Multisource Omic Alignment and Biological Feature Discovery with Performer Encoder and Triplet Networks.

Advances in single-cell sequencing technologies greatly enhance our understanding of molecular and cellular features. However, effectively leveraging these data to uncover key biological factors remains a major challenge in integrative analyses across multiomic types and comparative studies across species, particularly livestock species such as pigs and cattle. To address this, we develop AlignCell, a deep learning model designed to learn robust biological features by integrating multisource omic data across platforms, omic types, and species, thereby facilitating the discovery of key factors, such as conserved and species-specific genes in cross-species comparative studies. Across various applications and benchmarking compared with existing tools, AlignCell performs well. Notably, using AlignCell to integrate female gonad data across four species (human, mouse, pig, and cattle), including the bovine single-cell data generated in this study, AlignCell reveals the unexplored species-conserved gene CCT2 in primordial germ cells (PGCs). Additionally, it identifies unexplored species-specific genes PRICKLE4 and CTSV in pig and cattle PGCs, providing important insights for reproductive and developmental research.