3648 - A Novel PHyLiC Model Using Dynamic Circulating Biomarkers to Predict Efficacy in Radioimmunotherapy for Advanced Solid Tumors

Purpose/Objective(s):

Despite the success of combining PD-1 inhibitor with Radiotherapy and Granulocyte-macrophage colony-stimulating factor (PRaG), traditional static biomarkers face significant challenges due to tumor tissue sampling limitations and the inability to monitor changes dynamically. Therefore, this study aims to develop a novel efficacy prediction model, PHyLiC, by dynamically monitoring changes in peripheral blood lymphocyte subsets and cytokines. The model integrates heterogeneous graph neural networks (heterGNN) and long short-term memory networks (LSTM) to explore complex relationships between patient baseline characteristics, immune responses, and changes across treatment cycles.

Materials/Methods:

Data from the PRaG 1.0 (ChiCTR1900026175), PRaG 2.0 (NCT04892498), and PRaG 3.0 (NCT05115500) studies were analyzed to assess the objective response rate(ORR) using RECIST 1.1 criteria. The data collected included baseline characteristics, and dynamic changes in peripheral blood lymphocyte subsets across 35 lymphocyte subsets and seven cytokines throughout treatment cycles. A feature matrix was constructed, and then heterGNN was applied using this data. Additionally, lymphocyte subset information from 126 healthy patients was incorporated to construct stable lymphocyte pairs for PRaG patients. The PHyLiC model was trained by integrating heterGNN with LSTM. The model’s performance was evaluated using five-fold cross-validation and compared with traditional machine learning methods (logistic regression, random forest, deep neural networks), sequential models (LSTM, Bi-LSTM), and sequential graph models (LSTM-GNN, LSTM-GCN). The model’s efficacy was further validated using independent data from two additional PRaG studies (NCT05790447 and NCT06112041).

Results: Using data from four treatment cycles, the PHyLiC model achieved an ROC AUC (Receiver Operating Characteristic Area Under Curve) of 0.815, an F1 score of 0.741, accuracy of 0.713, recall of 0.803, and PR AUC of 0.794, outperforming models based on data from two or three cycles. Compared with other machine learning methods, the PHyLiC model demonstrated superior performance in predicting treatment efficacy. Independent validation further confirmed the model’s robust performance, with a ROC AUC of 0.801.

Conclusion: The PHyLiC model successfully integrates dynamic immune monitoring data to predict the efficacy of the PRaG regimen. It provides more accurate reflections of immune response changes during treatment, offering precise efficacy predictions for patients with advanced metastatic solid tumors. The PHyLiC model provides valuable support for personalized treatment and clinical decision-making. However, further validation with larger sample sizes is needed to confirm the model’s generalizability.