Dopamine neurons drive basal forebrain reward prediction in milliseconds

TL;DR: Two subcortical systems — midbrain dopamine neurons and a population of "bursting" neurons in the basal forebrain — have long been known to carry nearly identical reward-prediction signals, yet they were studied as separate machines. This paper shows they are not separate: dopamine neurons causally drive the basal forebrain cells, and they do it in about 10 milliseconds, far too fast for classical dopamine receptors. The likely wire is glutamate co-released from dopamine axons. It's a rare case where simultaneous recording, bidirectional optogenetics, pharmacology, and anatomical tracing all point the same way.

Two systems computing the same thing — why is that suspicious?

Start with the puzzle. When an animal sees a cue that predicts reward, dopamine (DA) neurons in the ventral tegmental area (VTA) fire a brief phasic burst whose size scales with expected value and shrinks as rewards become predictable. This is the textbook reward-prediction-error signal that drives reinforcement learning in the striatum.

Separately, a population of non-cholinergic neurons in the basal forebrain (BF) — the "BF bursting neurons" — do almost exactly the same thing. Same scaling with value, same learning-dependent changes. But these cells project to cortex, and their job is thought to be gain modulation: cortical activation, decision speed, attention.

So we have two populations, in different places, wired to different downstream targets, assigned different jobs — carrying what looks like the same signal. Three explanations are possible, and distinguishing them matters for how we think reward information is organized in the brain:

1. Independent DA BF

2. Common input up DA BF

3. Causal chain DA BF ~10 ms This paper's answer

The authors make a sharp methodological point: you cannot disprove the "common input" hypothesis by elimination, because there are too many candidate upstream sources. So instead of chasing that, they test the positive prediction of the causal hypothesis — that manipulating DA should move BF — and lean on timing to establish direction.

Step 1: the two populations move together, trial by trial

They recorded VTA and BF simultaneously in freely-moving rats doing a lick-schedule task with high-value (short-delay) and low-value (long-delay) reward cues. Both populations showed the expected value scaling: bigger response to the high-value cue, and — the signature of expectation — a smaller response to the reward itself when it was better predicted.

The key move is going beyond trial averages. On individual trials, when the DA population responded more strongly, so did the BF population. The trial-by-trial Pearson correlation was positive and significant at trial start, cue, and reward — and critically, it was selective: pairing DA cells with other BF neurons, or other VTA neurons with BF bursting cells, gave weaker correlations. So the coupling is between these two specific populations, not a general arousal tide lifting all boats.

This already kills the "fully independent" hypothesis. Shared moment-to-moment fluctuations mean shared information. But it doesn't tell you direction or mechanism.

Step 2: direction from millisecond timing

Now the elegant part. They computed spike-time cross-correlograms at 1-ms resolution across the whole session — every spike, not just task-evoked bursts. DA–BF bursting pairs showed a sharp correlation peak, and that peak sat at a 10-ms DA-leading lag. DA fires; ~10 ms later BF fires. This asymmetry held in all five rats.

Two controls make this convincing. First, restricting to inter-trial intervals — periods with no cue, no reward, no overt task event — preserved both the correlation and the 10-ms offset. So this isn't two systems independently locking onto the same external event; the coupling is intrinsic and ongoing. Second, a noise-correlation (jPSTH) analysis that explicitly subtracts the mean event-locked response still left a DA-leading residual peaking at one 10-ms bin. The coupling survives after you remove everything the stimulus could explain.

And it persists through learning: across acquisition of a new cue, extinction, and contingency degradation, the amplitude coupling held and the cross-correlation kept an 8-ms DA-leading peak, stable even within the two halves of a single session while predictions were actively being rewritten.

A ~10-ms consistent lead is exactly what you'd expect from a synaptic pathway from DA to BF. But a correlation, however clean, is still a correlation.

Step 3: cut the wire both ways

Here the paper switches to DAT-Cre mice for cell-type-specific optogenetics, and this is the section most worth your time (Figures 4 and 5).

Sufficiency. Express ChR2 in VTA DA neurons, then flash them outside the task — no cue, no reward — and record BF. Brief 10-ms light pulses to DA somata drove time-locked spiking in a subset of BF neurons at a mean latency of 11.4 ms, matching the natural 10-ms lead almost exactly. Then they stimulated the DA axon terminals inside the BF directly, bypassing the soma-to-terminal conduction delay — latency dropped to 4.2 ms. That shortening is the fingerprint of a local, monosynaptic connection: the DA axons in the BF are themselves the relay, not some detour through another DA target.

