An AI Agent for cell-type specific brain computer interfaces.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Biology

Background:

• Classifying neuronal subtypes from electrophysiological signals is crucial for understanding brain function but remains challenging.
• Existing methods require extensive training data and lack generalizability for in vivo recordings.

Purpose of the Study:

• To develop a novel AI framework, BCI AI Agent, for accurate and efficient neuronal cell-type classification from in vivo electrophysiological data.
• To enable few-shot learning for neuronal subtype inference without task-specific fine-tuning.
• To integrate cell-type classification with automated validation and dynamic neural decoding.

Main Methods:

• Repurposed pretrained vision-language models (VLMs) as few-shot learners for electrophysiological signal classification.
• Developed the BCI AI Agent, integrating vision-based inference, neuron tracking, and molecular atlas validation.
• Applied the framework to rodent motor-learning tasks and human Neuropixels recordings.

Main Results:

• Achieved robust and generalizable neuronal subtype inference with minimal supervision, validated against optogenetically tagged datasets.
• BCI AI Agent revealed stable, cell-type-specific neural trajectories in rodent motor learning.
• Successfully inferred and validated neuronal subtypes from human recordings using integrated atlases and literature.

Conclusions:

• BCI AI Agent offers a scalable, training-free approach for cell-type-specific inference of in vivo electrophysiology.
• This framework advances the dissection of neuronal population contributions to brain function and dysfunction.
• Enables new possibilities for neuroscience research and clinical brain-computer interfaces.