Giovanni Rabuffo, Irene Acero-Pousa, Tomas Berjaga-Buisan, Gustavo Deco
7 min
Brain stimulation is a powerful tool for modulating neural activity, yet its effects are notoriously variable across trials and individuals. This variability limits the clinical and research utility of stimulation protocols. The authors investigate the sources of this variability by developing personalized brain models for 100 participants using resting-state fMRI data. By virtually perturbing these models, the researchers map how stimulation responses depend on the target site, the timing of the pulse relative to the brain's ongoing state, and the use of multi-site (bifocal) stimulation.
The study reveals that the brain's response to stimulation is not random but follows a lawful, state-dependent organization. First, the magnitude of the response increases along the cortical hierarchy, from unimodal sensory regions to transmodal association areas. Second, the brain's instantaneous global energy—the total ongoing activity—acts as a gate for stimulation: lower-energy states consistently produce larger and more reproducible responses. This effect is a nonlinear property of the brain's dynamics, as it is absent in linear control models. Finally, the authors show that bifocal stimulation—targeting two regions simultaneously—can be used to amplify the response magnitude, particularly when both sites belong to the same functional network, and to improve the reliability of the stimulation effect.
These results suggest that much of the trial-to-trial variability in brain stimulation is not noise but a predictable consequence of the brain's current state. By identifying where, when, and how to stimulate, this framework provides a principled basis for designing closed-loop and circuit-level protocols. Such approaches could significantly enhance the reproducibility of neurostimulation, potentially improving outcomes for clinical applications in neurology and psychiatry.
Brain stimulation has transformed the treatment of several neurological and psychiatric disorders and is now widely used to read out causal interactions in the human brain. However, its effects vary from one trial to the next, even when the stimulation parameters are kept identical. Identifying the sources of this variability and developing new strategies to reduce it are key to achieving more controllable interventions. Yet, exhaustive testing in real experiments is unfeasible, because each participant can only be probed at a handful of sites under conditions the experimenter cannot fully control. Here we use personalized brain models to map how stimulation responses vary across target sites and with the brain's ongoing state. We show that stimulation responses are jointly determined by the target's position along the cortical unimodal-to-transmodal hierarchy and the brain's global ongoing activity, with lower-activity states yielding larger responses. Accordingly, timing stimulation to low-activity periods reduces trial-to-trial variability in response magnitude. Even without state information, joint stimulation of specific region pairs reduces variability compared with stimulating either region alone. Together, these findings identify where, when, and how to stimulate the brain to achieve more reproducible responses, with concrete predictions for closed-loop and circuit-level stimulation protocols.
Alex: [measured] Yes. They found that restricting stimulation to the lowest five percent of baseline energy states produced the largest, most reproducible effects. It’s a clear argument for closed-loop, state-dependent protocols.
Sam: [reflective] That’s the real potential. If a clinician can use this to optimize TMS, it moves us away from fixed-site protocols toward something much more precise.
Alex: [quiet confidence] That is the goal. For the full breakdown of their methodology and the specific network co-stimulation findings, check out the paper linked in the show notes.
Sam: [brightly] Definitely worth a deep dive. Thanks for walking me through it.
Alex: [steady, analytical] The brain’s responsiveness to stimulation is a dynamic variable gated by its instantaneous global energy level. This comes from a recent preprint using personalized neural network models of resting-state fMRI.
Sam: [curious, leaning in] So, if the brain's baseline activity is the gatekeeper, have we been misinterpreting stimulation data by treating it as a fixed input?
Alex: [nodding] Exactly. The global effect size of a perturbation is inversely proportional to pre-stimulus energy. Think of the brain as a guitar string: if it’s already vibrating wildly, plucking it produces a smaller, muddier sound than if it were still.
Sam: [thoughtful, processing] That’s a helpful analogy. How did they isolate that relationship without thousands of empirical trials?
Alex: [measured, explaining] They trained personalized neural networks on Human Connectome Project data, allowing them to perform in silico bifocal stimulation across 80,000 region pairs—an exhaustive search impossible in any clinical setting.
Sam: [probing] Okay, so they built a surrogate brain. Does that hold up against real-world neurostimulation?
Alex: [even pace, precise] They validated the model by showing it could reconstruct task-evoked activation patterns from resting-state data alone. It classified motor tasks with 91 percent accuracy, outperforming standard connectivity metrics.
Sam: [skeptical] Why is the whole-brain energy state more predictive than the local activity at the stimulation site?
Alex: [deliberate] The brain’s ongoing activity acts as a gain modulator. The model learns a nonlinear mapping where low-energy states provide a cleaner baseline, allowing the perturbation to propagate more effectively through the learned dynamics.
Sam: [thoughtful] So, could we theoretically time stimulation to hit these low-energy windows?
Alex: [measured] Yes. They found that restricting stimulation to the lowest five percent of baseline energy states produced the largest, most reproducible effects. It’s a clear argument for closed-loop, state-dependent protocols.
Sam: [reflective] That’s the real potential. If a clinician can use this to optimize TMS, it moves us away from fixed-site protocols toward something much more precise.
Alex: [quiet confidence] That is the goal. For the full breakdown of their methodology and the specific network co-stimulation findings, check out the paper linked in the show notes.
Sam: [brightly] Definitely worth a deep dive. Thanks for walking me through it.
Alex: [steady, analytical] The brain’s responsiveness to stimulation is a dynamic variable gated by its instantaneous global energy level. This comes from a recent preprint using personalized neural network models of resting-state fMRI.
Sam: [curious, leaning in] So, have we been misinterpreting stimulation data by treating it as a fixed input?
Alex: [nodding] Exactly. The global effect size of a perturbation is inversely proportional to pre-stimulus energy. Think of the brain as a guitar string: if it’s already vibrating wildly, plucking it produces a smaller, muddier sound than if it were still.
Sam: [thoughtful, processing] That’s a helpful analogy. How did they isolate that relationship without thousands of empirical trials?
Alex: [measured, explaining] They trained personalized neural networks on Human Connectome Project data, allowing them to perform in silico stimulation across 80,000 region pairs—an exhaustive search impossible in any clinical setting.
Sam: [probing] Okay, so they built a surrogate brain. Does that hold up against real-world neurostimulation?
Alex: [even pace, precise] They validated the model by showing it could reconstruct task-evoked activation patterns from resting-state data alone. It classified motor tasks with 91 percent accuracy, outperforming standard connectivity metrics.
Sam: [skeptical] Why is the whole-brain energy state more predictive than local activity?
Alex: [deliberate] The brain’s ongoing activity acts as a gain modulator. The model learns a nonlinear mapping where low-energy states provide a cleaner baseline, allowing the perturbation to propagate more effectively.
Sam: [thoughtful] So, could we theoretically time stimulation to hit these low-energy windows?
Alex: [measured] Yes. They found that restricting stimulation to the lowest five percent of baseline energy states produced the largest, most reproducible effects. It’s a clear argument for closed-loop, state-dependent protocols.
Sam: [reflective] That’s the real potential.