Results
First, we applied white noise odor waveforms and recorded the response of PNs innervating the DM4 glomerulus. Using the obtained odorant-response input/output pairs we constructed a one-dimensional Linear-Nonlinear (1D LN) cascade model that can faithfully predict the measured average spike rates. However, both the linear kernel and the nonlinearity of the model change with the mean concentration and the variance of a white noise stimulus. This dependence of model parameters on test stimuli suggested us to consider higher dimensional models.
Next, we used the spike-triggered covariance (STC) method to build a two-dimensional LN (2D LN) cascade model for the OSN spike train-PN spike train input/output pair. The first principal eigenvector of the STC matrix is monophasic and tuned to the slow amplitude component of the input. At the same time, the second principal eigenvector is biphasic and tuned to the variance component of the input. Although the model closely predicts PN output for white noise waveforms, it fails to predict PN output for non-stationary odor waveforms. By analyzing the trajectory of white noise stimuli, we determined that the predictive power of the 2D LN model is adversely affected by the limitations imposed on the set of input odor waveforms.
Finally, we estimated the non-linear block of the 2D LN model using non-stationary triangle odor waveforms with bigger trajectories and propose a model that can predict PN output to a larger set of stimuli. Combining the resulting model for the DM4 glomerulus with the OSN model in [1], we demonstrate that PNs most strongly encode information about both the first and the second derivative of odor concentration.