Filter-based models of suppression in retinal ganglion cells: Comparison and generalization across species and stimuli

Executive summary

• Research question and motivation: How do excitation and suppression interact within single retinal ganglion cells to shape spiking responses and the encoding of dynamic visual stimuli? The study evaluates how three canonical suppression mechanisms influence predictive power beyond a linear-nonlinear (LN) model, aiming to illuminate how suppression contributes to neural coding across species and stimulus contexts.

Methodology in brief

• Models and framework: Implemented four encoders in a common master structure:

• Subtractive suppression (two upstream branches: excitatory and suppressive; combined additively with a subtractive operation)

• Divisive suppression (two branches; suppressive impact multiplies or scales the excitatory input; suppressive nonlinearity is bump-shaped)

• Feedback suppression (suppressive signal driven by spikes via a lagged filter)

• LN baseline for comparison

• Common features: Each branch has a stimulus filter, a monotone or bump-shaped upstream nonlinearity, and a final output nonlinearity feeding a Poisson spike generator. Training used block-coordinate ascent to maximize the likelihood; filters normalized; suppressive nonlinearities constrained to maintain their suppressive interpretation.

• Data and evaluation: Recorded retinal ganglion cells from axolotl (n=607 cells, 15 recordings), mouse (n=335 cells, 8 recordings), and marmoset (n=370 cells, 10 recordings) under full-field white-noise stimuli; in axolotl/mouse, a frozen-noise segment tested generalization. Performance metric = average information per spike (test set); explained variance used when frozen-noise data were available.

Key findings

• Across species, all three suppressive models improved predictions relative to LN:

• Subtractive: better than LN for 97% of cells

• Divisive: better for 96%

• Feedback: better for 85%

• Relative performance across cells:

• Best model varied: subtractive best for 46% of cells, divisive best for 42%, feedback best for 12%.

• Cell-type and kinetics patterns:

• In axolotl, slow OFF cells favored subtractive; divisive and subtractive were more similar in fast cells.

• In mouse and marmoset, divisive often matched or outperformed subtractive; division tended to lag the excitatory input by about 20–70 ms in many cells, providing timely suppression of peak firing.

• Transient vs sustained: In a subset of primate cells, suppression was more beneficial for transient ON cells, suggesting suppression helps flatten or truncate brief response events.

• Generalization to non-white stimuli:

• At high contrast and high temporal frequency, divisive suppression better constrained predicted responses (preventing overshoot), whereas subtractive suppression predicted low-frequency or slow dynamics more faithfully.

• Across cells, divisive models generalized over wider contrast ranges; subtractive excelled at low-frequency components.

• Implications for coding: Subtractive suppression appears well-suited to slow or sustained components; divisive suppression better handles rapid dynamics and contrast adaptation.

Significance and applications

• Demonstrates that multiple, distinct suppression motifs can substantially improve single-neuron encoding models and that different motifs may dominate under different temporal regimes or stimulus statistics.

• Supports the relevance of divisive normalization-like computations as a general mechanism in early sensory processing, with potential extensions to cortical circuits.

• Provides a principled, cross-species comparison framework for dissecting suppression and informs computational neuroscience models of neural coding and stimulus adaptation.

Limitations and future directions

• Limitation: analyses rely on spatially homogeneous (full-field) stimuli, obscuring center-surround and local subunit suppression; biophysical interpretation of center vs. surround contributions remains unresolved.

• Future work: extend to spatially structured stimuli, independently manipulate center and surround, and test with naturalistic movies; relate suppression motifs to specific interneuron circuits; apply similar multi-branch frameworks to cortical data and other sensory modalities.