Results: from receptor to amygdala

3.1 Characteristics of the studies

The 119 publications span 1991 to 2026. By design, narrative/theoretical reviews form the largest group (40), followed by neuroimaging (24), behavioural/psychophysical studies (13), meta-analyses (10, including one mini-meta-analysis [32]), animal/molecular studies (10), human electrophysiology (9), systematic reviews without pooling (7), RCTs (2) plus one double-blind placebo-controlled experiment [106], non-randomised interventions (2), a cohort (1) and a lesion study (1). On the OCEBM scale, 58 papers are at level 5 (mechanistic, animal, expert opinion), 34 at level 4, 13 at level 3, 3 at level 2, 3 at level 1b and 8 at level 1a (Figure 2; Appendix A). Thus nearly 4/5 of the evidence base is mechanistic or small-scale experimental — appropriate for the mechanistic question (Q1), but a structural limitation when answering the applied question (Q3).

Typical sample sizes in the human literature are very small: fMRI usually 5–25 people (Herz 2004 n = 5 [31]), dense-sampling designs only 3 people with ~18 hours of scanning each [12, 13]; iEEG 7–17 epilepsy patients [11, 35, 51, 76]. By contrast, the applied literature has large syntheses: Donelli 2019 (65 RCTs, 7,993 people) [101], Roschk 2020 (71 samples, 15,447 people) [117], Hu 2020 (111 samples, N = 2,004) [83]. The epidemiological data comprise only one prospective cohort (WHICAP, n = 1,037) [96].

3.2 The periphery: receptors, transduction and signal termination

Gene family and numbers. Buck & Axel isolated a family of 7TM genes specific to the olfactory epithelium, estimating ≥100 and “possibly several hundred” genes, with variability concentrated in TM3–TM5 — the predicted ligand-binding region [40]. Estimates of the number of human OR genes vary considerably between sources: ~400 functional [1, 8]; 391 [6]; ~394 out of ~857 potential [9]; ~350–400 out of ~1,000 [33]; ~700 in total, about half functional [10]. The discrepancies reflect annotation criteria (intact gene vs pseudogene vs “possibly functional”) and date of publication; a reasonable consensus figure is ~400 functional ORs out of ~850–900 loci. Only 143–346 ORs have been observed to be expressed, depending on the technique [9].

Structure. Billesbølle 2023 published the first cryo-EM structure of a human OR (OR51E2–Gs, 3.1 Å): propionate binds in an enclosed 31 ų pocket between TM3–TM6, anchored by R262^6x59; enlarging the pocket with the F155A mutation (90 ų) shifted selectivity towards C8 fatty acids [42]. The result for one receptor (OR51E2) suggests that the tuning breadth of ORs depends on the geometry of the binding pocket, and that class I and class II ORs have distinct conserved motifs (R^6x59 in 71% of class I vs 7% of class II).

Combinatorial code and peripheral computation. Each OSN expresses one allele of one OR gene; each odour activates a combination of ORs, and the combination changes with concentration [8, 17]. The key change from the classical model: ~50% of ~800 odorants antagonise at least one receptor, mostly competitively; synergy dominates at low concentrations, antagonism at high concentrations [8]. Two important caveats: these experiments used mouse receptors (Olfr) and have not been replicated with human receptors; and the test concentration (300 µM) is many orders of magnitude higher than odour concentrations in natural air (usually below a few ppb), so the proportion of antagonism at physiological concentrations may be considerably lower. Consequence: a perfume blend is not a linear sum of its components, right from the epithelium.

Genetic variation → perception. Keller 2007 showed that OR7D4 is the only receptor out of 337 ORs that responds to androstenone; carriers of the RT/WM allele (24%) rated androstenone as weaker and more pleasant (less use of the descriptor “sickening”, more use of “vanilla”; 2/146 descriptors differed) [41]. Mainland (2014, cited in [1]) extended this: OR10G4 genotype explains 15% of the variance in intensity and >10% of the variance in valence of guaiacol. This is the most clearly demonstrated gene → receptor → valence causal chain in humans.

Transduction and termination. The cascade OR → Gα_olf → ADCY3 → cAMP → CNG channel → Ca²⁺ → ANO2 Cl⁻ channel is canonical [9, 10]; the Cl⁻ current accounts for up to ~90% of the receptor current [44]. Unlike retinal rods, OSNs provide almost no amplification: ~40 binding events (~1.2 pA) are needed for a 50% probability of one spike [44]. Stephan 2012 separated two processes that are often conflated: NCKX4 knockout did not change amplitude or latency but lengthened the decay time constant τ from 146 to 612 ms and caused excessive adaptation (WT recovered by 4 s; KO was still depressed at 15 s) [43]. Significance: the speed at which Ca²⁺ is extruded from the olfactory cilia determines the ability to keep pace with a 2–10 Hz sniffing rhythm; Reactome does not yet model this branch [9].

