Inference of neuronal functional circuitry with spike-triggered non-negative matrix factorization

Jian K. Liu, Helene M. Schreyer, Arno Onken, Fernando Rozenblit, Mohammad H. Khani, Vidhyasankar Krishnamoorthy, Stefano Panzeri, Tim GollischNature Communications 2017 · DOI

Abstract

Neurons in sensory systems often pool inputs over arrays of presynaptic cells, giving rise to functional subunits inside a neuron's receptive field. The organization of these subunits provides a signature of the neuron's presynaptic functional connectivity and determines how the neuron integrates sensory stimuli. Here we introduce the method of spike-triggered non-negative matrix factorization (STNMF) for detecting the layout of subunits within a neuron's receptive field. The method only requires the neuron's spiking responses under finely structured sensory stimulation and is therefore applicable to large populations of simultaneously recorded neurons. Applied to recordings from ganglion cells in the salamander retina, the method retrieves the receptive fields of presynaptic bipolar cells, as verified by simultaneous bipolar and ganglion cell recordings. The identified subunit layouts allow improved predictions of ganglion cell responses to natural stimuli and reveal shared bipolar cell input into distinct types of ganglion cells.

Introduction

In sensory systems, output neurons like Retinal Ganglion Cells (RGCs) act as integration hubs, pooling signals from an array of input neurons, such as bipolar cells. These inputs form functional subunits within the RGC's receptive field. Understanding this organization is crucial for deciphering how neurons process sensory information. However, these subunits are often invisible to standard analysis. The classic Spike-Triggered Average (STA) provides only a blurry, aggregate view of a neuron's sensitivity, failing to capture the fine-grained, nonlinear "pixels" that constitute its receptive field. While methods like Spike-Triggered Covariance (STC) exist, they often yield abstract features that lack clear biological correspondence.

The authors developed Spike-Triggered Non-negative Matrix Factorization (STNMF) to identify the physical layout and functional contribution of these subunits without prior assumptions about their shape or number.

Methodology: Spike-Triggered Non-negative Matrix Factorization

STNMF applies Non-negative Matrix Factorization (NMF) to neural spiking data through a multi-step process:

  1. Spike-Triggered Ensemble: Stimulus patterns preceding each spike are collected.
  2. Dimensionality Reduction: These patterns are collapsed into "effective" spatial images based on the neuron's temporal filter.
  3. Semi-NMF Decomposition: The ensemble is decomposed into a set of modules (subunits) and weights.
Schematic of the STNMF method, showing the decomposition of the effective spike-triggered stimulus ensemble into modules and weights.

A key innovation is the non-negativity constraint on the modules. Since biological subunits typically provide localized, uni-directional signals, this constraint forces the algorithm to find "parts-based" representations, effectively discovering the localized shapes of presynaptic cells. The algorithm uses iterative numerical procedures with random perturbations to avoid local minima and employs Moran’s I (spatial autocorrelation) to automatically distinguish structured subunits from noise.

Results and Validation

The method was tested on isolated salamander and mouse retinas recorded via multielectrode arrays (MEA). To validate the findings, the researchers used a "gold standard" dataset of simultaneous intracellular recordings from bipolar cells and extracellular recordings from RGCs.

Comparison between the standard Spike-Triggered Average (STA) and the modules (subunits) identified by STNMF. The subunits correspond to the receptive fields of presynaptic bipolar cells.

Key findings include:

  • Biological Accuracy: STNMF subunits showed high spatial overlap (up to 0.86) with the receptive fields of real, physically recorded bipolar cells.
  • Superior Prediction: Models using STNMF subunits significantly outperformed standard Linear-Nonlinear (LN) models in predicting RGC responses to natural images and "frozen" white noise.
  • Shared Circuitry: The study revealed that different types of ganglion cells (e.g., "Fast Off" and "Direction-Selective" cells) often share the same bipolar cell subunits, suggesting a highly interconnected mesh rather than independent processing pipelines.

Discussion

While robust, STNMF requires several thousand spikes, which may be challenging for neurons with low firing rates. The current implementation also simplifies the temporal dimension to focus on spatial layout. Future work could extend this to a full spatiotemporal STNMF and explore more complex subunit nonlinearities. Beyond the retina, this method holds promise for mapping functional subunits in other sensory areas, such as the primary visual cortex or the auditory system.

Summary updated

Abstract

Neurons in sensory systems often pool inputs over arrays of presynaptic cells, giving rise to functional subunits inside a neuron's receptive field. The organization of these subunits provides a signature of the neuron's presynaptic functional connectivity and determines how the neuron integrates sensory stimuli. Here we introduce the method of spike-triggered non-negative matrix factorization for detecting the layout of subunits within a neuron's receptive field. The method only requires the neuron's spiking responses under finely structured sensory stimulation and is therefore applicable to large populations of simultaneously recorded neurons. Applied to recordings from ganglion cells in the salamander retina, the method retrieves the receptive fields of presynaptic bipolar cells, as verified by simultaneous bipolar and ganglion cell recordings. The identified subunit layouts allow improved predictions of ganglion cell responses to natural stimuli and reveal shared bipolar cell input into distinct types of ganglion cells.How a neuron integrates sensory information requires knowledge about its functional presynaptic connections. Here the authors report a new method using non-negative matrix factorization to identify the layout of presynaptic bipolar cell inputs onto retinal ganglion cells and predict their responses to natural stimuli.
BibTeX
@inproceedings{liu2017inference,
  author = {Jian K. Liu and Helene M. Schreyer and Arno Onken and Fernando Rozenblit and Mohammad H. Khani and Vidhyasankar Krishnamoorthy and Stefano Panzeri and Tim Gollisch},
  title = {{Inference of neuronal functional circuitry with spike-triggered non-negative matrix factorization}},
  booktitle = {Nature Communications},
  year = {2017},
  doi = {10.1038/s41467-017-00156-9},
}
Figure