Shih-Chieh Lin

bsky.appbioRxiv Neuroscience (@biorxiv-neursci.bsky.social)A dynamical attractor computes reward prediction in the basal forebrain https://www.biorxiv.org/content/10.64898/2026.10.01.755868v1

Our 2nd preprint this week tackles a puzzle in reward prediction: BF bursts are both stable and flexible. Across conditions, neurons preserve relative response amplitudes, yet population burst amplitude scales with reward prediction. How can a circuit do both? bsky.app/profile/bior...

bsky.appbioRxiv Neuroscience (@biorxiv-neursci.bsky.social)A dynamical attractor computes reward prediction in the basal forebrain https://www.biorxiv.org/content/10.64898/2026.10.01.755868v1

biorxiv.orgDopamine neurons drive basal forebrain reward prediction in millisecondsPhasic reward prediction signals in ventral tegmental area (VTA) dopamine (DA) neurons and basal forebrain (BF) noncholinergic bursting neurons are remarkably similar despite the distinct circuit…

Finally, what circuit closes the recurrent loop? Our 1st preprint nominates VTA dopamine neurons: DA drives BF bursting within ~10 ms, while BF Npr3 neurons project to VTA. Together, the studies suggest recurrent BF↔VTA interactions may compute reward prediction. bsky.app/profile/scli...

Shih-Chieh Lin@sclin-lab.bsky.social

New preprint from our lab ⚡ Why do DA and BF bursting neurons carry such similar reward-prediction signals? Are they independent, driven by common input, or causally linked? We find a direct answer: DA activity rapidly and causally drives BF. www.biorxiv.org/content/10.6... 1/10

This study was led by my super-talented graduate student @hsiaochenliu.bsky.social. We’re very excited about this new framework for understanding phasic reward-prediction signals through recurrent circuit dynamics. Comments and suggestions are very welcome!

bsky.appbioRxiv Neuroscience (@biorxiv-neursci.bsky.social)A dynamical attractor computes reward prediction in the basal forebrain https://www.biorxiv.org/content/10.64898/2026.10.01.755868v1

Linear summation extended to behavior. During auditory detection, sound input and Npr3 stimulation also summed approximately linearly in BF burst amplitude, which encodes reward prediction. Larger cue bursts improved detection and were followed by smaller reward responses.

This solves a transformation problem: heterogeneous, distributed inputs can be converted into a common BF output. The attractor supplies a shared response format, while linear summation provides a simple algorithm for pooling sparse inputs into reward-prediction strength.

bsky.appbioRxiv Neuroscience (@biorxiv-neursci.bsky.social)A dynamical attractor computes reward prediction in the basal forebrain https://www.biorxiv.org/content/10.64898/2026.10.01.755868v1

Timing pointed to recurrent circuitry beyond the BF. Ipsilateral and contralateral bursts began after nearly identical delays, ~29 and ~28 ms. The directly stimulated side was not recruited first, arguing against simple local propagation.

How is burst amplitude determined? When left and right Npr3 inputs were activated together, their effects on population burst amplitude summed approximately linearly. The attractor fixes burst structure, while amplitude integrates linearly across concurrent inputs.

bsky.appbioRxiv Neuroscience (@biorxiv-neursci.bsky.social)A dynamical attractor computes reward prediction in the basal forebrain https://www.biorxiv.org/content/10.64898/2026.10.01.755868v1

Different Npr3 stimulation patterns produced BF bursts with similar duration, even as burst amplitude changed with input strength. Once the burst was underway, later pulses became ineffective, revealing nonlinear, state-dependent sensitivity.

The critical test: can partial access trigger the full burst? Unilateral Npr3 stimulation accessed only part of the system, yet stimulation from either hemisphere recruited the same structured burst across the bilateral BF ensemble.

bsky.appbioRxiv Neuroscience (@biorxiv-neursci.bsky.social)A dynamical attractor computes reward prediction in the basal forebrain https://www.biorxiv.org/content/10.64898/2026.10.01.755868v1

Our answer: a dynamical attractor. Inputs need only trigger the circuit, not impose the burst pattern. Recurrent dynamics generate the shared burst structure, while burst amplitude remains flexible and varies with input strength.

BF Npr3 neurons gave us an experimental entry point. Brief activation outside behavior reproduced natural cue-evoked bursting, including its temporal envelope, cell-to-cell amplitude relationships, and low-dimensional population trajectory.

bsky.appbioRxiv Neuroscience (@biorxiv-neursci.bsky.social)A dynamical attractor computes reward prediction in the basal forebrain https://www.biorxiv.org/content/10.64898/2026.10.01.755868v1

Phasic bursts are usually assumed to be feedforward: inputs directly impose the population response. But this creates an access problem. To preserve burst structure, every effective input would need to reach essentially the full BF ensemble with the right relative weights.

Congratulations to @deisseroth.bsky.social !!!

Nature@nature.com

BREAKING: The Nobel Prize in Physiology or Medicine has been awarded to Karl Deisseroth, Peter Hegemann and Georg Nagel “for their discoveries concerning light-gated ion channels and optogenetics.” More to follow.

