Michael Hawrylycz

Multimodal genomic datasets now profile mammalian brain cells across anatomical regions, developmental stages, disease states, and species. Learning shared representations that capture cellular diversity across these heterogeneous observations remains a central challenge. Here, we introduce Bloom, a transformer-based model that represents cells using gene expression, chromatin accessibility, or both. Bloom is trained on 300,000 adult mouse brain cells with paired measurements, randomly sampled with broad-subclass balancing—approximately 100-fold fewer cells than used by comparable pre-trained models. By combining genomic sequence embeddings with randomly sampled multimodal cell views, Bloom outperforms unimodal cell models in resolving brain-wide cell types and in annotation, regression, and label-transfer tasks. Its attention patterns reveal cell-type-dependent associations among genomic loci, including pairs separated by millions of base pairs. These results establish Bloom as a data-efficient framework for integrating multimodal brain cell measurements and generating hypotheses about cell-type-specific genomic relationships.

Integrating multimodal data is essential to advance neuroscience as a discipline. Large-scale collaborative efforts, exemplified by the BRAIN Initiative, together with advances in single-cell and spatial omics, multimodal imaging, human genetics, behavioral phenotyping, protein structure prediction, and related technologies, are transforming neuroscience into a data-rich, highly interconnected science. As experimental resources, scientific literature, and collaborative networks continue to expand, integrating discoveries across modalities, species, and levels of biological organization has become the field's central bottleneck. This Perspective argues that artificial intelligence (AI) is becoming essential for overcoming these constraints. Using six evolving approaches we outline strategies to accelerate AI adoption through community building and pilot projects that drive AI transformation at both the organizational and technological levels. Organizational transformation cultivates broad AI awareness and a pro-AI culture while technological transformation improves the efficiency, scalability, and quality of data analysis, allowing researchers to generate insights more rapidly and effectively. Together, these advances position AI as a key driver for advancing neuroscience at scale.

Single-cell transcriptomic atlases resolve thousands of putative brain cell types, yet common analytical workflows often rely on low-dimensional embeddings that can distort neighborhood relationships and obscure continuous variation between closely related populations. We introduce single-cell Discrete Morse Graph Construction (scDMGC), a deterministic framework based on topological data analysis and discrete Morse theory that extracts a compact graph from neighborhood structure in high-dimensional gene-expression space. Because graph edges are derived from the original k -nearest-neighbor complex, scDMGC provides an interpretable multiscale representation of transcriptomic organization without requiring a low-dimensional embedding. Local maxima represent coherent transcriptional states, saddles quantify their separation, and gradient paths capture intermediate states and continuous transitions. Applied to cortical and hippocampal datasets, scDMGC recovers established inhibitory-neuron classes and elucidates transcriptional gradients. In mouse whole-cortex data, Morse graph structure evaluates cell-type hierarchy and type robustness. Persistence across scales provides a quantitative measure of cell-type identity and discrete versus continuous relationships. Finally, in Alzheimer’s disease single-nucleus data, scDMGC distinguishes changes in cell-type abundance from disease-associated shifts in transcriptional state. scDMGC therefore provides a multiscale framework for quantifying whether transcriptional populations form persistent cell types, how strongly external annotations are supported by intrinsic data structure, how cell types relate to one another, and where discrete organization transitions into continuous variation.

Alzheimer’s disease (AD) is defined and staged by the stereotyped, progressive accumulation of amyloid-beta (Aβ) plaques and hyperphosphorylated tau (pTau) tangles across brain regions. These regions differ substantially in their architecture and function but share largely conserved cellular composition, with some regional specialization. As pathology accumulates, specific neurons are lost and non-neuronal cells shift toward disease-associated states, but whether the cell types affected in any one region are the same across the others has remained unclear. Here we extended the Seattle Alzheimer’s Disease Brain Cell Atlas (SEA-AD) to ten neo- and allocortical regions spanning the cortical arc of canonical AD staging, profiling approximately seven million nuclei from 84 donors with single-nucleus RNA-seq, ATAC-seq, and Multiome alongside quantitative neuropathology and whole-genome sequencing. Nuclei were mapped to an expanded BRAIN Initiative reference taxonomy of 207 cell types, and a hierarchical pseudo-progression framework derived continuous, donor-level measures of AD pathological burden within each region and across the brain by jointly modeling Aβ and pTau. Cellular changes were both highly selective and strikingly consistent: only ∼30% of cell types shifted in relative abundance, but those that did changed in a coherent direction across regions. Specific subsets of Sst, Lamp5, Vip, Sncg, and Pvalb inhibitory interneurons and myelinating oligodendrocytes were lost earliest in preclinical donors with minimal pathology, alongside initial emergence of AD-associated microglia; loss of L2/3 and selected deep-layer excitatory types, sharper microglial increases, and reactive astrocyte emergence followed in later-stage donors. Regionally specialized populations were also vulnerable, including expected allocortical types and, unexpectedly, primary visual cortex (V1C)-specialized layer 4 (L4 IT) excitatory neurons and intermixed Sst and Pvalb interneurons. Key changes replicated across three independent cohorts encompassing over 700 additional donors. We examined two complementary vulnerable populations in mechanistic detail: regionally specialized V1C L4 IT neurons lost late despite being widely considered resilient, and pan-cortical Sst interneurons lost earliest in disease. Applying a multi-agentic AI workflow that constructed literature-grounded hypotheses from differential expression to L4 IT neurons nominated hyperexcitability, mediated in part by high NMDA receptor expression, as a convergent vulnerability phenotype. Vulnerable Sst interneurons converged on hyperexcitability through partly distinct pathways, and were enriched for expression of AD GWAS-prioritized genes, linking their vulnerability to the genetic architecture of AD. These data, available at SEA-AD.org, provide a multiregional framework for the community to explore the molecular and cellular changes of AD progression.

