Abstract 1P
Background
Retroperitoneal liposarcoma (RPLS) is defined as an aggressive malignancy for which adjuvant modalities have not shown increased efficacy compared to surgery due to its highly complex and heterogeneous features. Recent work has highlighted important prognostic information captured by computed tomography (CT) and histopathology-based multi-omics, with the potential to non-invasively characterize the so-called radiological phenotype. However, little is known about the ability to combine features from these different sources to improve the prediction of treatment response.
Methods
Eight independent cohorts of RPLS patients were enrolled from our center (n=183) and validation cohort (n=184). To develop and validate a CT-based multi-omics model to predict response to NACT (n=15) and radiotherapy (n=20) by combining contrast-enhanced CT images, genomic (n=59), bulk transcriptomic (n=100), lipidomic (n=50), plasma metabolomic (n=370), scRNA-seq (n=4) and corresponding multiplex immunohistochemistry (mIHC) (n=39) from tumor samples to assess the immune and metabolic landscape.
Results
We have developed a novel approach based on a traditional CT radiomics model to to accurately predict OS and tumor heterogeneity in RPLS. Using annotations, we developed a machine learning approach to integrate multimodal features into a risk prediction model. Our multimodal model (area under the curve (AUC) = 0.782) outperformed unimodal measures, including the clinical (AUC = 0.743), pathological (AUC = 0.723) and radiomics (AUC = 0.661). The multimodal risk score was significantly reduced with post-NACT, with more tertiary lymphoid structures (TLS) and immune cell infiltration, such as plasma cells and B cells, compared to pretreatment, and PCK1 were most significantly altered in the high multimodal risk score group with unfavourable prognosis.
Conclusions
Our study provides a quantitative rationale for using multimodal features to improve the prediction of prognosis, TIME, and response to post-NACT in patients with RPLS using expert-guided machine learning with excellent performance. This may be a promising avenue if validated by further prospective randomized trials.
Clinical trial identification
Editorial acknowledgement
Funding
Natural Science Foundation of Fujian province (No. 2023J011698); Natural Science Foundation of Xiamen City (No. 3502Z20227279); Scientific Research Project of Shanghai Municipal Health Commission (20214Y0087, 20204Y0409); “Young Talents” Training Plan of Shanghai TCM-integrated Hospital (No. RCPY0063); Scientific Research Project of Hongkou District Health Committee (No. 2302-02).
Disclosure
All authors have declared no conflicts of interest.