Can measuring brain function help us choose an antidepressant?

Knowing which molecules an antidepressant acts on does not, by itself, tell us whether it will help the patient in front of us. Could measuring how that patient's brain functions provide a useful guide?

The SMART-D trial, published in July 2026, tested predictions built from brain imaging, behavior and clinical information by using them to guide medication selection. It did not demonstrate a benefit from choosing a drug according to these markers. But following the attempt helps clarify what still needs to be established before treatment can be reliably tailored to an individual. Original study ↗

From a drug's action to a patient's response

The two antidepressants studied, sertraline and bupropion, have different pharmacological actions. Sertraline primarily inhibits serotonin reuptake. Bupropion inhibits norepinephrine and dopamine reuptake, but not serotonin reuptake. The precise mechanism through which bupropion produces its antidepressant effects, however, remains incompletely understood. Sertraline prescribing information ↗, bupropion prescribing information ↗

Those differences do not mean that a prescription can be determined simply by dividing patients into those with a “serotonin problem” and those with a “dopamine problem.”

Even if the expression of relevant genes could be measured, separate questions would remain: how much of the protein is produced, where it is located in the cell, and how actively it functions. Basic research on the serotonin transporter has shown that transport activity can change with phosphorylation even without a change in the amount at the cell surface. Being present and being functional are not the same thing. Research on transporter regulation ↗

There is a further question: when that function changes, how do neural circuits respond, and how do symptoms change? Acting on the same molecule need not produce the same outcome in brains that are in different states. Understanding individual differences requires looking beyond the molecular target to the circuits and conditions in which it operates.

This is where an approach based on brain function becomes interesting. Before every step from molecules to symptoms has been explained, could the brain's functional state itself help predict treatment response? That is how we read the challenge this study takes on.

Testing past associations in actual treatment decisions

Earlier research offered a lead. In patients who had not improved sufficiently with sertraline and subsequently received bupropion, reward responses and resting-state functional connectivity between the nucleus accumbens and anterior cingulate cortex were associated with treatment response. Earlier study ↗

Building on these findings, SMART-D used prediction models combining fMRI with cognitive tasks, symptoms, employment status and other information. The main analysis included 47 people. Over eight weeks, treatment outcomes did not differ significantly between those assigned a drug consistent with their predicted response and those assigned an inconsistent drug. Marker-based classifications were associated with the course of symptom improvement, but that did not establish their usefulness for choosing between the two medications. SMART-D methods and results ↗

An important feature of this study is that it did not stop at fitting predictions to past data. It tested them in the treatment of new patients. At that stage, the expected benefit of marker-guided selection was not confirmed.

These markers therefore did not provide sufficient evidence to guide prescribing. Where the prediction fell short, however, remains unresolved.

Observing a brain state is not the same as predicting a drug response

fMRI detects changes such as blood oxygenation associated with neural activity. It does not directly measure the amount or activity of particular molecules. Basic research comparing fMRI signals with neural activity ↗

This creates a different difficulty from studying molecules individually. Complex processes can be observed together, but similar measurements may arise from different underlying mechanisms.

Suppose two patients show similar functional connectivity. Whether that state would change in response to the same drug is a separate question. In one patient, it might reflect a feature that influences treatment response; in another, it might merely accompany the symptoms. The present results cannot tell us which explanation applies.

The models may have captured a broader propensity to improve rather than differences related to the actions of particular drugs. Alternatively, associations found in the original population may not have transferred sufficiently to a new one. These are possible explanations, not established causes.

The treatment conditions under which the predictions were developed also matter. The earlier bupropion data came from patients who had not responded sufficiently to sertraline. A feature useful in that group need not have the same meaning for patients now choosing between the two drugs. The question is not only who was measured, but at what stage of treatment the evidence was obtained.

What needs to be measured next?

What follows is our interpretation.

A first priority is to compare the two drugs in patients at the same stage of treatment and establish which choice actually works better for people with a given feature. Simply collecting the characteristics of people who responded to each drug can mix features common to both treatments with those that should change the choice between them.

There may also be value in using changes after treatment starts, rather than relying solely on a single pretreatment measurement. Which functions change early, and do those changes relate to later improvement? Even if the best drug cannot be identified in advance, helping clinicians reach a decision sooner by observing the response could have clinical value. This is not an application demonstrated by the present study; it is a separate question to test.

Adding more measurements should not be treated as progress in itself. Combining molecular, imaging and behavioral data increases the information available. Whether that information changes a treatment decision and leads to better outcomes for patients still has to be established.

Nor must every aspect of the mechanism be explained before a prediction can be useful. Even without a complete explanation, a prediction that repeatedly proves useful in other patient populations could have a place in clinical care.

The markers developed from brain function and other information did not, in this trial, improve selection between the two drugs. Yet testing past associations in actual treatment made the next question more concrete.

Which differences in treatment response do the measured differences between patients represent? Only when that link is established can measuring the brain help us choose a medication suited to the patient.