Table S1: Leave-One-Experiment-Out Cross-Validation Results

Leave-one-experiment-out cross-validation was performed to assess model generalization. For each of 14 experiments, the factorial NB-GLM was fit on the remaining 13 experiments and used to predict event rates for the held-out experiment.

Table S2: Model Comparison

Table S2b: GLMM Robustness Check

A Negative Binomial GLMM with random track intercepts was fit using Bambi/PyMC to verify that the main findings are robust to hierarchical structure.

Table S3: Condition-Specific Suppression Amplitudes

The kernel amplitude varies across conditions while maintaining invariant shape and timescales:

Figure S1: Residual Diagnostics

Figure S2: Time-Rescaling Test

The time-rescaling test assesses whether the fitted hazard model produces inter-event intervals consistent with a Poisson process. Under the correct model, rescaled inter-event times should follow Exp(1), and their cumulative distribution should be uniform.

Diagnostic Statistics

• Pearson residuals: Mean = 0.0001, SD = 1.01, Skew = 40.7, Kurtosis = 2440

• Large residuals: 7,288 observations (0.09%) with \(|r| > 3\)

• Time-rescaling: KS statistic = 0.041, p = 0.17, mean deviation = 1.3%

The high skewness and kurtosis of residuals reflect the zero-inflated nature of event data (most frames have no events). The time-rescaling test passes at conventional significance levels, supporting the adequacy of the hazard model specification.

Table S4: Kernel Parameters with Bootstrap Confidence Intervals

Bootstrap resampling (100 track-level resamples) was used to estimate 95% confidence intervals for all kernel parameters.

Table S5: Turn Distribution Parameters

Turn angle and duration distributions from 319 filtered events (turn duration \(> 0.1\) s) were used to parameterize trajectory simulation.

Figure S3: Reverse Crawl LED Modulation

Reverse crawl detection using Mason Klein’s algorithm (SpeedRunVel \(< 0\) for \(\geq 3\) s) identified 1,853 reversal events across all 14 experiments. In contrast to reorientations (which are suppressed by LED), reverse crawls are increased during LED stimulation.

Figure S4: Reverse Crawl Detection Validation

Reverse crawl detection was validated against the original MATLAB code (mason_analysis.m) on Track 2 of the reference experiment.

Figure S5: PSTH Model Validation

The peri-stimulus time histogram (PSTH) provides a visual comparison of model predictions against empirical event rates aligned to LED onset.

Table S6: Per-Condition Kernel Parameters

The gamma-difference kernel was fit separately to each of the four experimental conditions in the \(2\times2\) factorial design. Bootstrap confidence intervals (95%) were computed from 200 resamples.

Key finding: The 50-to-250 Constant condition shows a 4-fold slower fast timescale (\(\tau_1 = 1.18\) s vs \({\sim}0.3\) s), suggesting that baseline neural excitation modulates sensory transduction speed.

Table S7: Model Comparison

The gamma-difference model achieves near-optimal fit quality (\(R^2 = 0.968\)) with half the parameters of the raised-cosine basis, while providing biological interpretability (timescales map to neural processes).

Note: The single exponential model shows \(R^2 < 0\) because it cannot capture the biphasic (suppression-then-recovery) kernel shape. A negative \(R^2\) indicates the model performs worse than predicting the mean, which is expected when fitting a monotonic decay to a non-monotonic target.

Figure S7: Event Duration Distributions

Event durations from Mason Klein run tables and trajectory segmentation characterize the temporal structure of larval behavior across conditions.

Figure S8: Fractional Behavior by Pulse

Behavioral state fractions (run, pause, turn, reverse crawl) were computed for each pulse across the 20-minute experiments. The stacked bar plots reveal systematic shifts in behavioral allocation over successive pulses.