Neural pathways and computations that achieve stable contrast processing tuned to natural scenes

Visual systems must maintain stable contrast perception despite rapid changes in luminance caused by self-motion, eye/gaze shifts, and object motion. While photoreceptors exhibit slower adaptation, behavioral and physiological evidence across species suggests a rapid post-receptor luminance gain control that preserves contrast computations within hundreds of milliseconds. The authors ask where in the Drosophila visual pathway this rapid luminance gain control arises, what computational principles enable stable contrast under dynamic luminance typical of natural scenes, and what circuit and molecular mechanisms implement it. They propose that normalization using locally pooled luminance enables luminance-invariant contrast processing and set out to identify the neuron types, spatial pooling extent, and inhibitory mechanisms involved.

Prior work indicates rapid luminance gain control in vertebrate pathways (e.g., cat Y-type RGCs under dim light, LGN neurons with stable contrast across fast luminance variations) and human perception. In flies, luminance-invariant visually guided behaviors at rapid timescales depend on post-receptor signals, notably luminance-sensitive L3 lamina monopolar cells. Direction-selective motion signals emerge in T4/T5 via inputs from lamina and medulla neurons (Tm1, Tm2, Tm4, Tm9). Normalization is a canonical computation in sensory systems and machine vision that can stabilize feature representations by scaling responses by pooled activity. However, the specific neural implementation of rapid luminance gain control and the role of potential wide-field inputs and spatial pooling scales remained unclear. The authors build on these findings to test whether divisive normalization with local luminance pooling underlies stable contrast coding in naturalistic contexts.

• In vivo two-photon calcium imaging (GCaMP6f) in Drosophila visual system across hierarchy: lamina monopolar cells (L2, L3), medulla Tm neurons (Tm1, Tm2, Tm4, Tm9), and direction-selective T4/T5. Visual stimuli included drifting sinusoidal gratings (1 Hz; 100% Michelson contrast; five mean luminances) and moving OFF edges (−100% Weber contrast) over varying backgrounds.
• Quantification: Fourier F1 amplitude for grating responses; peak responses for OFF edges; luminance dependence via linear fit slope vs log-luminance; ANOVA/Tukey tests across luminance per contrast; sample sizes specified per figure.
• Spatial pooling assays: Online receptive field (RF) mapping with ternary white noise (2.5° squares or 5° stripes). Single-neuron stimulation with centered 5° drifting grating plus background annuli of varying diameter and luminance to test pooling. Size-varying centered gratings (5–30°) at multiple luminances to assess gain control vs stimulus extent.
• Natural scene analysis and modeling: Voltage-membrane single-compartment model fit to LMC (L2/L3) responses to gratings, then used to simulate responses to natural scenes under different luminances along random trajectories. Implemented luminance gain control by normalizing LMC-like signals using locally pooled luminance over varying pooling extents. Assessed trade-off using a loss function combining distribution similarity across luminance and preservation of response range; repeated on multiple scenes.
• Circuit modeling: Divisive normalization via shunting inhibition. Tm1 modeled as main L2 input divided by wide-field pooled signal (linear normalization, pooling extent ~15°). Tm9 modeled similarly with L3 input but requiring non-linear (quadratic) normalization factor to capture enhanced low-luminance responses.
• Glutamatergic inputs: Dendritic glutamate imaging with iGluSnFR in Tm1 and Tm9; spatiotemporal RFs via reverse correlation to estimate spatial FWHM of glutamatergic input vs calcium outputs.
• Genetics: Cell-type specific knockout of GluClα using FlpSTOP in Tm1 or Tm9; verification via tdTomato.
• Optogenetics and connectivity: csChrimson expression in glutamatergic Dm12; recorded Tm9 calcium while activating Dm12 with 625 nm light in blind flies; characterized Dm12 visual responses and single-neurite RFs.
• Connectomics: FAFB (FlyWire) analysis to identify Tm9 presynaptic partners, neurotransmitter predictions, and Dm12 column span (~3 columns ≈15°).
• Dm12 silencing: Kir2.1 expression in Dm12 while recording Tm9 to test necessity.
• Statistical procedures included ANOVA with Tukey HSD, t-tests, Shapiro–Wilk and Levene tests; details and sample sizes per figure; code and data availability provided.

