Neuroscience AI and Brain Foundation Models
Neuroscience belongs in the Life Sciences AI Handbook because brain data sits at the boundary between cells, circuits, behavior, and disease. It is not well covered by molecular or cellular chapters alone. Neural recordings, connectomics, brain imaging, and neural decoding each have different data structures and validation requirements, so this chapter gives them a stable home.
- Distinguish neural activity models from clinical neuroimaging tools
- Identify the evidence standard for brain foundation model claims
- Separate connectomic prediction, neural decoding, and brain-imaging representation learning
- Recognize where model performance depends on species, preparation, task, and recording modality
- Link neuroscience AI claims to falsifying experiments or held-out neural recordings
Introduction
Brain data are not one modality. Calcium imaging, electrophysiology, connectomics, functional MRI, structural MRI, behavior, and clinical phenotypes measure different layers of the nervous system. AI systems trained on one layer should not be read as if they model the whole brain. The first question for any claim is: what neural signal, in what organism, under what task, with what validation?
A peer-reviewed Nature example shows why the area now deserves a chapter. Wang and colleagues trained a foundation model of neural activity from mouse visual cortex recordings and tested generalization to new stimulus domains and anatomical features (Wang et al., 2025). That is a serious signal for systems neuroscience. It is not evidence for general brain reasoning, human clinical prediction, or patient-level diagnosis.
Demonstrated
Demonstrated capability includes neural-activity prediction in bounded experimental systems when the held-out stimulus, animal, and neural response data are specified. The credibility comes from held-out recordings and from tests that cross stimulus classes or biological samples.
Demonstrated capability also includes AI-assisted analysis of brain images, neural recordings, and connectomic data as measurement-support tools. In those settings, the output is an annotation, embedding, segmentation, or prediction that still needs biological or clinical interpretation.
Theoretical
Theoretical capability includes brain foundation models that transfer across organisms, tasks, recording platforms, and disease states. The premise is plausible because neuroscience datasets are becoming larger and more multimodal. The validation burden is high because behavior, anatomy, physiology, and disease phenotype do not collapse into one representation.
Theoretical capability also includes neural decoding systems that generalize beyond tightly controlled tasks. Many decoding results are real for the studied setting; routine transfer to new people, disease states, and unconstrained behavior remains an evidence question.
Beyond current capabilities
Beyond current capabilities includes a general model of the human brain that predicts cognition, disease progression, treatment response, and behavior from arbitrary neural data. Current systems remain bounded by recording modality, experimental design, species, and clinical context.
Beyond current capabilities also includes reading subjective experience directly from neural recordings. A decoded stimulus, movement intention, or language signal is not a complete account of thought.
Practice Notes
- Name the organism, brain region, modality, task, and held-out condition.
- Keep animal-circuit evidence separate from human clinical evidence.
- Treat brain-imaging models as measurement and representation tools unless clinical validation is present.
- Require explicit privacy and consent review for human neural data.
- Avoid claiming general brain reasoning from a single sensory or imaging task.