Integrated Transcriptomic Landscape and Deep Learning Based Survival Prediction in Uterine Sarcomas

Area of Science:

• Genomics
• Oncology
• Bioinformatics

Background:

• Uterine sarcomas (USs) are rare malignancies with poorly understood genomic profiles.
• Comprehensive genomic characterization is crucial for improved diagnosis and treatment strategies.

Purpose of the Study:

• To elucidate the genomic landscape of uterine sarcomas.
• To identify distinct genomic signatures and potential diagnostic markers for different US subtypes.
• To develop a predictive model for patient survival.

Main Methods:

• RNA-sequencing was performed on 71 US samples.
• Analysis included gene fusions, differentially expressed genes (DEGs), pathway enrichment, and immune cell infiltration.
• A deep learning model (MMN-MIL) was developed for survival prediction.

Main Results:

• Distinct gene fusion signatures were observed in endometrial stromal sarcomas (ESS) and uterine leiomyosarcomas (uLMS).
• A novel gene fusion, MRPS18A-PDC-AS1, showed potential as a diagnostic marker.
• The MMN-MIL model achieved high accuracy (0.804) and AUC (0.909) in predicting US patient survival.
• Prognostic value was identified for LAMB4 and specific immune cell types.

Conclusions:

• Uterine sarcomas exhibit unique genomic and gene expression features differentiating subtypes like high-grade ESS (HGESS), low-grade ESS (LGESS), and uLMS.
• The MMN-MIL deep learning model demonstrates significant potential for predicting US patient survival.