• Luminance dependence in early vs late stages: L2 and L3 LMCs’ contrast responses vary strongly with mean luminance, failing to stably estimate contrast under rapid luminance changes. In contrast, T4/T5 direction-selective neurons show near luminance-invariant responses to the same stimuli (Fig. 1d–f; ANOVA and slope comparisons significant: e.g., L2–T4/T5 p≈4.9e−11, L3–T4/T5 p≈6.2e−08).
• Emergence in Tm neurons: Among Tm neurons connecting LMCs to T5, Tm2 and Tm4 remain luminance-dependent, whereas Tm1 and Tm9 exhibit rapid luminance gain control (Fig. 2). Tm1 is luminance-invariant; Tm9 shows negative slopes (enhanced responses at low luminance). Comparisons vs LMCs are significant (e.g., L2–Tm1 p≈3.49e−06; L2–Tm9 p≈1.16e−09). Across contrasts (20–100%), Tm1 remains invariant; Tm9 is invariant at high contrasts (≥80%) and enhances responses at low luminance for lower contrasts.
• Spatial pooling requirement: Annulus experiments reveal that modulating surrounding luminance changes the center neuron’s contrast response in a size-dependent manner, peaking around 10–15° of visual space for both Tm1 and Tm9, indicating local pooling. Size-scaling gratings show Tm1 becomes luminance-invariant at diameters of ~20–25°, while Tm9 is invariant for small sizes and develops negative slopes for larger sizes.
• Natural scene computations: Modeling shows that without normalization, LMC-like responses produce luminance-dependent contrast distributions; divisive normalization using local luminance pooling aligns distributions across luminance while preserving contrast range. There is a trade-off: too-narrow pooling collapses response range; too-broad pooling degrades background estimates due to rapid spatial decorrelation. Optimal pooling extents are local (on the order of several degrees), with scene-specific variation but consistent existence of an optimal local range.
• Mechanism via shunting inhibition: A normalization model with shunting inhibition and spatial pooling reproduces Tm1 and Tm9 behavior. Tm1 is captured with linear normalization (pooling ~15°); Tm9 requires non-linear (quadratic) normalization to account for low-luminance enhancement.
• Wide-field glutamatergic inputs: iGluSnFR imaging shows Tm1 and Tm9 dendrites receive glutamate signals with spatial FWHM ~15°, broader than their calcium RFs (~10°), consistent with multi-columnar pooling. Unlike output calcium signals, dendritic glutamate inputs scale with luminance, supporting a dendritic transformation implementing gain control. Slopes differ significantly between calcium and glutamate signals (e.g., Tm1: p≈7.2e−05; Tm9: p≈0.0056).
• Molecular substrate in Tm9: Cell-type specific knockout of GluClα in Tm9 abolishes luminance gain control, yielding responses that positively correlate with luminance (similar to LMCs). Luminance dependence increases significantly vs controls (control1–Tm9GluClαFlpSTOP p≈1.78e−05; control2–Tm9GluClαFlpSTOP p≈0.0022). GluClα loss in Tm1 does not alter its gain-control phenotype.
• Candidate wide-field neuron: Connectomics identifies glutamatergic Dm12 as a major Tm9 input; Dm12 spans ~3 medulla columns (~15°) and receives strong L3 input. Optogenetic activation of Dm12 suppresses Tm9 calcium (inhibitory connection), and Dm12 shows OFF-preferring responses with single-neurite RF FWHM ~17°, matching the pooling extent. Silencing Dm12 alone does not abolish Tm9 gain control, indicating distributed pooling across multiple inputs.
• Overall: Rapid luminance gain control arises in Tm1 and Tm9 dendrites via divisive normalization using locally pooled luminance and shunting inhibition; this computation is tuned to the statistics of natural scenes and stabilizes contrast for downstream motion processing.

The study locates the origin of rapid luminance gain control between lamina and direction-selective neurons, specifically in Tm1 and Tm9 dendrites. This directly addresses the question of how flies preserve stable contrast processing under rapidly changing luminance: by dividing center contrast signals by a locally pooled luminance estimate, implemented via shunting inhibition. The identified pooling extents (≈8–20°) align with natural scene statistics where luminance correlations decay with distance, explaining why local rather than global normalization is optimal. The differential behaviors of Tm1 (strict invariance) and Tm9 (low-luminance enhancement) suggest parallel channels that jointly support robust motion computation downstream in T5, ensuring stability across conditions and improving sensitivity in dim contexts. Mechanistically, glutamatergic wide-field inputs supply the pooled luminance signal; GluClα in Tm9 mediates inhibitory shunting for divisive normalization, while Tm1 likely uses a distinct inhibitory implementation. The convergence of modeling, physiology, genetics, and connectomics supports a general principle: local divisive normalization tuned to natural scenes underlies stable visual feature extraction, with potential analogs in vertebrate systems.

This work identifies where and how rapid luminance gain control is achieved in the fly: in Tm1 and Tm9 dendrites via local spatial pooling and divisive normalization. It clarifies algorithmic (local luminance pooling), circuit (wide-field glutamatergic inputs, notably Dm12 to Tm9), and molecular (GluClα-mediated shunting in Tm9) substrates. The computations are matched to natural scene statistics, optimizing the trade-off between luminance invariance and contrast resolution. Future research should delineate the full set of wide-field contributors beyond Dm12, clarify the molecular implementation of normalization in Tm1, explore interactions with contrast gain pathways (e.g., Tm2/Tm4), and assess how state or behavioral context modulates pooling scales. Extending analyses across broader luminance regimes and naturalistic behaviors will test generality across conditions and species.

• Dm12 contributes to Tm9 normalization but its silencing alone does not abolish luminance gain control, indicating distributed implementation across multiple presynaptic neurons that remain to be identified.
• The molecular implementation differs between Tm1 and Tm9; GluClα is required in Tm9, but the corresponding mechanism for Tm1 is unresolved.
• Experiments focused on luminance ranges typical of daytime activity; broader luminance regimes in nature may involve additional mechanisms or scaling.
• Not all third-order neurons exhibit luminance gain control (Tm2, Tm4 remain luminance-dependent), and integration of parallel gain mechanisms in downstream T5 is inferred rather than directly measured in this study.
• Spatial pooling optima show scene dependence; while the principle of local pooling holds, precise extents may vary across environments and were modeled for limited scenes.