Documentation

Quality Control and Library Filtering

Our in-house profiling produced 109 single-nucleus libraries spanning 52 vertebrate species, generated with two parallel techniques (SeekOne, 90 libraries, and 10x Chromium, 19 libraries). Nuclei were filtered with a uniform, permissive quality threshold, and libraries with low read-mapping rates were set aside before any downstream analysis.

For species profiled with both techniques, the two platforms captured comparable cell numbers and gene counts, and cell-type composition and expression patterns were reproducible between them, indicating that both profiling methods are reliable and robust.

Batch Correction and Unsupervised Clustering

Batch effects were corrected with Harmony using library identity as the sole batch covariate, and the integrated cells were clustered with Leiden. This produced the atlas-level cluster annotation used throughout the resource.

After integration, cells clustered by cell type rather than by species of origin: the median species iLISI was 7.76 versus a cell-type iLISI of 1.07, and the large clusters each spanned a median of 99 of the 100 analysis-layer species, indicating that the same cell type shares similar molecular blueprints across species.

A Uniform Cross-Species Matrix

The full integrated matrix comprises 165 libraries and 4,414,988 cells (including the 109 newly constructed libraries with 945,408 newly profiled cells). From this, a uniformly processed analysis layer was derived for all cross-species analyses, capped at 5,000 cells per species to balance representation across the tree.

The analysis layer is anchored to human homologs retained across species, which defines the gene universe for all cross-species analyses in this resource. For a full species-by-species breakdown, see the species catalogue.

Cell Program Scores and Detection

Across the atlas, cell programs (for example excitatory, inhibitory, astrocyte, or myelin programs) are scored at single-cell resolution, then summarized to species-level calls so that presence and absence can be compared across the phylogeny.

Strict and Inclusive Ortholog Mapping

All cross-species comparisons are evaluated under two ortholog versions that differ in how ambiguously mapped homologs are handled. Reporting both makes the results robust to orthology-mapping stringency.

The choice shifts how many genes are carried and how much signal is retained per gene. The gene-age conclusions are stable across both versions: for example, the top neuronal-subtype markers remained 100% Euteleostomi-age under strict one-to-one (n = 261) and one-to-one-plus-many-to-one (n = 314) ortholog sets, and the four cross-species marker programs behaved consistently when assayable orthologs were compared.

Cell-Type Annotation Curation

The 43 atlas clusters were assigned to major cell types by expert manual curation of cluster-level marker expression, and the curated assignments were then executed programmatically as a cluster-to-type assignment table (the full curation table is provided in Supplementary Table 2).

Twelve of the 27 neuronal clusters rest on lower-confidence marker evidence and are retained as curated assignments pending orthogonal support. Unassigned clusters were annotated as co-expressing lineages or left undefined. In total, 483,837 of the 492,121 analysis-layer cells (98.3%) received a unified cell-type annotation. Annotation robustness was stress-tested by a full-pipeline rerun with an alternative highly-variable-gene set: cluster-level concordance was on par with or below the primary run, leaving all main conclusions unchanged.

Core identity rests on a set of 46 canonical markers selected from an extensive literature review for the major brain cell types, including neuronal (pan-neuronal, inhibitory and excitatory) markers, and all 46 core cell-type markers trace to the oldest phylostratum (Euteleostomi-age) within the cross-species-retained ortholog space.