3.3 Olfactory bulb and piriform cortex: from chemical features to odour objects

Olfactory bulb. OSN → glomerulus convergence is ~10³ axons in the mouse [10]; histological counts in humans give roughly 5,000–8,000 glomeruli per olfactory bulb (5,568 ± 830 in a sample aged 39–89; about 8,000 in young people), with large inter-individual variation and a decline with age; the frequently cited figure of 1,100–1,200 is not based on any count in humans (see Appendix C). Each glomerulus has ~20 mitral and ~50 tufted cells, with different tuning among “sister” cells — the bulb’s output is not a copy of its input [45]. Tufted cells (low threshold, responding early in the sniff cycle, projecting to anterior cortex) and mitral cells (high threshold, late, projecting to the whole olfactory cortex) form two parallel temporal channels [45]. Go/no-go learning increases the number of activated glomeruli for up to 5 weeks [14]; in mice, odour fear conditioning increases the number of OSNs expressing the corresponding receptor [26] — plasticity occurs right at the first stage.

APC vs PPC. Gottfried 2006 described a double dissociation using fMRI cross-adaptation: the APC encodes structure, the PPC encodes perceptual quality [48]. Howard 2009 confirmed this with MVPA: odour identity could be decoded from multivoxel patterns in the PPC (P = 0.030) and OFC (P = 0.028) but not in the APC or amygdala, and not from mean activity; the distance space of the PPC matched perceptual space (goodness-of-fit 0.53) but not 32 molecular descriptors [49]. At the single-cell level, Kehl 2024 showed that the piriform cortex decodes chemical identity better than perceived identity (75.8% vs 22.1%), whereas the hippocampus shows the reverse (26.7% vs 69%) [11] — a chemical → conceptual gradient along the piriform → hippocampus axis.

Pattern completion and odour objects. Expert wine tasters can reliably identify only ~3 components in a mixture [46]. In rats, the piriform cortex performs pattern completion when a 10-component mixture lacks one substance, but decorrelates strongly when several are missing or a distractor is present — matching behaviour [47]. 20–40% of APC/PPC neurons respond supra-additively and 40–60% sub-additively to mixtures [46]. Converging conclusion: the perceptual unit of olfaction is a configural “odour object”, synthesised in the piriform cortex on the basis of experience, not a molecular spectrum.

The piriform cortex as associative memory and predictive model. Optogenetic activation of only ~500 cells (~0.5% of the piriform population) is enough to drive learned approach/avoidance behaviour [47]; silencing the ensemble active during learning impairs fear memory, and reactivating it evokes fear [27]. In humans, Zelano 2011 showed that before an odour appears, the PPC has already formed a “template” matching the pattern evoked by the target odour (P < 0.04), and that template strength predicts odour-search accuracy across individuals (R = 0.64) [50]. Lyons & Gottfried proposed a predictive coding framework: the piriform cortex sends predictions down to granule cells, and mitral cells carry prediction errors [15] — a well-grounded hypothesis that has not yet been causally tested in humans.

Limits of resolution. The claim that “humans can discriminate 1 trillion odours” [55] rests on a sphere-packing extrapolation; Gerkin & Castro showed that the same formula gives results ranging from ~4.6×10³ to 1.5×10²⁹ depending on the parameters, so the figure is only an upper bound; Meister showed that this method assigns ~9×10¹¹ “odours” to a system with only 3 percepts, and that olfactory perceptual space may have only 1–20 dimensions [55]. Yeshurun & Sobel argued that odour perception is almost one-dimensional along valence [57], with the valence axis accounting for ~30% of the variance (Khan 2007, cited in [1]). A GNN model predicted odour labels closer to the panel mean than the median rater did for 53% of molecules [56], showing that structure–odour relationships are predictable at the population level but poorly at the individual level (DREAM: r = 0.21 individual vs 0.41 population [1]).