The mammalian brain consists of diverse neuron types with various functions. Recent single-cell RNA sequencing approaches have led to a whole-brain taxonomy of transcriptomically defined cell types¹. Patch-seq experiments augment these cell-type descriptions by linking transcriptomic profiles with local morphological and electrophysiological properties^2-7. However, linking transcriptomic identities to long-range axonal projections remains a major unresolved challenge. Here, to address this, we collected two datasets from the mouse visual cortex consisting of: (1) 1,528 excitatory Patch-seq neurons, with local morphological, electrophysiological and transcriptomic data collected from each cell, and (2) 341 excitatory, whole-neuron morphologies. From the Patch-seq data, we defined 17 morphoelectric-transcriptomic types and built a multistep classifier to integrate cell-type assignments with whole-neuron morphology and interrogate cross-modality relationships. We find that transcriptomic variation within and across morphoelectric-transcriptomic types corresponds with morphological and electrophysiological phenotypes. In addition, these gene expression patterns, along with the anatomical location of the cell, can be used to predict projection targets of individual neurons. We observed novel multimodal cell-type signatures for layer 5 intratelencephalic and extratelencephalic neurons and shed new light on their axonal circuitry, including interhemispheric intratelencephalic projections. With this approach, we establish a comprehensive, integrated taxonomy of cortical, excitatory neuron types, and create a system for high-dimensional cell-type classification that can be extended to the whole brain and potentially across species.

The NIH BRAIN Initiative Cell Atlas Network (BICAN) aims to generate a standardized, integrated cell atlas of the human, macaque, marmoset, and mouse brain that serves as a foundational community reference for the classification and study of brain cell types. Here we present the first major component of this effort: a cross-species, multimodal atlas of the basal ganglia, a group of subcortical nuclei central to motor control and implicated in a broad range of neurological disorders. Grounded in large-scale single-cell transcriptomic classification and integrated with epigenomic and spatial genomic modalities, this resource is enabled by coordinated cross-species sampling and harmonized analytical frameworks. It provides extensive phenotypic characterization of cell types, incorporates community-informed annotation, and establishes a highly curated, data-driven taxonomy with standardized nomenclature. The atlas is anchored to species-specific anatomical reference frameworks and linked across species through unified structural ontologies, enabling consistent cross-species comparisons. Multiple complementary datasets are mapped to this reference, including multiomic profiles and developmental trajectories aligned to adult cell states. Realization of this resource has required coordinated standards for tissue processing across human and model organisms, harmonization of donor metadata across brain banks, and the development of unified anatomical reference systems. To support these advances, BICAN has established an integrated ecosystem comprising standardized sequencing pipelines, neuroanatomically grounded data infrastructure, scalable visualization and mapping tools, and interoperable metadata standards. Analogous to the standardization achieved in genome science, this ecosystem provides a FAIR (findable, accessible, interoperable, and reusable) framework that enables researchers to map, compare, and interpret diverse datasets against a shared reference and associated knowledge base. The BICAN reference system is now being extended to the whole brain, with principles that are readily generalizable to other organ systems.

Single-cell mapping methods convert raw, heterogeneous single-cell datasets into interpretable and comparable representations of biological identity. As reference cell-type taxonomies mature, mapping new datasets to shared references has become a central strategy for enabling cross-study integration, reproducible annotation, and cumulative biological knowledge. Here we present MapMyCells , an open-source framework designed to align diverse single-cell omics datasets to hierarchical reference taxonomies with minimal preprocessing. MapMyCells provides out-of-the-box support for an expanding set of high-quality brain cell-type references generated by the Allen Institute for Brain Science, the BRAIN Initiative, and the Seattle Alzheimer’s Disease Brain Cell Atlas, including whole-brain mouse and human atlases, aging and Alzheimer’s disease cohorts, and a cross-species consensus taxonomy initially focused on the basal ganglia. MapMyCells enables efficient mapping of hundreds of thousands of cells on standard workstations without specialized hardware, providing a deterministic, scalable, and modality-agnostic approach that is robust across species and molecular assays. The framework produces interpretable confidence metrics and quantitative summaries of mapping performance, allowing users to evaluate assignment precision and accuracy. We demonstrate the mapping of unlabeled transcriptomic, epigenomic, and spatial datasets to reference taxonomies and describe a general workflow for preparing arbitrary hierarchical taxonomies for reference-based mapping. As the ecosystem of single-cell reference atlases expands, MapMyCells offers a practical and reproducible solution for community-scale cell-type annotation and cross-dataset integration, supporting the development of unified and extensible brain cell atlases.

SUMMARY: Aβ presence in the caudate nucleus (Ca) partially defines Thal stage III in Alzheimer’s disease (AD), but little is known about AD’s cellular impact on the region. Leveraging a public basal ganglia taxonomy of cellular populations, we generated a cellular resolution atlas of AD-associated pathological changes in Ca. Unlike cortex, we found that Ca AD pathology is dominated by two key features: phosphorylated tau (pTau)-containing neuropil threads enriched near oligodendrocytes in white matter tracts and amyloid-β diffuse plaques enriched in gray matter. Although AD pathology in affected cortical regions results in neuronal loss, we find no AD-driven reductions in neuron proportions in Ca. However, there were observable changes in multiple cellular populations. Protoplasmic astrocytes and FLT1+/IL1B+ microglia increased in abundance with global pTau levels. We also observe gene expression changes in fast-spiking PTHLH-PVALB interneurons indicative of disrupted signaling pathways and altered intrinsic physiological properties. This work provides a cellular-resolution framework for understanding AD pathology in Ca.