Predicting the outcome of psilocybin treatment for depression from baseline fMRI functional connectivity

Depression is a leading cause of disability worldwide and many patients fail to respond to first-line treatments; those who do not improve after at least two adequate antidepressant trials are commonly labelled as having treatment‑resistant depression (TRD). Psilocybin, administered with psychological support in controlled settings, is under active investigation as a novel treatment for TRD and major depressive disorder. Although prior trials indicate clinically meaningful antidepressant effects and an association between certain acute subjective experiences (for example, mystical-type and emotional breakthrough experiences) and clinical improvement, treatment responses are heterogeneous and it remains important to identify baseline markers that predict who will benefit. This study set out to test whether baseline resting‑state functional connectivity (FC) measured with fMRI can predict subsequent symptomatic outcome after psilocybin treatment. Copa and colleagues applied a machine learning pipeline to whole‑brain FC features derived from an open‑label TRD sample (n=16) who received two psilocybin doses (10 mg then 25 mg, one week apart) and tested generalizability in an independent sample (n=22) from a separate psilocybin trial. The investigators aimed to classify responders versus non‑responders up to 24 weeks after treatment and to identify the FC patterns most predictive of early and late clinical improvements.

Overall design and datasets: The primary data came from an open‑label clinical trial of psilocybin for TRD (original imaging reported previously), yielding usable baseline resting‑state fMRI for 16 participants. An independent dataset of 22 patients with moderate‑to‑severe major depression was used to test generalizability, and analyses were also repeated on the combined 38‑subject sample. Both studies had ethical approvals and informed consent consistent with Good Clinical Practice. Participants and clinical labels: Subjects in the TRD dataset met HAM‑D criteria for moderate‑to‑severe depression (HAM‑D‑17 ≥ 17) and had failed at least two different adequate antidepressant courses within the current episode. Responders were defined at multiple follow‑up times as those achieving at least a 50% reduction in QIDS‑SR16 relative to baseline; binary labels were constructed for time points immediately post‑treatment and at weeks 1, 2, 3, 5, 12 and 24. The extracted text reports responder counts by time point (e.g. 13 immediately after the second dose, 12 at week 1, 11 at week 2, 12 at week 3, 9 at week 5, 7 at week 12, 7 at week 24). fMRI acquisition and preprocessing: Both datasets were acquired on a 3 T Siemens Tim Trio scanner with eyes‑closed resting‑state T2* echo‑planar imaging (3 mm isotropic voxels). Preprocessing steps included removal of initial volumes, de‑spiking, slice‑time and motion correction, brain extraction, registration to anatomy and MNI space, scrubbing based on framewise displacement (FD) with exclusion of subjects who had >20% scrubbed volumes using FD threshold = 0.5, spatial smoothing (6 mm FWHM), band‑pass filtering (0.01–0.08 Hz), detrending and regression of nuisance signals (motion, ventricular, draining‑vein and local white‑matter signals). The mean and maximum percentage of scrubbed volumes in baseline recordings are given (mean 4.6±5% and maximum 17.3%). Feature construction and reduction: BOLD time series were averaged within 401 predefined cortical and subcortical ROIs. Pairwise linear correlations between ROI time series produced the full FC matrix (≈80,200 independent features). To reduce dimensionality and focus interpretation, features were restricted to connections between all ROIs and ROIs belonging to seven major resting‑state networks (RSNs): primary visual (PV), extrastriate visual (ESV), auditory (AUD), sensorimotor (SMN), default mode network (DMN), executive control (EXC), and salience (SAL). This retained inter‑areal FC between each RSN and the rest of the brain while substantially cutting feature numbers. Machine learning procedure and evaluation: A supervised binary classifier based on gradient boosting (ensemble of decision trees) was implemented in scikit‑learn. Owing to limited sample size, stratified 5‑fold cross‑validation was used, with folds balanced for class labels. Within each fold a K‑best univariate feature selection was applied; K was chosen among power‑of‑two settings (p = 3, 4, 5, 7). Performance was quantified by area under the ROC curve (AUC). To estimate statistical significance, the entire CV procedure was repeated 1000 times both with true labels and with labels randomly shuffled; p‑values were computed as the fraction of shuffled runs that achieved AUC greater than the true‑label runs. Feature importance scores were averaged across trees, folds and iterations to highlight FC connections driving classification.

