Extended Data Fig. 5: Selective odour responses with stable within-day statistics and across-day drift in piriform cortex.

a, b, Fraction of single units that exhibit a significant modulation (a) and an increase (red) or decrease (blue) (b) in firing rate during the odorant stimulus epoch in response to 0–8 odorants (Wilcoxon rank-sum on firing rate during the odorant stimulus epoch versus spontaneous baseline firing rate, α = 0.001). c, Cumulative distributions (left) and mean coefficient of variation (c_v; right) of response magnitude computed on each odour test day across all trials for each odour–unit pair. Mean (95% CI) across days, c_v = 0.88 (0.87, 0.89), n = 19,356 odour–unit pairs. d, Cumulative distributions (left) and mean fraction of responses preserved per responsive single unit (right) across 8–32-day intervals (8 days: 35.0% (27.3%, 42.5%), 16 days: 19.8% (13.5%, 27.5%), 24 days: 16.9% (10.3%, 26.7%), 32 days: 6.6% (1.9%, 17.5%); ρ = –0.25, P = 5 × 10⁻⁶, n = 318 single units). Non-responsive and broadly responsive single units were excluded from the analysis by setting a threshold on lifetime sparseness (0.65). e, Left, fraction of preserved responses per single unit across 32 days versus lifetime sparseness threshold. Right, fraction of single units stable across 32 days versus lifetime sparseness threshold. A single unit was considered stable over 32 days if all significantly modulated responses to the odorant panel were preserved. These quantities do not depend on lifetime sparseness threshold (0.2–0.65, 40th–95th percentile across all single units). f, Classification accuracy (8-way, SVM, linear kernel, L2 regularization, trained and tested on data stitched across 3 mice, random draws of 231 single unit subsets from 286 total single units to avoid saturation, 1-s sliding window, 250-ms steps). Grey box, 4-s odorant stimulus epoch; vertical dotted line, onset of odour response at odour port (mean time across all stimuli at which the PID signal reached 5% of maximum); horizontal dashed line, chance performance for 8-way classification. g, Classification performance for fifteen temporal binnings of the odour response epoch, measured by maximum classification accuracy (top) and number of single units required to reach 50% of maximum accuracy (bottom) (n = 286 single units recorded within-day, stitched across 3 mice). Black shading, binning used for all subsequent classification and for computation of pairwise population vector correlations, angles and drift rate, unless otherwise indicated. For classification using single bins, the window start was set to 500 ms after stimulus initiation so the quantification windows did not begin before odorant stimulus onset as measured by the PID signal. h, Classification accuracy as a function of number of single units used, using the highest performance binning in g (four 2-s bins). Dashed arrow, number (21) of single units required to achieve >50% classification accuracy. i, Classification accuracy for a classifier trained on earlier days and tested on later days (‘Forward’, replotted from Fig. 2b) compared with a model trained on later days and tested on earlier day (‘Reverse’). Dotted lines, mean; shading, s.d.; limit of 41 single units per animal with 100 permutations. j, Classification accuracy of a classifier trained on responses on day 24 alone (all 56 trials) and tested on day 32 compared with a model trained on 75 random subsets of 56 trials drawn from days 0–24 and tested on day 32; P = 2.6 × 10⁻⁵, Wilcoxon rank-sum, 100 random subsets of 23 single units per mouse. A classifier trained on concatenated data from days 0–24 will assign high weights to single units with stable (less variable) responses across all days and low weights to single units whose responses varied. Thus, if there is a special population of neurons whose responses are informative about stimulus class and are more stable than others, a model trained on a concatenation of days 0 through 24 ought to perform better when tested on day 32 than a model trained on day 24 alone. However, we do not observe this: thus, it is not possible to establish single units that are most informative about odour identity on day 32 based on their responses across days 0–24. This finding argues against the presence of an informative stable subpopulation. k, Representational drift between a pair of days can be estimated by measuring the difference in odour-evoked population responses across days after correcting for within-day variability²⁵. Top left, variability across days (across-day drift + within-day variability), estimated by computing the angle (θ_p,q) between trial-averaged population vectors u_p and u_q for each odour across each pair of days p and q. Bottom left, variability within a day (noise), estimated by measuring the mean of the angle between the trial-averaged population vectors (\(\bar{\theta }\)) for each odour within each day on odd trials versus even trials (θ_k, over all days k). Right, the drift rate (r_p,q) is the corrected angle \(({\theta }_{p,q}-\,\bar{\theta }\)) divided by the time between days p and q (Δt_p,q). l, Cumulative distributions (left) and mean angles (right) between trial-averaged population vectors within-day and across 8–32-day intervals (n = 180, n = 144, n = 108, n = 72 and n = 36 pairs, respectively). Blue dotted line, exponential regression fit with \(\theta \,=\,C\,-\,(C-R)\,{e}^{\frac{-t}{\tau }}\), where \(\theta \) is the variability (angle), C the asymptote, R the intercept at t = 0 (within-day variability), and τ the time constant of the exponential in days. The mean rate of change of the exponential fit over the 32-day interval is 1.0 ° per day. m, Cumulative distributions (left) and mean within-day angles (right) between trial-averaged population vectors (n = 72 pairs per day). No pair of within-day angles differs significantly (P ≥ 0.56 for all pairs, Wilcoxon rank-sum). Black crosses, mean ± 95% CI; blue dotted line, linear regression; blue shading, 95% CI. Classification performed on the three mice presented with an eight-odorant panel. Otherwise, n = 6 mice.