3722 - Liquid Biopsy-Driven HER-2 Subtyping in Breast Cancer: A Multimodal Fusion Model Integrating Raman Spectroscopy and ClinLabOmics

Purpose/Objective(s): Accurate HER-2 subtyping remains a clinical challenge due to tumor heterogeneity and the limitations of conventional methods, such as immunohistochemistry (IHC) and fluorescence in situ hybridization (FISH). While liquid biopsies, such as circulating tumor DNA (ctDNA), lack metabolic resolution, spontaneous Raman Spectroscopy (RS) suffers from limited clinical interpretability. It is hypothesized that integrating RS with routine clinical tests (clinlabomics) can provide a noninvasive and metabolically informed approach to HER2 subtyping, reducing subjective assessments and resolving indeterminate HER2 cases.

Materials/Methods: This retrospective study included 2133 breast cancer patients diagnosed and treated at a single institution between 2021 and 2023. To overcome limitations in HER-2 assessment, an Omics Fusion Strategy (OFS) was developed that integrates serum RS with clinlabomics. The cohort comprised 1292 HER2-positive and HER2-negative cases. Importantly, 395 IHC 2+ cases with FISH confirmation were included to address discordance between IHC and FISH. Serum samples underwent RS analysis, with spectral data processed using a Convolutional Neural Network. Clinlabomics features were extracted using Random Forests, Gradient Boosting Machines, ElasticNet, Generalized Linear Models, and deep learning methods. Cross-attention mechanisms were employed to reorganize routine test features based on Raman-derived data. Model performance was evaluated using Accuracy (ACC), Area Under the Curve (AUC), Sensitivity, Precision, and F1-score.

Results: The multimodal fusion model outperformed single-modality approaches, achieving an AUC of 0.97 compared to RS alone (AUC = 0.89) and clinlabomics alone (AUC = 0.71). For ambiguous HER-2 cases (IHC 2+), the model predicted gene amplification with an AUC of 0.89, closely aligning with FISH results. Detailed results shown in Table 1. Key biomarkers prioritized via cross-attention included 5'-NT, GGT, AST, and ALP, along with lipid metabolism-associated Raman peaks (1080–1120 cm?¹, r = 0.82) correlated with HER-2 protein expression (p < 0.001), thereby enhancing interpretability and diagnostic accuracy.