scLAP: A Liver-Adapted Contrastive Shared-Latent Framework for Pseudo-Pairing Single-Cell Transcriptomes and Chromatin Accessibility Profiles

Matched single-cell RNA sequencing (scRNA-seq) and single-cell assay for transposase-accessible chromatin sequencing (scATAC-seq) profiles from the same cells remain limited in hepatocellular carcinoma (HCC), hindering cell-resolved integration of transcriptional programs and regulatory chromatin states. To address this challenge, we developed single-cell Liver-adapted Alignment and Pseudo-pairing (scLAP), a liver-adapted contrastive shared-latent framework for cross-study RNA-ATAC pseudo-pairing. scLAP was trained on a paired liver single-nucleus multiome reference using 2000 highly variable RNA features and 697 ATAC features, including 647 transcription factor motif activity scores and 50 latent semantic indexing components. On a prespecified cell-level test partition (n = 3235), scLAP achieved an ATAC-to-RNA Top 10 exact-pair retrieval rate of 21.48% and a mean fraction of samples closer than the true match (FOSCTTM) of 0.0431. In six leave-one-sample-out folds, the corresponding macro means were 5.23% and 0.2278, respectively, indicating above-random ranking but limited exact-pair transfer to unseen samples. A validation-selected scButterfly configuration using 20,000 native ATAC peaks achieved 8.93% Top 10 retrieval and a mean FOSCTTM of 0.0811 on the same test cells; comparisons with Con-AAE and scMODAL remained study-specific adaptations. Applying reciprocal retrieval, a validation-derived cosine-similarity threshold, and greedy one-to-one assignment to independent HCC datasets yielded 4007 similarity-filtered candidate RNA-ATAC pseudo-pairs. Downstream analyses identified an NRCAM-associated accessibility candidate and separate exploratory MET-, CAV1-, and AP-1-associated observations.