Overall pattern: Copa and colleagues report that baseline FC in distinct RSNs predicted clinical response to psilocybin at different post‑treatment latencies. Visual networks predicted early symptom improvement, fronto‑parietal networks predicted early changes around two weeks, and salience network FC predicted sustained response at later follow‑up. Primary numerical results: For the primary visual network (PV) and extrastriate visual network (ESV) the classifiers attained high AUCs for early outcomes: PV at week 2 AUC = 0.93 ± 0.08 (p = 0.035) and PV at week 5 AUC = 0.95 ± 0.06 (p = 0.021); ESV at week 1 AUC = 0.90 ± 0.09 (p = 0.034), ESV at week 2 AUC = 0.92 ± 0.09 (p = 0.028), and ESV at week 3 AUC = 0.90 ± 0.09 (p = 0.04). The DMN and executive control (EXC) networks predicted response at week 2 (DMN AUC = 0.883 ± 0.10, p = 0.041; EXC AUC = 0.90 ± 0.09, p = 0.04). Long‑term outcome at week 24 was predicted by salience network FC (SAL AUC = 0.89 ± 0.10, p = 0.047). Feature anatomy and directionality: Feature importance analyses showed that responders tended to have greater fronto‑parietal FC relative to non‑responders. The visual network predictors were driven by long‑range occipital‑to‑frontal connections, strong interhemispheric occipital links and occipito‑temporal and occipito‑parietal connections (especially for the ESV). DMN‑related predictive features involved frontal and frontal‑occipital links (predominantly left frontal). EXC importance mirrored the visual networks' distribution. SAL features that predicted long‑term response were concentrated in within‑frontal, fronto‑limbic and insular‑limbic ROI pairs, mainly intrahemispheric. Generalizability and combined sample: Applying the classifier to the independent 22‑subject dataset yielded AUCs above chance for some RSNs (salience reached values near 0.7) but p‑values did not meet conventional significance thresholds, which the authors attribute to low sample sizes and between‑dataset heterogeneity. Analyses on the combined 38‑subject sample again identified significant predictive accuracy in a subset of networks (notably ESV and EXC) and predominantly for early post‑treatment weeks (for ESV at weeks 0, 1 and 3, and for EXC at week 0). Feature importance patterns overlapped substantially between the original and combined analyses.

Interpretation of findings: The authors interpret these results as evidence that baseline resting‑state FC contains signals predictive of response to psilocybin treatment for depression, with different networks predicting improvement at different latencies. Early improvements (within about five weeks) were most strongly predicted by FC of primary and extrastriate visual networks, characterised by increased long‑range connections between occipital and frontal regions as well as intra‑occipital coupling. Two‑week improvements were associated with higher baseline FC in DMN and executive control networks, while salience network FC predicted sustained improvement at 24 weeks. Relation to prior work and possible mechanisms: Copa and colleagues relate the visual network findings to prior reports that visual FC can predict response to other antidepressant treatments (for example electroconvulsive therapy) and to known sensory and visual alterations produced acutely by psychedelics. They suggest a plausible mediating pathway whereby baseline visual‑frontal coupling predisposes individuals to particular subjective acute experiences (e.g. mystical‑type or emotional breakthrough experiences) that have been linked with positive outcomes after psilocybin. DMN involvement is discussed in the context of rumination; altered DMN activity is implicated in depressive rumination and reductions in DMN connectivity during the acute psychedelic state may relate to therapeutic effects, although the timing here (prediction at two weeks) does not imply a direct acute DMN action. The salience network result is framed as potentially indexing network properties that relate to longer‑term clinical course and may overlap with mechanisms implicated in placebo responses. Limitations highlighted by the authors: The investigators emphasise several limitations. Sample sizes were small, reducing statistical power and limiting the robustness of generalizability tests; classifier p‑values did not survive correction for multiple comparisons in some analyses. Heterogeneity between datasets and lack of standardised acquisition procedures complicates cross‑dataset comparisons. Crucially, both datasets lacked a placebo control arm, so some predictive features could reflect expectancy or placebo effects rather than psilocybin‑specific mechanisms. Short scanning durations in both datasets are noted as an additional constraint. The authors also acknowledge practical limitations of neuroimaging as a predictive tool in real‑world settings, including susceptibility to artefacts, between‑scanner differences and cost. Implications and next steps: The study is presented as a proof‑of‑concept that baseline neurophysiological measures can help predict psilocybin treatment outcome. The authors recommend larger samples, inclusion of placebo or active control conditions, standardised imaging protocols and multimodal approaches combining neuroimaging, behavioural, subjective and genetic markers to improve prediction and to better disentangle placebo from drug‑specific effects. They also propose further investigation of whether baseline FC predicts the acute subjective effects of psychedelics and whether those acute effects mediate longer‑term antidepressant benefit.