AIM: an accurate and explainable model for ATAC to GEX translation and pathway analysis

• DOI: 10.1007/s12293-025-00442-w
• Publication Date: 2025-03-22T15:18:34Z
• Authors: Quang H. Nguyen; Hoang V. Tran; Huu Tien Nguyen; Phuong T. M. Le; Phi Le Nguyen; Binh P. Nguyen

Abstract The development of multimodal technologies has enabled the simultaneous measurement of various cellular modalities, such as chromatin accessibility (ATAC), gene expression (GEX), and surface protein abundance in single cells. However, the lack of multimodal datasets requires the development of robust algorithms that can translate data between different modalities. In this study, we present AIM, a framework for accurate and interpretive multimodal translation, specifically designed for the conversion of ATAC data into GEX profiles. AIM introduces a novel two-tier modeling architecture. The upper tier captures the global relationships between ATAC and GEX, generating an initial estimate of gene expression. The lower tier performs a finer-grained analysis by modeling inter-chromosomal interactions to refine the generated GEX representation. This modular structure enhances both the accuracy and adaptability of AIM. Additionally, an integrated attention mechanism provides interpretability by highlighting critical chromatin regions influencing specific gene expressions. Our experimental results demonstrate that AIM achieves state-of-the-art performance, with a per-chromosome RMSE of 0.2206, outperforming existing approaches (0.2232). Furthermore, the attention maps generated by AIM offer a pathway analysis capability, uncovering biologically significant gene-gene interactions such as ARHGAP24-ARAP2 and SYK-PAX5. These findings validate AIM’s effectiveness not only as a data translation tool but also as a platform for deriving mechanistic insights into gene regulatory dynamics.