Speciesformer learns conserved cellular states for cross-species generative virtual cell modeling

Abstract

Cells across species are governed by evolutionarily conserved biological programs, yet their molecular states and responses are reshaped by species, tissue and cellular context. A central challenge for virtual cell modeling is therefore to learn cellular states and state transitions that separate transferable biological principles from context-specific variation. Existing single-cell foundation models have advanced cellular representation learning, but most remain focused on single-species analysis, discriminative tasks or specialized forms of generation. Here we present Speciesformer, a cross-species generative single-cell foundation model that integrates evolutionary representation learning with virtual cell-state generation. Speciesformer is pretrained on SpeciesCorpus, comprising 131 million cells from 11 species, 154 tissues, and more than 923 cell types, and maps species-specific genes into a shared evolution-informed gene space. Its encoder learns transferable cell and gene representations that support biological representation probing and cross-species knowledge transfer, providing a common state space for resolving conserved and context-dependent cellular programs. Building on this shared representation, Speciesformer uses a unified generative architecture to model cellular states and state transitions under semantic and interventional conditions, enabling bidirectional generation between transcriptomic states and biological text descriptions, as well as prediction of post-perturbation transcriptomes from initial cell states and intervention descriptions, including in previously unobserved cellular contexts. By unifying evolutionary variation, biological semantics and conditional state transitions, Speciesformer extends cross-species foundation modeling toward a generative virtual cell framework for representing, describing and predicting cellular states across biological contexts.