scAICT: Adversarial Integration and Contrastive Translation for Unpaired scRNA-Seq and scATAC-Seq

Unpaired single-cell gene expression and chromatin accessibility data lack the cell-level correspondence needed to directly link measurements across modalities. We present scAICT, a two-stage framework that infers this correspondence through integration and uses it to guide bidirectional molecular prediction. Modality-specific variational autoencoders combine adversarial alignment with structural and anchor constraints to learn a shared representation. Inferred pseudo-pairs then supervise directional mapping networks through cross-modal reconstruction and contrastive matching. We evaluate scAICT on naturally unpaired Muto-2021 kidney data and PBMC-10k data under unpaired training, with a complementary paired PBMC-10k translation benchmark. scAICT achieves the highest overall integration scores among the evaluated methods, reaching 78.62% on Muto-2021 and 84.61% on PBMC-10k, and supports cross-modal annotation transfer. Under unpaired training, its mean biological conservation scores exceed Monae in both directions, with larger gains for gene expression prediction from chromatin accessibility. PBMC-10k molecular evaluations also show higher median cell-level correlations than Monae in both directions, including in cells held out from translation training. Predicted RNA retains cell-type-associated marker patterns. These findings support molecular prediction guided by inferred correspondence, with evidence for cell-type conservation in kidney data and molecular-profile concordance in PBMC-10k.