Olfactory predictive codes and stimulus templates in piriform cortex

Abstract

Neuroscientific models of sensory perception suggest that the brain utilizes predictive codes in advance of a stimulus encounter, enabling organisms to infer forthcoming sensory events. However, it is poorly understood how such mechanisms are implemented in the olfactory system. Combining high-resolution functional magnetic resonance imaging with multivariate (pattern-based) analyses, we examined the spatiotemporal evolution of odor perception in the human brain during an olfactory search task. Ensemble activity patterns in anterior piriform cortex (APC) and orbitofrontal cortex (OFC) reflected the attended odor target both before and after stimulus onset. In contrast, prestimulus ensemble representations of the odor target in posterior piriform cortex (PPC) gave way to poststimulus representations of the odor itself. Critically, the robustness of target-related patterns in PPC predicted subsequent behavioral performance. Our findings directly show that the brain generates predictive templates or "search images" in PPC, with physical correspondence to odor-specific pattern representations, to augment olfactory perception.

Figure 1

Experimental design and behavioral results. A. Subjects took part in an olfactory attentional search paradigm that conformed to a two-way factorial design in which either odor target or odorant stimulus was varied. B. At the start of each fMRI run, subjects were informed whether the target smell would be odor A or odor B for that run. On a given trial, a countdown cue began 3 seconds prior to sniff (thick black bar), then one of the odor stimuli was presented and subjects indicated whether or not the assigned target odor was present in the stimulus (“y” or “n”).

Figure 2

Behavioral results. A. Overall detection accuracy for the target odor did not significantly differ across A and B runs, though performance was better with A and B stimuli than with the AB stimulus mixture (mix.). A cross-over interaction was observed, in which subjects were more accurate on target A runs when the stimulus was A (congruent condition; cong.) rather than B (incongruent condition; incong.), and were more accurate on target B runs when the stimulus was B (cong.) rather than A (incong.). B. Performance accuracy, collapsed across A and B runs, was significantly higher when the target was present in the stimulus (cong.) compared to when the target was not present (incong.). C. Reaction times followed a similar profile, such that subjects were faster for congruent trials compared to incongruent trials. *, P < 0.05.

Figure 3

Odor-specific predictive codes in the olfactory system. A. We made use of a single comparison (same target/different stimulus vs. same stimulus/different target) to look for both target-related effects (green) and stimulus-related effects (blue). These comparisons were computed for all target A runs (A|A compared to A|B vs. B|A; A|AB compared to A|A vs. B|AB) and for all target B runs (B|B compared to B|A vs. A|B; B|AB compared to B|B vs. A|AB), both before and after odor onset. B. Target- and stimulus-related effects were computed by testing whether multivoxel correlations between same-target/different-stimulus conditions (green bars) differed from different-target/same-stimulus conditions (blue bars). Same-target correlations significantly exceeded different-target correlations in APC and OFC both before and after stimulus arrival. In PPC, same-target conditions were more correlated than different-target conditions before stimulus onset, but same-stimulus conditions were more correlated than different-stimulus conditions after stimulus onset. *, P < 0.05. Error bars denote between-subject S.E.M. for each comparison.

Figure 4

The temporal evolution of predictive codes and stimulus representations differs across olfactory cortical regions. A. In all three regions, correlations increase prior to stimulus arrival for same-target conditions (blue lines), and remain elevated over same-stimulus conditions (red lines) in APC and OFC. In contrast, the correlation time-course in PPC exhibits two peaks: an early peak for same-target conditions, and then a later peak for same-stimulus conditions, reflecting the observed double dissociation between pattern type (target vs. stimulus) and time (pre vs post) in this region. Each plotted point represents the mean over two consecutive TRs. Black stars indicate time-points where same-target correlations exceed same-stimulus correlations; black diamonds indicate time-points where same-stimulus correlations exceed same-target correlations (at P < 0.05). B. To better visualize these effects, the correlation difference between the blue and red lines in panel A was plotted for each region at each time-point (ST(r), same-target r-value; and SS(r), same-stimulus r-value). A shift in pattern coding from target to stimulus is apparent in PPC at the crossing of the x-axis. Red stars, significant target-related effect; red diamonds, significant stimulus-related effect; P < 0.05.

Figure 5

Predictive odor templates in PPC resemble the stimulus response to the target smell. (A) In PPC, pre-stimulus target-specific ensemble patterns more closely resembled post-stimulus odor patterns when the target matched the stimulus (e.g., A|A vs. B|A) compared to when it did not (e.g., A|A vs. A|B). Subject-averaged correlations for matching and non-matching conditions, averaged across both A and B targets, are shown (means ± s.e.m.). (B) A scatterplot of matching versus non-matching conditions in PPC indicates that the pre-stimulus target template was more highly correlated to matching (vs. non-matching) post-stimulus odor representations for 10/12 subjects.

Figure 6

Feature-specific pre-stimulus patterns augment olfactory perception. (A) The strength of pattern correlation between same-target conditions in PPC predicted identification accuracy on the odor search task. (B) The extent to which PPC ensemble overlap was greater for same-target conditions than for different-target conditions was also positively correlated with performance, on a subject-by-subject basis. Note, each dot represents one subject.

Figure 7

Univariate fMRI analysis reveals that the mean level of odor-evoked activity in MDT is reduced in response to an expected or predictable stimulus. The group-averaged mean percent signal change is plotted over time within each ROI. In MDT, unexpected conditions elicited a higher response magnitude than did expected conditions.