Predicting and Monitoring the Toxicity of Immune Checkpoint Inhibitors: Breakthrough Application of the DysPred Deep Learning Framework

Academic Background

Immune checkpoint inhibitors (ICIs) represent a major breakthrough in cancer immunotherapy in recent years, enhancing the body’s antitumor immune response by inhibiting immune checkpoint signaling pathways. However, ICIs can trigger a wide range of immune-related adverse events (irAEs) during treatment, which not only affect patients’ quality of life but may also lead to organ dysfunction or even death. Due to the high heterogeneity of irAEs across clinical settings, tumor types, tissue specificity, and patient demographics, there is an urgent need for a robust and scalable method to predict and manage these adverse effects.

Although previous studies have explored irAEs through clinical trials and traditional adverse drug reaction datasets (such as SIDER and OFF SIDES), these approaches are often limited by small sample sizes, conflicting data, and the inability to predict future adverse drug reactions. Therefore, leveraging large-scale pharmacovigilance data, especially through post-marketing surveillance, to identify and understand the toxicity of ICIs has become a critical research priority.

Paper Source

This study was conducted by Yan Dongxue, Bao Siqi, Zhang Zicheng, Sun Jie, and Zhou Meng from the National Clinical Research Center for Ocular Diseases, Eye Hospital, Wenzhou Medical University. The research was published in Nature Computational Science in 2024, with the DOI 10.1038/s43588-024-00748-8. The research team developed the DysPred framework, which for the first time applies dynamic graph convolutional networks (DGCN) to pharmacovigilance data analysis, aiming to predict the toxicity risks of ICIs at the population level.

Research Process and Results

1. Data Preprocessing and Pharmacovigilance Analysis

The research team extracted 13,754,811 reports from the FDA Adverse Event Reporting System (FAERS) between 2014 and 2022. After quality control, 1,539,445 high-quality reports were selected, including 56,209 reports related to ICIs. ICIs primarily include anti-PD-1 drugs (e.g., nivolumab, pembrolizumab), anti-PD-L1 drugs (e.g., atezolizumab, avelumab), and anti-CTLA-4 drugs (e.g., ipilimumab).

The researchers used disproportionality analysis methods, including the reporting odds ratio (ROR) and the empirical Bayesian geometric mean (EBGM), to assess the association between ICIs and adverse events. The results showed that strong or very strong signals induced by ICIs were primarily concentrated in the immune and endocrine systems.

2. Construction of the DysPred Framework

The core of the DysPred framework is a deep learning model based on dynamic graph convolutional networks, designed to predict the toxicity risks of ICIs through pharmacovigilance data analysis. The framework consists of three key steps:

• Toxicity Landscape Generation: Generate toxicity landscapes for ICIs based on reported adverse events and compare them with the overall pharmacovigilance cohort.
  
• Toxicity Risk Prediction: Predict potential toxicity landscapes for the next five years using disproportionality analysis graphs combined with node semantic similarity. The model employs k-core graphs and long short-term memory (LSTM) networks to generate node representations in low-dimensional latent space and produce predicted safety risk scores (PRSS) for each ICI-induced toxicity.
  
• Risk Classification and Output: Convert the safety risk distribution of each ICI-induced toxicity into signal safety risk classes through repeated model construction processes.

DysPred demonstrated exceptional predictive performance across multiple cancer types and demographic cohorts, particularly showing strong robustness in small-sample scenarios. For example, DysPred achieved a prediction accuracy of 81.2% for CTLA-4 therapy in melanoma (MEL) patients.

3. Logical Relationships and Contributions of the Results

The results indicate that DysPred can accurately capture toxicity signals induced by ICIs and exhibit stable predictive performance across different time points. Moreover, DysPred shows high consistency with known physiological systems and transcriptional profile (CTP) changes, further validating its clinical relevance. By analyzing the toxicity risks of different ICI treatment regimens, DysPred provides clinicians with an important tool for optimizing patient treatment plans.

4. Research Highlights and Significance

The innovations of the DysPred framework include:
- Efficient Data Integration: Combining large-scale pharmacovigilance data with semantic similarity analysis to achieve precise predictions of ICI toxicity.
- Dynamic Temporal Analysis: Capturing the evolution of toxicity signals through time-series dynamic graph convolutional networks, offering a new perspective for post-marketing surveillance.
- Broad Application Prospects: DysPred is not only applicable to ICIs but can also be extended to the toxicity risk assessment of other antitumor and non-antitumor drugs.

5. Research Conclusions and Value

The DysPred framework provides a novel approach to assessing the toxicity risks of ICIs, enabling the prediction and management of potential toxicity risks through large-scale pharmacovigilance data. Its scientific value lies in revealing the dynamic evolution of ICI toxicity and offering clinicians evidence to optimize treatment strategies. Additionally, DysPred’s success provides a technical pathway that can be referenced for toxicity prediction of other drugs.

Other Valuable Information

The research team also explored DysPred’s potential in predicting rare and new drug-adverse event associations. Through temporal evaluations of dynamic disproportionality analysis graphs, DysPred demonstrated excellent performance in predicting small-sample and emerging drug-adverse events, further expanding its value in practical applications.

The DysPred framework not only achieves technical breakthroughs but also provides critical decision-making support for the clinical application of ICIs, advancing the safety and efficacy of cancer immunotherapy.