Regulatory-prior-guided attention preserves biological structure during unpaired single-cell RNA-ATAC integration

scHPGT uses modality-specific encoders to model RNA and ATAC signals, a prior-guided cross-modal Transformer that constrains gene-peak attention using regulatory links, and a domain-adversarial objective to reduce modality-specific discrepancies in a shared latent space. Relative to existing approaches that frame integration as distribution matching and can over-align biologically distinct or condition-specific cell states, this work writes regulatory priors directly into the attention constraint rather than relying on distribution alignment alone. The abstract explicitly names three components (modality-specific encoders, prior-guided cross-modal Transformer, domain-adversarial objective) and reports results on four benchmarks; specific numbers and ablation details are not given in the abstract text.

Across PBMC3k, mouse spleen, CITE-seq/ASAP-seq PBMC and PBMC10k benchmarks, scHPGT improves clustering agreement, label transfer and biological structure preservation while maintaining effective modality alignment. Evaluation is extended beyond alignment alone to clustering agreement, label transfer and structure preservation, spanning multiple datasets and platform combinations. Based on comparisons across four named benchmarks, this is an abstract-level overall conclusion; concrete metric values, statistical tests and the baseline list are not listed in the text.

In partial-overlap and condition-shift settings, scHPGT aligns shared populations without forcing unmatched or condition-specific states into inappropriate correspondence. Addresses the failure mode of distribution-matching integration that can over-align, describing behavior under partial-overlap and condition-shift settings. Partial-overlap and condition-shift settings serve as the evidence scenarios, described qualitatively; no quantitative metrics for these settings are given.

Attention-derived links recover regulatory relationships, highlight marker-gene regulatory regions, recover transcription factor programs and produce regulatory activity profiles consistent with cell-type-specific transcriptional programs. Uses the integration model's attention weights as an interpretable regulatory readout rather than only as a byproduct of latent-space alignment. Supported by consistency between attention-derived links and known regulatory relationships, marker-gene regulatory regions, transcription factor programs and cell-type-specific transcriptional programs; the specific validation approach and scale are not expanded in the abstract.

The work targets integration of unpaired single-cell RNA and ATAC profiles in settings with shared populations, partial overlap or condition shift, aiming to align shared populations while preserving condition-specific and modality-specific structure; beneficiaries include single-cell analysts needing cross-modal label transfer, clustering and regulatory interpretation. The method relies on regulatory links as a prior, so its applicable scope relates to the coverage of the regulatory prior used; code and datasets are public, supporting reproduction and extension on other data.

This assessment rests on abstract-level text without figures or supplementary material, so the specific magnitude of improvement in clustering agreement, label transfer and structure preservation, the baseline list and statistical tests cannot be confirmed here; quantitative behavior under partial-overlap and condition-shift settings, the degree of agreement between attention-derived regulatory links and known regulatory relationships, and method behavior when the regulatory prior is missing or inaccurate are open questions worth checking in the full text.