Ketamine improves short-term plasticity in depression by enhancing sensitivity to prediction errors

Sumner and colleagues frame ketamine as a rapid and robust antidepressant whose downstream effects on neural signalling and plasticity likely contribute to its therapeutic action. While ketamine acts as an NMDA receptor antagonist, the investigators note that its antidepressant onset (around 2 h) and durability (about a week) point to multilayered mechanisms beyond acute receptor blockade, including enhanced Hebbian long-term potentiation (LTP) and other pro-glutamatergic, AMPA receptor-mediated changes that emerge within hours. The authors position predictive coding — a hierarchical, recursive process in which the brain issues top-down predictions and uses bottom-up prediction errors to update an internal generative model — as a complementary plasticity mechanism to Hebbian LTP that could be relevant to depression and to ketamine's effects. The study sets out to test whether ketamine modulates short-term predictive-coding processes in people with major depressive disorder (MDD). Specifically, the investigators used a roving auditory mismatch negativity (rMMN) paradigm and dynamic causal modelling (DCM) of EEG to probe MMN, P3a, repetition suppression, and the effective connectivity underlying deviant-stimulus processing 3–4 hours after a single subanaesthetic ketamine infusion. The primary aim was to determine whether ketamine, in the early window when antidepressant effects first emerge, increases sensitivity to prediction errors and alters short-term plasticity in patients with treatment-resistant depression.

The study used a randomised, double-blind, active placebo-controlled crossover design in 30 participants meeting DSM‑V criteria for MDD who had failed at least two prior recognised treatments. Each participant received racemic ketamine on one visit and an active placebo (remifentanil, target plasma concentration 1.7 ng/mL) on the other; order was randomised and counterbalanced. The clinical primary outcome for the broader trial was change in Montgomery–Åsberg Depression Rating Scale (MADRS) at 24 h post-infusion, and linear mixed models were used to assess drug-by-time effects. The EEG/MMN experiment was conducted approximately 3–4 h after infusion, a period chosen to capture early downstream neural changes that coincide with the emergence of antidepressant effects. Ketamine was administered intravenously by an anaesthetist as a 0.25 mg/kg bolus followed by a 0.25 mg/kg·hr infusion for 45 minutes; dosing for participants with BMI >30 kg/m2 used ideal body weight. Remifentanil was delivered via a target‑controlled infusion based on the Minto model. Participants were scanned in an MRI environment and provided informed consent under ethical approval. The electrophysiological probe was a roving auditory oddball (rMMN) task presented 3–4 h post-infusion. Trains comprised 1–11 identical tones; the first tone in a train served as the deviant and tone 5 served as the standard. Train lengths were pseudorandomised to prevent anticipation. Tones were 70 ms with 500 ms interstimulus interval, frequencies varied between 500–800 Hz in 50 Hz steps, and participants performed a visual distractor task to keep the auditory stimuli unattended. Continuous EEG was recorded with a 64‑channel system and preprocessed in SPM12; averaged ERPs were constructed for tones 1–11 and converted to NIfTI images covering 0–400 ms, smoothed with a 6×6×6 mm kernel. Main ERP analyses comprised a 2×2 ANOVA testing deviant (tone 1 versus tone 5) and drug (ketamine versus placebo), and a 2×4 ANOVA probing repetition suppression using difference waves (deviant minus tones 2–5). Family-wise error correction controlled multiple comparisons for main effects (p < 0.05 FWE‑c); a more liberal uncorrected threshold (p < 0.001) was used for interactions with interpretation limited to contrasts surviving FWE‑c. Source reconstruction used Multiple Sparse Priors to identify generators in the MMN (200–300 ms) and P3a (300–400 ms) windows; selected sources (including bilateral primary auditory cortex coordinates from the literature) defined nodes for the DCM. DCM employed the standard ERP neural mass model to estimate modulatory effects associated with deviant versus standard and ketamine versus placebo. Six model architectures were specified (forward only; backward only; forward+backward; each with or without intrinsic connections), yielding 36 combinations for main effects plus additional interaction models, for a total of 42 inverted models. Bayesian model selection was applied at the family level and then within the winning family to select the best model (using protected exceedance probabilities and Bayesian Omnibus Risk). Posterior parameter estimates were tested with Wilcoxon signed‑rank tests (FDR correction) and correlated with percent change in MADRS at 24 h using Spearman's rho (FDR corrected).

