Dopamine and serotonin in human substantia nigra track social context and value signals during economic exchange

The study investigates how fast dopamine and serotonin signaling in the human brain contributes to social decision-making. Prior human research relied largely on non-invasive imaging with limited temporal resolution, leaving neuromodulator dynamics during social interactions unclear. Pharmacological manipulations implicate serotonin and dopamine in social behaviors (for example, serotonin modulating reactions to unfairness in the ultimatum game and dopamine affecting harm aversion), and animal studies show subsecond neuromodulator fluctuations encode value-related signals such as reward prediction errors. The authors hypothesized that in humans, during social exchange, dopamine and serotonin would exhibit distinct fast signals reflecting social context and value. They leveraged intraoperative human electrochemistry to measure subsecond fluctuations in substantia nigra pars reticulata (SNr) while patients played the ultimatum game against human versus computer proposers to test whether overall neuromodulator levels reflect social context and whether trial-by-trial relative changes encode value statistics.

- Non-invasive human neuroimaging has mapped a social brain network but cannot resolve fast chemical dynamics, leaving gaps in understanding neuromodulator roles at subsecond timescales. - Pharmacology in humans: lowering serotonin via acute tryptophan depletion increases rejection of unfair ultimatum offers; SSRIs reduce such rejections; levodopa can reduce harm aversion, indicating causal roles of serotonin and dopamine in social decisions. - Animal studies: fast dopamine transients encode reward prediction errors for learning; this motif extends to social prediction errors in mice; serotonin neurons signal reward and punishment across timescales and may encode aspects of reward processing and learning rate. - Human electrochemistry during DBS surgery has shown subsecond dopamine and serotonin in striatum encode RPEs and sensory/action signals in non-social tasks. However, whether these roles are modulated by social context and how they manifest in human social interaction remained unknown.

Participants: Four Parkinson’s disease patients (one female; mean age ± s.e.m. 71.3 ± 3.4 years) undergoing bilateral deep brain stimulation (DBS) surgery participated. Each performed the task in two surgical sessions (14–28 days apart), yielding eight datasets (4 patients × 2 sessions). Parkinson’s medications were withheld during surgery. Task: One-shot ultimatum game with 60 trials per session: 30 human-proposer trials and 30 computer-proposer trials, blocked within session with counterbalanced order across sessions. Each trial presented a unique human avatar (image and name) or a fixed computer avatar as proposer. Offers were preprogrammed between US$1–9 from a US$20 stake; the same randomized set of offers was used in both conditions within a session. Participants decided to accept or reject each offer. On ~one-third of trials, participants rated mood on a visual analogue scale (negative to positive). Feedback screens followed choice. Electrochemistry and recording site: A carbon-fibre electrode was temporarily inserted into substantia nigra pars reticulata (SNr) along the clinical guide cannula. Fast-scan cyclic voltammetry (FSCV) was used with a triangular waveform (−0.6 V to +1.4 V at 400 V/s and back) applied at 10 Hz during the task (base 100 kHz sampling). The approach yields ~10 samples per second. Conditioning at 97 Hz preceded task recording. Signal prediction model: An ensemble of deep convolutional neural networks (modified InceptionTime architecture with two residual blocks and multi-kernel convolutional layers) was trained on large in vitro datasets to predict concentrations of dopamine, serotonin, norepinephrine, and pH from differentiated current sweeps. Training data: 64 in vitro datasets with mono- and mixed-analyte solutions across concentrations (0–2,500 nM) and pH 7.0–7.8; 59 datasets for training and 5 held out for testing (7,260 concentration combinations; >1 million sweeps). Models used Adam optimization, MSE loss, learning rate scheduling, and ensemble averaging of five submodels. Model evaluation on held-out in vitro data demonstrated high sensitivity and specificity. Neuromodulator estimates: Overall estimates were computed as the sum of samples within 1 s after offer presentation. Relative estimates subtracted the sample at offer onset (local baseline) and then summed the 1 s window. Behavioral and neural analysis: Trials with RT >14 s were excluded (~5%). Mixed-effects models at the trial level were used: - Choice (logistic): C-M1 included value, condition (human vs computer), and their interaction; C-M2 additionally included value difference (current − previous offer). - Reaction time (linear): RT-M1 included choice, value, condition, and their interactions. - Emotion ratings (linear): E-M1 included choice, value, condition, and their interactions. - Neural models (linear): N-M1 predicted overall dopamine/serotonin from choice, condition; N-M2 added within-session order; N-M3 predicted relative changes from value and value difference; N-M3* orthogonalized value difference against value; N-M4 added absolute value difference; N-M5 included condition interactions. Binary variables were contrast coded (−0.5/0.5); continuous predictors were z-scored within dataset. Model comparison used AIC. Analyses were performed in MATLAB (fitglme).

