DGAT: A Dual-Graph Attention Network for Inferring Spatial Protein Landscapes from Transcriptomics

Our plain-English summary of DGAT: A Dual-Graph Attention Network for Inferring Spatial Protein Landscapes from Transcriptomics, by H. Wang, B. Cody, H.U. Osmanbeyoglu (bioRxiv, 2025).

The study

• Authors: H. Wang, B. Cody, H.U. Osmanbeyoglu
• Journal: bioRxiv
• Published: 2025
• Cited by: 2 articles, as counted by NIH iCite
• Read it: PubMed · Publisher · Free full text

Abstract

Spatial transcriptomics (ST) technologies provide genome-wide mRNA profiles in tissue context but lack direct protein-level measurements, which are critical for interpreting cellular function and microenvironmental organization. We present DGAT (Dual-Graph Attention Network), a deep learning framework that imputes spatial protein expression from transcriptomics-only ST data by learning RNA-protein relationships from spatial CITE-seq datasets. DGAT constructs heterogeneous graphs integrating transcriptomic, proteomic, and spatial information, encoded using graph attention networks. Task-specific decoders reconstruct mRNA and predict protein abundance from a shared latent representation. Benchmarking across public and in-house datasets-including tonsil, breast cancer, glioblastoma, and malignant mesothelioma-demonstrates that DGAT outperforms existing methods in protein imputation accuracy. Applied to ST datasets lacking protein measurements, DGAT reveals spatially distinct cell states, immune phenotypes, and tissue architectures not evident from transcriptomics alone. DGAT enables proteome-level insights from transcriptomics-only data, bridging a critical gap in spatial omics and enhancing functional interpretation in cancer, immunology, and precision medicine.

In plain English

This preprint, posted to bioRxiv and not peer reviewed, described DGAT, a computer method that estimates protein levels across a tissue slice from RNA (gene activity) measurements and the position of each measured spot. It did not study asbestos exposure. The authors trained and tested the method on six samples measured for both RNA and about 30 proteins: two tonsil, one breast cancer, one glioblastoma, and two malignant mesothelioma samples that the authors processed themselves from tissue supplied by the National Mesothelioma Virtual Bank. Each sample was predicted by a model trained on the other five. For the first tonsil sample, the median correlation between predicted and measured protein rose from 0.308 using RNA alone to 0.610 with DGAT. For the two mesothelioma samples, the median correlations were 0.312 and 0.284. The authors suggested that the mostly immune protein panel may explain why prediction was harder in glioblastoma and mesothelioma. DGAT had a higher median correlation than four existing methods in all six held-out samples. The authors concluded that DGAT can give protein-level insight where protein measurements are unavailable, and they noted that suitable training datasets remain limited.

We wrote this plain-English summary from the full article with the help of AI, and a second AI model checked each sentence against the article. It is not the authors' wording; the abstract and the article are their own account.

What the study did and found

• Design: Computational methods study (deep learning model development and benchmarking); bioRxiv preprint not certified by peer review
• Who was studied: Six tissue samples with paired spatial RNA and protein measurements: tonsil (2), estrogen receptor-positive breast cancer (1), glioblastoma (1), and malignant mesothelioma (2, in-house, from formalin-fixed paraffin-embedded blocks from the National Mesothelioma Virtual Bank). Also public RNA-only datasets from lymph node, triple-negative and ER-positive breast cancer, melanoma and prostate cancer.
• Main results:
  • In the tonsil1 held-out sample, median protein-wise Spearman correlation between predicted and measured protein rose from 0.308 plus or minus 0.219 (RNA baseline) to 0.610 plus or minus 0.230 (DGAT), and RMSE fell from 1.61 to 0.226.
  • With the full DGAT model, median Spearman correlations were 0.312 for mesothelioma sample 1 and 0.284 for mesothelioma sample 2, compared with 0.610 to 0.658 for the tonsil samples and 0.579 for breast cancer. Removing the spatial graph lowered them to 0.266 and 0.178.
  • The authors suggested that prediction was harder in glioblastoma and mesothelioma because the protein panel is mostly immune markers and these tissues typically have lower immune infiltration or more complex tumor microenvironments.
  • Compared with four existing methods that do not use spatial information (Seurat v4, sciPENN, cTP-net, scLinear), DGAT had higher median correlation in all six held-out samples.
  • In a lymph node dataset without protein measurements, DGAT-based clustering matched annotated germinal centers with an F1 score of 0.823, compared with 0.806 for CTP-net, 0.782 for scLinear, 0.291 for sciPENN, 0.667 for Seurat and 0.575 for transcriptome-only clustering.
  • Agreement with expert region annotations (adjusted Rand index) was 0.504 for DGAT versus 0.447 for transcriptome-only clustering in triple-negative breast cancer, and 0.683 for DGAT versus 0.638 for the mRNA-only model in prostate cancer.
• The authors' conclusion: The authors concluded that DGAT can infer protein-level spatial maps from transcriptomic data alone, offering proteome-level insights where protein measurements are unavailable. They stated that its application to in-house malignant pleural mesothelioma tissue demonstrates its value beyond well-characterized diseases.
• Limitations the authors state:
  • Broader use depends on spatial CITE-seq training datasets, which remain relatively limited.
  • Because mRNA–protein relationships vary by tissue and cell state, the model may need retraining or fine-tuning for tissues unlike those it was trained on.
  • Only about 30 immune-related proteins were evaluated; larger protein panels have not been systematically benchmarked.
  • Each measured spot often contains several cells, mixing their signals and limiting fine-grained inference, especially in tumors.

Topics

Medical terms in this study, explained in our Lexicon: Mesothelioma.

Related studies

• Spatial Landscape of Malignant Pleural and Peritoneal Mesothelioma Tumor Immune Microenvironments — X. Ma, D. Lembersky, et al. (2024)
• Comparison of imaging based single-cell resolution spatial transcriptomics profiling platforms using formalin-fixed paraffin-embedded tumor samples — N. Ozirmak Lermi, M. Molina Ayala, et al. (2025)
• Identification and Characterisation of Infiltrating Immune Cells in Malignant Pleural Mesothelioma Using Spatial Transcriptomics — D. Shek, B. Gloss, et al. (2023)
• Comparison of imaging-based single-cell resolution spatial transcriptomics profiling platforms using formalin-fixed, paraffin-embedded tumor samples — N.O. Lermi, M.M. Ayala, et al. (2025)
• The Immunological Landscape of M1 and M2 Macrophages and Their Spatial Distribution in Patients with Malignant Pleural Mesothelioma — C. Laberiano-Fernandez, C.M. Baldavira, et al. (2023)
• Current Landscape of Immunotherapy Trials Involving the Programmed Cell Death Protein 1/Programmed Death-Ligand 1 Axis in Intrathoracic Tumors — S.A. Kareff, S. Samtani, et al. (2021)

Browse return to medical research.