Taken together, DA neurons exert millisecond-scale causal control over BF bursting, from phasic reward-prediction responses to ongoing activity. This links two major subcortical systems previously studied largely independently into a directionally coupled circuit. 9/10

We think this reveals a new organizing principle for reward prediction: DA neurons can rapidly recruit a second broadcast system with distinct downstream targets. Glutamate co-release may provide the noncanonical DA-neuron output that enables this real-time coupling. 10/10

But what carries a signal this fast? Canonical D1/D2 receptor signaling is predominantly metabotropic and difficult to reconcile with ~10-ms coupling. Indeed, rapid BF recruitment by DA-neuron activation persisted during combined D1/D2 receptor blockade. 7/10

Anatomical tracing provided a plausible mechanism: VTA DA neurons formed glutamatergic synapses onto predominantly noncholinergic BF neurons. Together with the D1/D2 result, this points to glutamate cotransmission as a leading candidate for the rapid DA→BF signal. 8/10

We next tested causality by activating DA neurons. Brief excitation rapidly evoked BF bursting at the same ~10-ms timescale. Direct stimulation of DA axon terminals in BF produced similarly rapid responses, with even shorter latencies. DA activity is sufficient to recruit BF bursting. 5/10

The converse was even more striking. Inhibiting DA neurons, at cell bodies or axon terminals in BF, suppressed BF bursting with the same rapid latency, including ongoing firing outside task events. DA neurons therefore exert continuous control over a subset of BF bursting neurons. 6/10

Simultaneous recordings in behaving rats revealed strong trial-by-trial coupling in DA and BF response amplitudes. This relationship persisted across behavioral events and learning stages, arguing for a robust functional link between the two systems. 3/10

Timing revealed an even tighter relationship. Spike-time cross-correlation showed DA activity leading BF bursting by ~10 ms. The same offset persisted during ITIs and across learning stages, and noise-correlation analysis recovered the same offset after removing event-locked responses. 4/10

DA and BF reward-prediction signals have largely been studied in separate frameworks. Yet their activity is strikingly similar. This was unexpected because BF receives relatively sparse dopaminergic innervation and is not considered a canonical DA target. 2/10

New lab paper in #ScienceAdvances: we identify a modality-common selective-attention signal in noncholinergic #BasalForebrain (BF) “bursting” neurons, distinct from the modality-specific attention signals typically described in corticothalamic circuits. 1/n www.science.org/doi/full/10....

We also see a negative correlation between stimulus-locked bursts and outcome responses—an RPE-like signature consistent with prior reports—suggesting BF bursts may transform stimulus-onset RPE into an attention-for-performance amplification signal. 9/n

Huge thanks to lead author Sz-Wen Liu (@szwenliu.bsky.social)—freshly minted PhD 🎓—for leading this project from start to finish. Also grateful to our colleagues at NYCU, lab members, and reviewers who helped sharpen this study. 10/end

BF population activity mirrors behavior trial by trial, tracking attention switches during block transitions as well as spontaneous A↔V shifts, and aligning neural dynamics with the animal’s current attentional set. 7/n

Bimodal and unimodal trials show similar behavior and BF patterns, consistent with prior BF work: bigger bursts → faster reward seeking; suppression → nogo/rapid stopping. Extends a bidirectional gain-control view to selective attention. 8/n

When A and V oddballs co-occur but only one is attended, BF shows neither divisive normalization nor multisensory synergy. Instead, attended and ignored streams are processed in parallel and summed (near) linearly. 5/n

Interpretation: selection of the attended modality occurs upstream of BF; BF then provides a modality-common, post-selection/target-detection signal—an amplification that links attended targets to reward-seeking actions. 6/n

Single noncholinergic BF "bursting" neurons show strong, short-latency bursts to attended targets and minimal responses to identical stimuli when ignored. BF activity tracks behavior in real time, even during spontaneous shifts of attention. 3/n

Converge, not diverge: the same BF neurons carry attended A and attended V signals. Across units, amplitudes are tightly correlated (r≈0.8); variability is essentially a single scalar (amplitude), not modality tuning—unlike modality-specific attention signals in corticothalamic circuits. 4/n

@malcolmgcampbell.bsky.social

We built a new cross-modal oddball task in rats that pits auditory (A) vs. visual (V) stimuli simultaneously and allows both instructed and spontaneous attention shifts. Choices are unaffected by the unattended stream—clean behavioral evidence for cross-modal selective attention. 2/n

Shih-Chieh Lin reposted

Malcolm Campbell

19 Sept 2025

🚨Our preprint is online!🚨 www.biorxiv.org/content/10.1... How do #dopamine neurons perform the key calculations in reinforcement #learning? Read on to find out more! 🧵

Shih-Chieh Lin reposted

Zero-Noise Lab

@zeronoiselab.bsky.social

10 June 2025

Now out in @natcomms.nature.com‬: Mice and monkeys spontaneously shift through comparable cognitive states - and it's written all over their faces! (1/7) www.nature.com/articles/s41...

nature.comInferring internal states across mice and monkeys using facial features - Nature CommunicationsHere, the authors created a virtual reality task for monkeys and mice to explore if internal states like attention are similar across species. Their facial expressions during the task were similar, su...

Shih-Chieh Lin reposted

Paul Masset

@paulmasset.bsky.social

4 June 2025

Replying to @paulmasset.bsky.social

Together with this paper by the group of Joe Paton www.nature.com/articles/s41... our results across the two papers complement each other providing strong evidence that distinct dopamine neurons discount future rewards at different rates providing a substrate for multi-timescale RL in the brain

nature.comA multidimensional distributional map of future reward in dopamine neurons - NatureAn algorithm called time–magnitude reinforcement learning (TMRL) extends distributional reinforcement learning to take account of reward time and magnitude, and behavioural and neurophysiological...

Shih-Chieh Lin reposted

Paul Masset

@paulmasset.bsky.social

4 June 2025

Our work with Pablo Tano, @hyunggoo-kim.bsky.social Athar Malik, Alexandre Pouget and @naoshigeuchida.bsky.social exploring how dopamine neurons could enable multi-timescale reinforcement learning in the brain is out in @nature.com www.nature.com/articles/s41...

nature.comMulti-timescale reinforcement learning in the brain - NatureIndividual dopaminergic neurons encode future rewards over distinct temporal horizons.