The Cellular and Synaptic Actions of Dopamine During Behavior

TL;DR: In a technical tour de force — patch-clamp recordings from >300 identified striatal neurons in awake mice, with simultaneous readout and optogenetic control of dopamine — the authors find that physiological dopamine transients do almost nothing to a striatal neuron's moment-to-moment voltage or excitability. What dopamine actually does is gate synaptic plasticity during learning, and it does so selectively: learning strengthens cortex→striatum synapses onto both cell types, but only the direct-pathway (D1) synapses need dopamine to change. This is a direct in-vivo challenge to the textbook "dopamine rapidly tunes striatal activity" model, and it maps cleanly onto a three-factor, credit-assignment view of dopamine rather than a real-time control view.

Two jobs dopamine was supposed to have

The striatum is the input stage of the basal ganglia, and its principal cells — spiny projection neurons (SPNs) — split into two populations: D1-expressing SPNs (the "direct/Go" pathway) and D2-expressing SPNs (the "indirect/NoGo" pathway). Dopamine, released by midbrain neurons, is the neuromodulator that everyone agrees is essential here: lose it and you get Parkinson's; dysregulate it and you get addiction.

Decades of slice physiology built a two-part model of how dopamine acts. First, it rapidly tunes intrinsic excitability — a dopamine pulse makes D1 neurons more excitable, biasing which SPNs fire right now. Second, it enables synaptic plasticity — dopamine paired with cortical input strengthens or weakens corticostriatal synapses, the presumed substrate of reinforcement learning.

The problem: nearly all of that mechanistic evidence comes from acute brain slices, where you apply dopamine or drugs to tissue that has been cut off from its normal, continuously fluctuating dopamine supply. Does the same machinery actually get engaged by the real, sub-second dopamine transients that occur during behavior? Nobody had measured it, because doing so requires recording the membrane potential of a genetically identified SPN, in an awake behaving animal, while independently controlling both its cortical input and dopamine release. That is the experiment this paper pulls off.

The preparation: three independent knobs

The elegance is in the wiring. In a single triple-transgenic mouse they got three orthogonal handles:

• Cortical input: the red opsin ChrimsonR in layer-5 neurons of somatosensory cortex (S1), so a light pulse fires the corticostriatal axons feeding the striatum.
• Cell-type ID: ChR2 in D2 SPNs, so a blue pulse through the patch pipette ("optopatch") tags a recorded cell as D2; unresponsive connected cells are putative D1.
• Dopamine: the opsin ChRmine in midbrain (SNc) dopamine neurons, calibrated with a fluorescent dopamine sensor (GRAB-DA) so that the evoked transient matches the size of a natural reward response — this calibration matters, because it means they are testing physiological, not supraphysiological, dopamine.

So they can whisper a defined cortical input into an identified D1 or D2 neuron, release a reward-sized puff of dopamine at a chosen moment, and read the postsynaptic voltage directly.

Result 1: physiological dopamine does almost nothing acutely

Fire a reward-sized dopamine transient right at a corticostriatal synapse and the EPSP (excitatory postsynaptic potential) barely moves: a small but significant bump in D1 neurons, nothing in D2, and no change in paired-pulse ratio (arguing the small D1 effect is postsynaptic, not a change in release). Inject a current ramp and measure rheobase — the minimum current to fire a spike — and dopamine changes it in neither cell type. Drive dopamine during a sustained depolarization and firing rate is unchanged, during and for tens of seconds after. And when they simply recorded membrane potential while dopamine fluctuated spontaneously, the two were uncorrelated.

Here is the move that makes the null credible. They took the same rig — same pipettes, same protocol — into acute slices and got the classic textbook result: dopamine robustly and durably boosts D1-SPN firing, no effect on D2. So the machinery exists and their setup can detect it. It just isn't engaged by physiological dopamine in the living brain.

Their explanation is occlusion: in the intact brain, tonic midbrain firing keeps D1 receptors and downstream cAMP/PKA signaling continuously engaged, holding the neuron near a dopamine-set excitability ceiling. A slice, cut off from dopamine for an hour, has fallen off that ceiling — so bath-applied dopamine "restores" a state that in vivo was never lost. The dramatic slice effect is real but it's measuring the gap created by deprivation, not a signal the brain uses moment to moment.

