Novel dynamic circulating biomarkers for predicting therapeutic efficacy of PRaG regimen in advanced refractory solid tumors.

Abstract

2548 Background: Common biomarkers for predicting the efficacy of immune checkpoint inhibitors (ICIs), such as programmed death-ligand 1 (PD-L1) expression, face notable challenges with tumor tissue sampling and the inability to enable dynamic monitoring. Circulating T lymphocyte subset classification and cytokines offers a promising alternative, reflecting T cell functionality and predicting ICI responses. The PRaG regimen, combining PD-1 inhibitors, radiotherapy, and granulocyte-macrophage colony-stimulating factor (GM-CSF), has shown efficacy in patients with metastatic or refractory solid tumors unresponsive to standard therapies. This study seeks to develop an efficacy evaluation model by intergrating dynamic peripheral blood lymphocyte subsets and cytokines, based on comprehensive analysis of clinical data from the PRaG studies. Methods: Data from the PRaG 1.0 (ChiCTR1900026175), PRaG 2.0 (NCT04892498), and PRaG 3.0 (NCT05115500) studies were analyzed to evaluate the objective response rate (ORR) by RECIST 1.1. Machine learning models, including linear, sequential, attention-based, and hybrid models, were employed to predict disease progression. These models utilized dynamic peripheral blood data from thirty-five lymphocyte subsets and seven cytokines, collected across treatment cycles. Model efficacy was further validated using independent data from two additional PRaG studies (NCT05790447 and NCT06112041). Results: As of November 30, 2023, 132 patients were included in the study, with a median age of 63 years. Patients over 65 accounted for 41.7%, and 60.4% had more than five metastatic sites. Patients with an ECOG score of 2-3 made up 59.7% of the cohort. The ORR was 20.13%, and the disease control rate was 48.19%. Dynamic monitoring of peripheral blood features across treatment cycles facilitated the development of an LSTM-HeterGNN model, which integrates long short-term memory (LSTM) networks with heterogeneous graph neural networks (heterGNN). This model outperformed ten other models, achieving a ROC AUC of 0.818. Independent validation further demonstrated robust performance, with a ROC AUC of 0.801. Conclusions: This study underscores the potential of the PRaG regimen as an effective salvage therapy for advanced solid tumors after the failure of standard treatments. The LSTM-HeterGNN model, leveraging dynamic peripheral blood biomarkers, provided precise efficacy predictions, surpassing traditional models. These findings lay the groundwork for dynamic treatment monitoring and optimization. Larger sample sizes are required to further validate the model’s generalizability.