Bloom learns brain-cell representations from 300,000 cells, spanning RNA and chromatin accessibility at 91%–94% subclass accuracy

Gene expression and chromatin accessibility of brain cells can now be mapped into one representation space, so the same cell embedding can come from an RNA measurement or an accessibility measurement. Earlier single-cell pretrained models mostly took gene expression as input and represented genes by arbitrary identifiers or protein language model embeddings, which do not directly cover non-coding regions and often separate cells in latent space by profiling technology rather than biological correspondence. On adult mouse whole-brain multiome data, linear classifiers on Bloom embeddings reach 91% cross-validated adjusted balanced accuracy for 315 subclasses and corresponding results for 1,203 supertypes; the same model reaches 94% and 94% on scRNA-seq (310 subclasses) and snRNA-seq (297 subclasses) that were not used in training.

Each gene or accessibility peak is described by a combination of its DNA sequence, its relative position on the chromosome and its measured value in that cell, so the encoding does not depend on the measurement modality. Describing genes through protein embeddings applies only to protein-coding genes and does not extend directly to non-coding RNAs or chromatin accessibility peaks. The feature library covers 32,129 genes and 1,747,701 peaks, each peak a 500 bp interval; replacing Evo 2 or AlphaGenome sequence embeddings with random embeddings lowers classification accuracy, and shuffling Evo 2 embeddings across features returns accuracy to the random-embedding level.

During training each cell yields two randomly sampled feature views, built from expressed genes, accessible peaks or a mix, and the model jointly learns to reconstruct query genes and to pull the two views of the same cell together. On tabular single-cell data, simply adding more cells of the same kind has not produced clear gains, unlike the scaling behavior seen on natural-language corpora. Training used 292,221 mouse multiome cells balanced to at most 1,000 cells per broad subclass; raising the cap to 10,000 cells per subclass or adding unimodal RNA data produced no substantial improvement in downstream classification accuracy.

Gene–peak associations read from the model's first attention layer can serve as hypotheses, and the same gene–peak pair receives statistically significantly different attention across cell types. Earlier descriptions of gene–peak relationships relied largely on dataset-level co-expression or co-accessibility statistics. Between the Pvalb and L5-ET-CTX subclasses, after multiple-testing correction 25% of L5-ET-CTX-side pairs and 32% of Pvalb-side pairs show significantly different mean attention (q<0.05); 15 of 30 peaks around the Pvalb gene receive significantly higher attention in Pvalb cells.

The demonstrated scope is adult mouse whole-brain multiome training data plus linear classification, clustering and label transfer on mouse multiome, scRNA-seq, snRNA-seq, Patch-seq and aging datasets and on human snRNA-seq and Alzheimer's disease snRNA-seq evaluation sets; the authors propose that the architecture can also accommodate non-genomic features such as spatial location, morphology, connectivity and physiology, but those were not evaluated here.

The authors state that computationally inferred gene–peak associations should not be treated as true explanations before experimental validation; sequence embeddings improve performance, but which sequence properties, such as transcription factor binding motifs, drive the gain is not established; performance on the human Alzheimer's disease dataset trails some baselines, which the authors suggest may relate to the highly similar and continuous nature of human cortical supertypes.