the drop of decoder accuracy (F1 loss). Because the F1 loss for the decoders did not significantly differ when either ERɑ+ and ERɑ- populations were shuffled, the authors propose that the two populations contribute similarly to the encoding of most social contexts (except home territory AGG-). While comparable F1 loss could suggest that both populations contribute to decoding, the term ‘similarly’ may be misleading, as it implies a shared underlying encoding implementation that is not directly demonstrated here. In addition, ERɑ+ and ERɑ- populations are likely to be interconnected, raising the possibility that the contribution of one population reflects its coupling to and readout of activity in the other population. The authors could strengthen this conclusion by first quantifying covariance between ERɑ+ and ERɑ- populations and compare F1 loss difference from sites with the least correlated ERɑ+ and ERɑ- population activity. For sites with highly correlated ERɑ+ or ERɑ- populations, the authors may consider reversibly silencing either ERɑ+ or ERɑ- populations (e.g. chemogenetics) and examining how it affects both the decoding accuracy based on the other population and the magnitude of F1 loss. 

4. The authors argue that social contexts are encoded in a distributed fashion across the social behavior network from analyzing how network decoding accuracy suffers from the removal of individual components. Removing individual regions caused a gradual, not modular, loss of accuracy, indicating that the encoding is more distributed across the network rather than concentrated in specific regions.However, many of the 11 regions recorded are highly interconnected. With each region sending and receiving dense projections from one another, there is likely significant redundancy across these nodes. Activity in any single site may reflect not only its own local processing but also network-wide dynamics propagated through recurrent connectivity. This covariance across regions could confound the interpretation that encoding is distributed, since holding out one region’s activity may not fully remove its influence on the network. To more directly test the contribution of individual regions, the authors may consider region-specific perturbations, at least of sites with the largest decoder weights, and assess how it impacts network decoding of social context.

5. While the authors showed that circulating gonadal hormones are necessary and sufficient for gain-control of social-action code, it remains unclear whether testosterone and its estradiol metabolite act directly through the ERɑ+ and ERɑ- neuron populations that were recorded. The proposed significance of gonadal hormones on social-action encoding would be strengthened by further mechanistic experiments demonstrating how these hormones causally influence ERɑ+/ERɑ- activity across different regions. The authors could combine pharmacological perturbation of ERɑ receptors, such as administrating an ERɑ antagonist, with hormonal perturbation to test whether the hormonal modulation of the network encoding depends specifically on ERɑ signaling. 

6. The divergent behavioral and neural effect of gonadectomy in male and female subject mice could be explained by sex-specific conversion of testosterone to estradiol via aromatase. Moreover, the efficacy of testosterone replacement also depends on aromatase activity. These issues can be addressed by controlling or manipulating aromatase activity and by quantifying circulating hormone levels to validate the efficiency of hormone depletion and replacement in both sexes. 

7. ERɑ+ neurons within many of the recorded regions have been shown to be highly heterogeneous. It is well established that different subsets of ERɑ+ neurons are active during different social contexts. Such heterogeneity is masked by the coarseness of fiber photometry recording. It may be worthwhile to spot check one or two regions with well-established  heterogeneity (e.g. VMHvl) with single-cell resolution calcium imaging to confirm that it is consistent with the authors’ observation in bulk population recording. 

Minor comments:

1. To help readers contextualize the analyses for this high-dimensional dataset and establish better intuition for the behavioral and neural codes proposed in this manuscript,  the authors should consider including  behavioral occupancy plots for all 6 social interactions for male and female subjects as well as neural activity heatmaps for all recording sites as supplementary figures. 

2. In figure 3, the colors chosen for “All Shuffled” x “ERɑ- Only Shuffled” and “Male Subjects” x “ERɑ+ Only Shuffled” are a little difficult to distinguish,especially given the small symbol size.Adjusting to more visually distinct color palettes or increasing the marker size would improve readability and clarity. 

Competing interests

The authors declare that they have no competing interests.

Use of Artificial Intelligence (AI)

The authors declare that they did not use generative AI to come up with new ideas for their review.