Behavioural/clinical outcome: Across the trial cohort 70% of participants achieved at least a 50% reduction in MADRS at 24 h post‑ketamine; the primary MADRS analyses showed ketamine reduced scores relative to placebo at 3 h, 1 day and 7 days post‑infusion, and no carry‑over effects were detected in the crossover. ERP (sensor‑level) findings: The 2×2 ANOVA confirmed a robust MMN, peaking at 243 ms with a frontal distribution (F(1,116)=114.33, p < 1.00e‑17 FWE‑c) and bilateral temporal peaks at 233 ms; a P3a component appeared at 333 ms (F(1,116)=66.38, p = 9.62e‑9 FWE‑c). There was no significant main effect of drug in the initial ANOVA, though a ketamine×deviant interaction appeared at 225 ms at an uncorrected threshold (p = 0.001 uncorrected) and did not survive FWE correction. When repetition effects were explicitly modelled (deviant minus subsequent tones), no main effect of tone number (repetition suppression) was found. A main effect of ketamine was observed at multiple loci: a left‑lateralised frontal MMN‑consistent peak at 224 ms (F(1,232)=56.10, p = 8.10e‑5 FWE‑c), a right posterior negativity at 160 ms (F(1,232)=21.02, p = 0.027 FWE‑c), a right temporal peak at 222 ms (F(1,116)=29.07, p = 0.001 FWE‑c), and a central positivity at 352 ms consistent with P3a (F(1,116)=26.28, p = 0.003 FWE‑c). Post‑hoc contrasts showed the MMN and 160 ms negativities were more negative in the ketamine session (e.g. MMN t(232)=7.49, p = 1.04e‑8 FWE‑c), while the temporal peak and P3a were more positive post‑ketamine. ERP–MADRS covariate: An exploratory voxelwise test relating the ketamine effect on deviant−standard difference waves to percent change in MADRS at 24 h revealed no significant relationships. Source analysis: In the 200–300 ms window, bilateral inferior temporal cortex (ITC) and a right inferior frontal gyrus (IFG) source were identified and used for DCM nodes; primary auditory cortex (A1) was modelled as input using coordinates from prior literature. Comparing sessions, the right inferior temporal gyrus (ITG, MNI 48 −14 −38) showed significantly greater activation post‑ketamine (F(1,116)=23.99, p = 0.022 FWE‑c), whereas left fusiform gyrus activation was lower overall post‑ketamine (F(1,116)=22.31, p = 0.041 FWE‑c). The 300–400 ms window did not justify adding sources for P3a modelling. DCM results: Family‑level Bayesian model selection identified the forward+backward (FB) family as best for explaining the deviant effect (family exceedance probability fxp = 0.780). Within the FB family the FB model was also selected as the winning architecture for the ketamine effect (pxp = 0.753, BOR = 0.16). For the FB model explaining the deviant effect, Wilcoxon signed‑rank tests on posterior parameter estimates showed the right A1→right ITC connection was significantly stronger for the standard than the deviant (M = 0.23, SD = 0.36; Z = 3.61, p = 0.002 FDR); no other connection survived FDR correction for deviant or ketamine effects. DCM–clinical correlation: Modulation of the right A1→ITC connection (the same connection that differed between deviant and standard) was significantly negatively correlated with percent change in MADRS at 24 h (Spearman r_s = −0.578, p = 0.006 FDR), indicating participants with greater antidepressant responses showed stronger modulation of this forward connection for standard tones.

Sumner and colleagues interpret their findings as evidence that a single subanaesthetic ketamine infusion increases sensitivity to prediction error several hours post‑dose in people with MDD. The ERP analyses showed a larger MMN and altered P3a several hours after ketamine compared with the active placebo, and source‑level results implicated increased right inferior temporal activation. In the DCM framework the winning model included both forward and backward connections, and modulation of the forward right A1→ITC connection related to clinical antidepressant response, which the authors argue reflects increased sensitivity to prediction errors rather than improved repetition suppression. The investigators emphasise that, despite increased MMN amplitude, they observed a remarkable lack of repetition suppression across tone repetitions in this depressed sample and found no ketamine effect on repetition suppression at 3–4 h. They note this may suggest ketamine restores error sensitivity without immediately restoring short‑term calibration of the generative model; synaptic and structural processes supporting long‑term plasticity (spine and synapse growth) occur later, around 24 h, and repetition suppression might therefore require a longer follow‑up to reveal change. Limitations highlighted by the authors include the absence of a healthy control group, which prevents asserting specificity of the effects to depression; the fact that 20/30 participants were taking antidepressant medications (mitigated partly by the within‑subject crossover design); and analytic constraints in SPM that limited some covariate modelling to simpler one‑sample tests. The choice to use inferior temporal cortex sources for DCM — driven by data‑driven source localisation in this cohort — may also complicate comparison with many MMN DCM studies that use superior temporal gyrus sources, although the authors report a supplementary replication using STG. In conclusion, the study reports that ketamine increases neural sensitivity to unexpected sensory input and prediction errors 3–4 h after infusion, and that modulation of a forward auditory→temporal connection correlates with antidepressant response at 24 h. The authors propose that enhancement of prediction‑error signalling may be one mechanism contributing to ketamine's rapid antidepressant effects, while changes in repetition suppression and model calibration may emerge on a longer timescale and warrant further investigation.