Overview
Neural computation emerges from interactions across molecular, synaptic, circuit and behavioral scales. The Liu Lab combines quantitative experiments with interpretable artificial intelligence to discover how neural systems transform molecular events into circuit dynamics, movement and adaptive behavior. We integrate optical neurotransmitter measurements, multi-region neural recording, causal perturbation, super-resolution microscopy and data-driven modeling to build mechanistic—and ultimately predictive—accounts of brain function in health and disease.
Our research is organized around three connected directions:
01 — AI-Enabled Striatal Computation & Movement Control
02 — Systematic Presynaptic Profiling
03 — Molecular Architecture to Predictive Function
Together, these directions create a multiscale path from molecules to computation—linking fundamental mechanism, open quantitative tools and disease-relevant phenotypes.
01 — AI-Enabled Striatal Computation & Movement Control
We have developed synthetic-to-real multi-view 3D pose estimation, behavioral state-space models, multi-fiber photometry and causal perturbations to determine how striatal populations generate movement. We show that D1- and D2-MSN movement codes are spatially heterogeneous, while dopamine is broadcast more broadly; dopamine depletion can preserve overall MSN activity yet disrupt action representations. This program is informed by our discoveries that acetylcholine can initiate action potentials locally in distal dopamine axons (Science, 2022) and that rapid dopamine dynamics promote reward anticipation and vigor while baseline dopamine can sustain movement (Nature, 2024).
02 — Systematic Presynaptic Profiling
We are building a functional atlas of the presynapse by pairing high-throughput CRISPR perturbations with optical glutamate and GABA release measurements, readily releasable pool and release-probability assays, protein localization, network activity and transcriptomics. Our database (TransmissionGO) standardizes spontaneous-event analysis, while interpretable machine learning organizes genes into functional manifolds and prioritizes mechanisms linked to autism and schizophrenia. This direction extends our work showing that dopamine is released at sparse active-zone-like sites containing Bassoon and RIM (Cell, 2018), and that active-zone scaffolds assemble even without presynaptic CaV2 channels, although evoked release is abolished (Neuron, 2020).
03 — Molecular Architecture to Predictive Function
Our third direction asks whether molecular organization can predict neural function across scales. The dopamine domain-overlap framework links micrometre-scale release and receptor territories to local versus synchronized neuromodulation (Nature Reviews Neuroscience, 2021). Building on this spatial logic, the superresolution technology (me4Pi-SMLM) will map presynaptic proteins at approximately 2–3 nm localization precision, while the release reporter technology (RELEASE) is designed to connect molecular identity with functional output at scale. Integrating nanoscale architecture, optical physiology, functional sequencing and interpretable AI will move the field from descriptive maps toward experimentally testable predictions of release, circuit dynamics and behavior.