Necessity. Express halorhodopsin, silence DA on a random half of trials. Inhibition didn't just blunt the cue-evoked BF burst — it nearly abolished ongoing BF firing on catch trials (no cue at all), with an estimated onset of 9.0 ms. Terminal inhibition reproduced this at 8.5 ms. The ongoing-activity result is the surprising one: DA isn't merely forwarding phasic reward events, it's supplying continuous excitatory drive to these BF cells.

Across all four manipulations the same population was targeted. The affected BF neurons were enriched for bursting cells, and the latencies cluster tightly:

The selectivity for bursting neurons is real but partial. Among BF bursting neurons, 56.4% were directly excited by somatic DA activation versus 20.5% of other BF neurons (p=0.016); for somatic inhibition, 45.1% vs 14.5% (p=0.031). The effect is enriched, not total — a point to keep in mind.

Step 4: not the dopamine receptors you think

Now the mechanism twist. Canonical dopamine signaling runs through D1/D2 receptors — G-protein-coupled, second-messenger cascades, operating on hundreds of milliseconds to seconds. A 10-ms effect simply cannot be that.

So they blocked D1 and D2 receptors systemically (SCH23390 + raclopride, at doses known to alter striatal firing — and confirmed here to alter BF LFP and locomotion). Fast DA-evoked BF recruitment survived: 69 of 77 responsive neurons kept their responses for a full 60 minutes post-blockade, indistinguishable from PBS controls (65/70). Whatever carries the signal, it isn't the metabotropic dopamine receptors.

The candidate they land on is glutamate co-release. A subset of VTA DA neurons express VGLUT2 and release glutamate as a fast ionotropic co-transmitter, well documented in the ventral striatum. To look for the anatomical substrate in the BF, they used ATLAS — an anterograde transsynaptic tracer that labels postsynaptic neurons specifically across GluA1-containing (glutamatergic) synapses. Expressing ATLAS in VTA DA neurons labeled BF neurons downstream, and 97.7% of the labeled cells were ChAT-negative (non-cholinergic) — the right cell class. So DA axons form glutamatergic synapses onto predominantly non-cholinergic BF neurons.

blocking D1/D2 leaves the fast path intact

What to believe, and what to hold loosely

The causal core is strong. Bidirectional optogenetics with matching latencies, the soma-vs-terminal latency drop pinning the connection to local DA axons, and the ongoing-activity suppression together make a convincing case that VTA DA neurons directly and rapidly drive a population of BF bursting neurons. This genuinely reframes two "parallel" reward systems as, in part, a serial one — and shows a broadcast nucleus (DA) recruiting a second broadcast nucleus (BF→cortex) with millisecond precision. If it holds, the same predictive event could simultaneously feed striatal learning and cortical gain control through one upstream source.

Where to stay skeptical:

• "Putative" DA neurons in rats are functionally defined, not opto-tagged. The rat correlation data rest on identifying DA cells by baseline rate and phasic cue response. The mouse optogenetics uses genetically defined DA neurons, so the causal chain is solid there, but the beautiful 10-ms natural-coupling number inherits the identification caveat.

• Glutamate co-release is a candidate, not a demonstration. The receptor blockade rules out canonical D1/D2 signaling, and the ATLAS tracing shows glutamatergic synapses exist. But no one here knocked out VGLUT2 in DA neurons to show the fast BF response disappears. The authors say as much. Other fast substrates (e.g., gap junctions, a non-canonical ionotropic dopamine action) aren't formally excluded, though glutamate is the parsimonious bet.

• The affected population is a substantial subset, not all. Only ~30–37% of BF neurons responded to somatic manipulation; terminal manipulations reached fewer. This could be incomplete viral/optical coverage, or genuine heterogeneity among BF bursting cells. Either way, "DA drives BF bursting" is true of an enriched subset, not uniformly.

• "Common input" is softened, not slain. The authors are careful: their data show a causal route exists; they do not claim it's the only source of BF reward-prediction activity. Both routes may coexist.

The most efficient read is Figures 4–5 for the causal optogenetics and Figure 6 for the receptor-blockade-plus-tracing mechanism; Figure 2 is worth a look for how cleanly the 10-ms lead falls out of the cross-correlation, including during inter-trial intervals. The headline worth carrying away: a reward-prediction signal computed in one subcortical broadcast system can be injected into another on a synaptic timescale, through dopamine neurons' fast glutamatergic side-channel rather than dopamine itself.