3.4 The amygdala: valence, intensity and the gateway to emotion

Input anatomy. The amygdala receives direct olfactory bulb input via its “olfactory” nuclei (MeA, CoA, PAC). Yang 2025 (tractography, n = 25) ranked connection density with the olfactory bulb as MeA > CeA > ACo > PAC > LA > BLA (F = 9.31; p = 6.6×10⁻¹⁰); the olfactory group of nuclei connects more strongly with the OFC, insula and ventral striatum, whereas the BLA/LA group connects more strongly with the hippocampus [22]. Noto 2021 reconstructed the MeA/CoA/PAC boundaries from resting-state connectivity alone (z ≥ 6.05) [23]. Hakim 2019 noted that odours do not go straight into the BLA as sounds/images do; the periamygdaloid complex may be the gateway [26]. An ALE meta-analysis (81 studies, 1,744 people) identified the largest activation cluster for all odour types as the left amygdala/piriform (k = 1,163 voxels) [4].

Valence or intensity? This is a historical controversy. Zald & Pardo (PET, n = 12) found that the left amygdala correlated with unpleasantness (r = −0.51) [60]; Anderson 2003 argued that the amygdala encodes intensity and the OFC encodes valence [52]. Winston 2005 resolved this with a 3 odour types × 2 concentrations design: the amygdala increased with concentration only when the odour had valence (pleasant +0.14%; unpleasant +0.23%; neutral 0.0001%; interaction p = 0.017), whereas the posterior piriform and medial OFC encoded intensity independently of valence [62]. The amygdala therefore encodes emotional significance (valence × intensity) — compatible with both earlier results.

Single-cell evidence. Kehl 2024 (2,416 units, 17 patients) provides the strongest evidence: amygdala neurons fired more strongly for liked odours (Z = 2.9; P = 0.004) and correlated with the normative valence scale (Spearman r = 0.56); no other region showed this effect [11]. The amygdala decoded identity from 400 ms, only 50 ms later than the piriform cortex. Sagar 2026 added evidence at the voxel-population level: aversive odours were decoded better than appetitive odours in the piriform cortex and amygdala (0.174 vs 0.059; z = 4.58) but not in the OFC; the piriform cortex and amygdala use separate codes for the two valence classes, and no region has an integrated valence axis [13].

Valence at an earlier stage. Iravani 2021 recorded non-invasive olfactory bulb EEG (n = 19): the gamma code correlated with valence from ~250 ms (r = 0.60), negative odours produced early beta synchronisation (50–200 ms), and positive odours late synchronisation (690–780 ms); a pre-registered experiment (n = 44) confirmed that unpleasant odours cause people to lean away within 0.5 s, whereas pleasant odours do not cause leaning forward [65]. In mice, Root 2014 demonstrated causality: silencing olfactory bulb terminals in the CoA abolished both TMT avoidance (−65 → −8) and 2-phenylethanol preference (+48 → −8), whereas silencing in the tubercle or piriform cortex did not; optogenetic activation of the tagged CoA ensemble reproduced the behaviour [66]. Thus innate valence has its own hard-wired circuit via the CoA, in parallel with learned valence via piriform–BLA–OFC [3, 16].

The reverse influence: emotional state shapes perception. Krusemark 2013 induced anxiety (15.6 → 40.4), which made neutral odours rated as more unpleasant (the change in valence of neutral minus negative odours correlated with the increase in anxiety, r = −0.61) and created a new effective connectivity pathway APC → amygdala → PPC/OFC [64]. Fear conditioning made two indistinguishable odours distinguishable (accuracy more than doubled), accompanied by decorrelation of PPC patterns [63]. In both cases, emotional state rewrites the odour representation in sensory cortex; the driving role of the amygdala is seen directly in the DCM analysis [64], whereas in fear conditioning the amygdala response to the CS+ actually declined over time [63].

Emotional chemosignals. The de Groot & Smeets meta-analysis (26 studies, N = 1,652) gave g = 0.36 (95% CI 0.31–0.41; I² = 2.8% after removing outliers); the p-curve was right-skewed and p-uniform detected no publication bias [71]. Sweat from first-time skydivers activated the left amygdala in recipients who could not distinguish the odour (independent replication: Z = 3.88), and made the psychometric function for identifying threatening faces ~43% steeper [73]. Fear sweat increased sniffing and visual scanning; disgust sweat did the opposite [70]. The volatile-compound sensor (PID) signal was higher for fear sweat (10.3 vs 5.1 mV); a model including pad weight, emotion and sample freshness explained R² = 0.98, but from only 3 donors [74]. A systematic review of 45 studies found the strongest evidence for negative emotions, with women more sensitive [72]. Limitations: small original sample sizes (mean 35, range 7–156), and the active molecules have not been identified.

Scientific reference material, not medical advice. Figures are redrawn from published data, not copied from the original figures. GRADE and risk-of-bias ratings are the compilers’ own. Every number traces back to detailed reading notes on each paper.