Huawen Zhong, Wenkai Han, Guoxin Cui, David Gómez-Cabrero, Jesper Tegnér, Xin Gao, Manuel Aranda
Combines RNA expression with embeddings from protein and general-purpose language models to learn universal cell embeddings across species, correcting batch effects and preserving conserved biological signals. Reconstructs more accurate multi-species cell-type evolutionary trees and uncovers convergent gene programs when applied to species separated by over 700 million years. Enables more accurate prediction of perturbation responses across species, such as from mouse to human, establishing Unify as a powerful method for comparative single-cell genomics and evolutionary biology.
Abstract Integrating single-cell RNA-sequencing (scRNA-seq) data across species is hindered by evolutionary divergence, technical batch effects, and the reliance on one-to-one orthologs. Here, we present Unify, a transfer learning methodology that learns universal cell embeddings by defining functionally coherent, multi-modal macrogenes. This is achieved by combining RNA expression with embeddings from protein language models and general-purpose language models. Unify transcends species boundaries, enabling cross-species comparisons beyond strict gene-level homology. Unify corrects batch effects while preserving conserved biological signals across vast evolutionary distances and enables more accurate prediction of perturbation responses across species, such as from mouse to human. Applied to species separated by over 700 million years, Unify reconstructs more accurate multi-species cell-type evolutionary trees and uncovers convergent gene programs. Together, these results establish Unify as a powerful method for comparative single-cell genomics and evolutionary biology.