MAGNET: Multi-view graph autoencoder with cell-gene attention for cell interaction network reconstruction from...

Area of Science:

• Computational Biology
• Genomics
• Systems Biology

Background:

• Inferring cell-cell interactions from spatial transcriptomics is difficult due to tissue complexity.
• Current deep learning models often separate external cell communication and internal gene regulation.
• This separation limits understanding how cellular environments influence internal states.

Purpose of the Study:

• To develop a novel framework, MAGNET, for reconstructing cell-cell interaction networks.
• To unify the representation of cell-cell communication and internal gene activity.
• To improve the accuracy of cell-cell interaction inference in spatial transcriptomics.

Main Methods:

• Developed MAGNET (Multi-view Graph Autoencoder with Cell-Gene Attention Network).
• Constructed multiple biological graphs representing cellular information.
• Integrated a Cell-Gene attention module to link external environment with internal gene activity.

Main Results:

• MAGNET achieved superior performance in reconstructing cell-cell interaction networks on benchmark datasets (seqFISH, MERFISH, STARMAP).
• Achieved an Average Precision (AP) of 0.901 on the seqFISH dataset, outperforming TENET by 0.185.
• The Cell-Gene Attention module was critical, with its removal dropping AP from 0.901 to 0.521.

Conclusions:

• MAGNET effectively reconstructs cell-cell interaction networks by unifying cellular environment and gene activity.
• Applied to breast cancer data, MAGNET identified functional heterogeneity in cancer cells, distinguishing immune evasion and autonomous growth signatures.
• MAGNET offers a powerful tool for dissecting complex cellular communication in tissues.