Recurrent inhibition within olfactory networks shapes noise correlations and stimulus discrimination

Wang, Angelia Feiyang

doi:https://doi.org/10.21985/n2-b6pk-c215

Work

Recurrent inhibition within olfactory networks shapes noise correlations and stimulus discrimination

Public

Download PDF

Download All Files (.zip)

Each second, living organisms take in sensory input from an ever-changing environment and respond appropriately. Identifying and contextualizing stimuli is critical for survival, and it often necessitates distinguishing between sensory experiences that are similar to each other. Pattern separation characterizes the mechanisms by which neuronal networks extract and highlight differences between similar stimulus-induced activity patterns, making it easier for higher-order brain centers to distinguish between them. Within the olfactory system, pattern separation is made possible by its extensive circuitry through which neurons communicate and interpret each other’s activities. Previous work discusses how specific shared connectivity can induce pairwise correlations between neurons and how correlations reduce the total information a network can carry, but the roles of specific recurrent connectivities and mechanistic interactions between cells within the bulb are not well understood. In this thesis, we investigate how different motifs of recurrent connectivity shape noise correlations, how variances and pairwise covariances are impacted by stimulus shape and measurement duration, and how all these quantities impact performance in a two-class stimulus discrimination problem. The olfactory bulb contains recurrently-connected mitral cells (MCs) and granule cells (GCs), whose excitatory-inhibitory interactions generate network oscillations in the gamma frequency range. In Chapter 2, we find that recurrent inhibition from the GCs induces a gamma rhythm characterized by alternating volleys of MC and GC spikes and that MC fluctuate on the scale of the gamma period. On short timescales, MCs fire synchronously (i.e. within the same volley of spikes); over longer timescales, they reduce each other’s activity via shared inhibition. Correspondingly, pairwise covariance between two MCs is positive for short measurement durations and negative in measurements larger than the period of the gamma rhythm. In Chapter 3, we quantify network performance using linear discriminability, which is governed by the average difference in response (∆μ) between two stimuli and the sum of response covariances (Σ). Linear discriminability measures the dissimilarity between two stimulus-evoked patterns and may represent a cortical neuron’s readout of MC activity. We find that inhibition worsens stimulus discrimination in a network comprising independent pairs of MC-GC reciprocal connections (“single connections”), but that this reduction is largely dominated by average difference in response ∆μ. Inhibition also reduces variance in these single connected networks, which benefits discriminability and partially offsets the effect of ∆μ. Conversely, in a network with all-to-all coupling (“global connections”), small amounts of inhibition improve discriminability despite reducing ∆μ, and noise correlations in Σ become increasingly beneficial as the stimuli become more similar. We assess both optimal and exploratory linear discriminability, which constitute different ways a network can perform discrimination. Optimal discriminability Fopt represents the best-case-scenario performance of a network that optimally weights each MC input in order to maximize stimulus separability, and random (exploratory) discriminability Frandom represents the performance of a network that has not learned how to discriminate an odor pair and weights MC inputs randomly. Networks connected with either single connections or global connections perform similarly in both metrics – inhibition delivered through single connections worsens discriminability and small inhibition delivered through global connections improves discriminability. However, a key difference between the two metrics is that Frandom accounts for noise in dimensions (MC inputs) which have the same average activity across both stimuli. As a result, preferentially inhibiting these distracting cells reduces the total noise in the system, which improves Frandom. In Chapter 4, motivated by the idea that cortical responses may result from a few spikes soon after stimulus onset, we implement a sniff cycle and take measurement windows sequentially along the sniff. We find that discriminability in the steady-state regime matches discriminability over the later portion of a sniff, but that there may also be substantial improvements during early inhalation. Specifically inhibition delivered through either single-connected and all-to-all networks improves both Fopt and Frandom, but only at the specific times when MCs with the largest differences between stimuli are active, whereas inhibition that preferentially suppresses strongly-spiking uninformative cells produces strong sustained improvement in discriminability over the entire inhalation. We also find that inhibition that reduces overall discriminability in the steady-state may produce alternating periods of improved and worsened discriminability, even if the overall time-averaged discriminability is reduced. Chapters 5 and 6 include model implementation and discussion of this thesis work, respectively.In Chapter 7, I describe my CAR-T modeling project completed as a Clinical Pharmacology summer intern at Takeda Pharmaceuticals. While not directly related to the primary thesis work, this research nevertheless showcases an example of using mathematical modeling to further understanding of biological mechanisms. In this project, I selected a case study of CD19-targeting CAR-T used to treat both immunodeficient mice and a cohort of clinical cancer patients, and I constructed a physiology-based mechanistic model that simultaneously captured data from both cohorts. In doing so, the model predicted key mechanistic differences in how preclinical (non-human) species and clinical patients respond to this specific CAR-T construct. Specifically, T-cells expand approximately four times more rapidly in clinical patients compared to immunodeficient mice, and it also has approximately five-fold higher tumor-killing efficiency after infusion. This work may inform clinical dosing strategies in the future and provides a framework for other modeling efforts in translational medicine. The appendix includes more detailed analyses of some of the methods we use, as well as various exploratory analyses that informed model development.

Creator