Behavioral: - Choice depended positively on offer value (C-M1: t(454)=2.43, P=0.016, β=1.77 ± 1.43) and negatively on human versus computer condition (participants accepted fewer human offers; t(454)=−2.23, P=0.026, β=−3.60 ± 3.18); no value×condition interaction. - Adding value difference (C-M2) did not yield significant history effects (value difference: t(423)=0.38, P=0.707) though model fit improved (AIC 2843 vs 3016). - Reaction times were longer for reject than accept decisions (RT-M1 choice: t(450)=−2.56, P=0.011); no significant effects of value or condition on RTs. - Emotion ratings showed no significant effects (all |t(147)|<1.70, P>0.093). Neural – Overall levels (1 s post-offer): - Dopamine showed higher overall levels in the human versus computer condition (N-M1 condition: t(454)=3.04, P=0.002, β=0.85 ± 0.55); no choice effect and no interaction. Serotonin showed no effects (condition: t(454)=−0.01, P=0.993). - Controlling for within-session order (N-M2) replicated the dopamine condition effect (t(450)=2.98, P=0.003) with no order-related effects; serotonin remained non-significant. Models including order fit worse by AIC. - The dopamine context effect unfolded gradually over time across trials. Neural – Relative changes (baseline-subtracted 1 s post-offer): - Dopamine tracked value difference (current − previous offer), consistent with RPE-like signaling: value difference effect (N-M3: t(426)=2.82, P=0.005, β=0.21 ± 0.14); no effect of current value (t(426)=−1.47, P=0.144). Orthogonalized analysis (N-M3*) replicated the value difference effect. - Serotonin tracked current offer value: value effect (N-M3: t(426)=3.15, P=0.002, β=0.22 ± 0.14); no significant value difference effect (t(426)=−1.82, P=0.070). N-M3* replicated the value effect. - Adding absolute value difference (N-M4) did not yield significant effects for dopamine or serotonin and worsened model fit (AIC: dopamine 1243 vs 1231; serotonin 1244 vs 1231). - Social context did not modulate these relative value-related effects (N-M5: no significant condition interactions for dopamine or serotonin; models fit worse than N-M3). Overall: Participants rejected more human than computer offers despite identical offer sets; this social-context effect was mirrored by higher overall dopamine (not serotonin). Relative dopamine encoded trial-by-trial value changes (RPE-like), while relative serotonin encoded current value, both independent of social context.

The study demonstrates that subsecond neuromodulator dynamics in human SNr relate to both social context and value computations during economic exchange. Behaviorally, participants exhibited stronger rejection of offers believed to be from humans, consistent with fairness norms and social context effects. Neurochemically, overall dopamine levels were elevated in the human condition, suggesting dopamine contributes to setting a social context or motivational frame in human interactions, without directly predicting single-trial choices. In contrast, serotonin overall levels did not vary with social context. On a trial-by-trial basis, baseline-relative dopamine increased with better-than-previous offers and decreased with worse-than-previous offers, consistent with reward prediction error signaling extended to social offers. Serotonin, by contrast, tracked the immediate value of the current offer, supporting a complementary role in evaluating present value rather than comparative changes. These roles were stable across human versus computer contexts, indicating generalized value coding. The findings fit with basal ganglia models and prior animal and human electrochemistry work, and suggest that dopamine and serotonin provide complementary computations—comparison to recent past versus evaluation of current state—that can support norm-based social behavior. The authors discuss SNr connectivity, noting inputs from SNc dopamine and raphe serotonin neurons and outputs to thalamocortical circuits involved in decision-making, and consider whether observed signals reflect broader broadcast neuromodulator activity versus SNr-specific processing. They argue that despite the Parkinson’s disease sample and surgical context, several factors (withheld medication, within-subject design, absence of major comorbidities, consistency with prior work) support generalizability to healthy neuromodulatory function.

This work provides direct human evidence that fast dopamine and serotonin fluctuations in SNr encode distinct but complementary signals during social exchange: overall dopamine reflects social context (human versus computer), relative dopamine encodes trial-by-trial value changes (RPE-like), and relative serotonin encodes current offer value. These findings extend neuromodulatory theories of value coding to human social decision-making and suggest a mechanism by which dopamine may help set a social interaction frame while dopamine and serotonin jointly support value-based computations. Future research should: (1) examine repeated and strategic social interactions (multi-round games) involving richer inference; (2) record across multiple brain regions to assess regional specificity and network dynamics; and (3) expand to larger and more diverse participant samples, including non-clinical populations, potentially leveraging depth electrodes in epilepsy monitoring.

- Small sample size (n=4 patients; 8 datasets) limits statistical power and generalizability. - Clinical population (Parkinson’s disease undergoing DBS) and intraoperative setting may not fully represent healthy brain function, though medication was withheld and prior work suggests comparable neuromodulator dynamics. - Recording site constrained to SNr; signals may reflect broader neuromodulator broadcast rather than SNr-specific processing; single-site recordings preclude network-level conclusions. - One-shot, blocked ultimatum game design limits analysis of strategic adaptation over longer histories; absence of multi-round interactions may underestimate complex social inference effects. - Analyses focused on a 1 s post-offer window (though robustness to other windows was checked); temporal dynamics beyond this window were not extensively characterized. - Human versus computer context manipulation, while counterbalanced, may involve unmeasured factors (e.g., arousal, attention) not fully disentangled from social context.