Early selection of task-relevant features through population gating

Article https://doi.org/10.1038/s41467-023-42519-5 Early selection of task-relevant features through population gating Received: 24 October 2022 Accepted: 12 October 2023 1234567890():,; 1234567890():,; Check for updates Joao Barbosa Srdjan Ostojic 1 , Rémi Proville2, Chris C. Rodgers 1 & Yves Boubenec 5 3 , Michael R. DeWeese4, Brains can gracefully weed out irrelevant stimuli to guide behavior. This feat is believed to rely on a progressive selection of task-relevant stimuli across the cortical hierarchy, but the specific across-area interactions enabling stimulus selection are still unclear. Here, we propose that population gating, occurring within primary auditory cortex (A1) but controlled by top-down inputs from prelimbic region of medial prefrontal cortex (mPFC), can support across-area stimulus selection. Examining single-unit activity recorded while rats performed an auditory context-dependent task, we found that A1 encoded relevant and irrelevant stimuli along a common dimension of its neural space. Yet, the relevant stimulus encoding was enhanced along an extra dimension. In turn, mPFC encoded only the stimulus relevant to the ongoing context. To identify candidate mechanisms for stimulus selection within A1, we reverseengineered low-rank RNNs trained on a similar task. Our analyses predicted that two context-modulated neural populations gated their preferred stimulus in opposite contexts, which we confirmed in further analyses of A1. Finally, we show in a two-region RNN how population gating within A1 could be controlled by top-down inputs from PFC, enabling flexible across-area communication despite fixed inter-areal connectivity. The informational value of different stimuli can change dramatically depending on the context, but animals can adapt with impressive flexibility to virtually any contingency change. A classical example of this feat is the so-called “cocktail party effect”, which refers to our ability to focus on a specific, currently relevant conversation while ignoring all the others. Understanding how stable neural circuits implement this kind of flexible, context-dependent behavior has proven challenging. While there is a growing consensus that it emerges from the interaction between different regions along the brain hierarchy1–4, the specific interactions are unclear. One possibility is that regions early in the hierarchy merely represent the incoming stimuli and propagate their representations downstream, where context-dependent rules are applied to effectively guide behavior5–8. In line with this view, pioneering work combining artificial neural networks and neurophysiological recordings from monkeys performing a canonical context-dependent task9, shows that both relevant and irrelevant stimuli are encoded as late as the frontal cortex, suggesting that the selection of relevant stimuli indeed may occur late in the cortical hierarchy. Empirical evidence demonstrates however that primary sensory areas are modulated by behavioral context4,10–13, potentially through feedback interactions with downstream areas that could control the selection of the relevant stimulus upstream14,15. This evidence supports early models of parallel distributed processing16, that proposed that task-relevant stimuli encoding could be enhanced by top-down inputs to sensory neurons. The prefrontal cortex17 is deemed essential in providing these inputs, which push task-irrelevant units to 1 Laboratoire de Neurosciences Cognitives et Computationnelles, INSERM U960, Ecole Normale Superieure - PSL Research University, 75005 Paris, France. Tailored Data Solutions, 192 Cours Gambetta, 84300 Cavaillon, France. 3Department of Neurosurgery, Emory University, Atlanta, GA 30033, USA. 4 Department of Physics, Helen Wills Neuroscience Institute, and Redwood Center for Theoretical Neuroscience, University of California, Berkeley, CA, USA. 5 Laboratoire des Systèmes Perceptifs, Département d’Études Cognitives, École Normale Supérieure PSL Research University, CNRS, Paris, France. e-mail: 2 Nature Communications | (2023)14:6837 1 Article the low-gain region of their dynamic range, thereby effectively reducing their sensitivity. While an attractive possibility, the specific mechanisms through which different cortical areas cooperate to select the relevant stimuli earlier in the cortex are unclear. Here, we examine the population dynamics in the rat primary auditory cortex (A1) and the prelimbic region of medial prefrontal cortex (PFC), and propose a mechanism through which interactions between these two areas flexibly select relevant stimuli within A1 in a context-dependent task13. We found that both relevant and irrelevant stimuli were encoded within a sensory subspace of A1, in line with other studies of humans and other animals performing contextdependent tasks2,4,13. However, we found that the relevant stimuli were furthermore projected along an additional dimension, which we named ‘selection axis’. On the other hand, PFC encoded only the decision, fully determined by the selected stimuli. Both areas encoded context robustly throughout the trial. To investigate how this contextual information could drive stimulus selection in A1, we trained recurrent neural networks (RNN) on a similar task. Using the same analyses, we found that the geometry of the relevant and irrelevant stimuli representations resembled those of the rat’s A1. Reverseengineering the mechanisms employed by these networks18–20 predicted that context-modulated populations selectively gate the relevant stimuli in a context-dependent fashion, with different populations selecting specific stimuli in their preferred context. Further analyses of neural recordings revealed a similar population structure in A1, validating the model prediction and suggesting it could subserve the flexible communication of the selected stimulus with mPFC. A possible interpretation of our within-area modeling and data analyses is that context-dependent gain modulation occurring within A1 could be controlled by top-down inputs from PFC16,17. A recent hypothesis posits that different regions communicate through lowdimensional subspaces21–23, but how the information being communicated could alternate flexibly to solve a context-dependent task is unclear. Our final contribution is to show through network modeling that within-area gain modulation19, controlled by across-area inputs, could sub-serve such flexible communication along low-dimensional subspaces. Specifically, we demonstrate that a previously proposed class of RNNs constrained to have within-area low-dimensional dynamics18–20 can be naturally extended to account for across-area communication subspaces. In a two-region RNN, we show that relevant stimuli information can be transmitted between A1 and PFC in a context-dependent manner, despite fixed inter-area connectivity. Our model is a neural implementation of the communication subspace hypothesis22,23 that solves a cognitive task and (...truncated)