Result 2: learning strengthens both pathways, but only D1 needs dopamine

Now the constructive half. They built a behavioral task out of the very same S1→striatum synapse they can record. Mice learn that an optogenetic pulse to S1 predicts water; they lick, they get reward. Over ~10 days they cross 80% hits with under 40% false alarms. Fiber photometry shows the reward-dopamine response migrating backward in time onto the cue — textbook reward-prediction-error transfer, confirmed by omitting reward and seeing the cue response persist. So the task naturally aligns dopamine release with the cortical input, exactly the pairing plasticity is supposed to need.

Learning strengthened the corticostriatal synapse in both D1 and D2 SPNs: larger early EPSPs on hit trials in experts, larger responses to 10 Hz S1 trains even with the reward spout removed, a drop in paired-pulse ratio, and — in slices from trained mice — a lower AMPA/NMDA ratio, a classic plasticity signature. And it matters: deleting the NMDA receptor subunit NR1 locally in the striatum impaired learning without breaking locomotion, tying task acquisition to plasticity in this exact circuit.

The clean dissociation comes from a four-group experiment (unstimulated / repeated S1 stimulation / S1 + dopamine neuron activation / S1 + D1&D2 antagonists):

✓ ✓ ✗ –

✓ ✓ ✓ –

only D1 loses plasticity when DA is blocked

Two things stand out. Extra dopamine (co-stimulating SNc) did not add anything beyond S1 stimulation alone — the endogenous dopamine recruited by cortical activity already saturates the D1 plasticity mechanism. And D2 plasticity survived complete dopamine blockade, implying it runs on a different substrate (the authors point to adenosine A2A signaling and possibly cholinergic interneurons).

Why an ML reader should care

This is a physiological vote for the three-factor learning rule picture and against dopamine as a real-time control input. In RL terms: dopamine is the term in the weight update, not a term in the policy's forward pass. The synaptic change requires pre-synaptic activity × post-synaptic activity × dopamine — exactly the eligibility-trace-gated, reward-modulated Hebbian form that actor–critic models assume. The result that this gating is asymmetric — mandatory for D1/direct/"Go" learning, dispensable for D2/indirect/"NoGo" learning — is a concrete biological constraint that the usual symmetric two-actor caricatures of the basal ganglia don't capture, and it's the kind of detail worth building into models.

It also reframes the "dopamine sets vigor/policy in real time" debate. If reward-sized transients don't budge membrane potential, then the fast behavioral effects people attribute to dopamine may be indirect, and the huge structural changes seen in addiction likely come from dopamine pushed outside its physiological range engaging different mechanisms — consistent with their saturation finding.

What to be skeptical about

The honest weakness is the one the selection flagged: this is a paper of nulls and directions, and the main text reports strikingly few effect sizes, sample sizes, or confidence intervals. "No detectable change," "small but significant," "robustly strengthened" — the quantitative backbone is in supplementary tables not included here. For the acute-manipulation nulls, the whole argument rests on the claim being adequately powered, and in-vivo whole-cell yields are low, so I'd want the per-group n and effect-size bounds before treating "dopamine does nothing acutely" as settled rather than "we couldn't detect an effect." The slice positive control is the strongest thing propping up the null — it shows sensitivity — but sensitivity in a slice with 20+ minute stable recordings doesn't fully transfer to noisier, shorter in-vivo epochs.

Two more caveats. Putative D1 cells are defined by absence of an optogenetic response, so that population is inferred, not tagged. And the whole story is one pathway (S1→dorsolateral striatum) in a sensory-cued licking task; whether the D1/D2 dopamine asymmetry generalizes to associative or ventral striatum is open.

What to read

If you read one thing, read the Result 1 excitability section plus its slice control (Fig. 2) — that in-vivo-null-against-slice-positive contrast is the paper's rhetorical and scientific hinge. Then jump to Fig. 5, the four-group plasticity dissociation, which carries the constructive claim. The Discussion's occlusion argument and its extension to hyperdopaminergic/addiction states is the most speculative and the most interesting to argue with.