GeneCytNet: an interpretable deep learning framework for rheumatoid arthritis classification and in silico cytokine perturbation modeling

C. L. Philip Chen, Dagang Li, Lujia XuFrontiers in Immunology 2026 · DOI

TL;DR

This paper introduces GeneCytNet, a novel deep learning framework that combines Variational Autoencoders (VAEs) and Graph Attention Networks (GATs) to accurately classify Rheumatoid Arthritis (RA) from gene expression data. Crucially, GeneCytNet goes beyond prediction by offering mechanistic insights through "in silico" cytokine perturbation simulations, revealing the relative impact of key inflammatory cytokines on disease risk.

What Problem Does This Paper Address?

Rheumatoid Arthritis (RA) is a complex autoimmune disease characterized by significant heterogeneity in its underlying molecular pathways and clinical presentation. While transcriptomic data offers a rich source of information for understanding RA, existing machine learning models often struggle to provide interpretable insights into the biological mechanisms driving the disease. Specifically, it's challenging to decipher the precise roles of key cytokines, like IL-6, TNF-α, and IL-1β, in orchestrating gene networks and influencing disease progression. This lack of interpretability hinders accurate diagnosis, precise patient stratification, and the development of targeted therapies.

Methods

The core of this work is the GeneCytNet framework, a deep learning architecture designed for both classification and mechanistic interpretation. It comprises three main components:

  1. Variational Autoencoder (VAE) Encoder: This component takes the high-dimensional gene expression data and compresses it into a lower-dimensional "latent space." This process helps to denoise the data and capture non-linear relationships, creating a more robust representation.
  2. Graph Attention Network (GAT): GeneCytNet constructs a gene co-expression graph where genes are nodes and edges represent their correlated expression. The GAT then uses attention mechanisms to learn how these gene interactions influence RA classification, effectively modeling the network structure of biological pathways.
  3. Classification Head: A standard multi-layer perceptron (MLP) that takes the processed gene embeddings and predicts the probability of a sample belonging to the RA class.

A novel aspect is the in silico cytokine perturbation module. This module simulates the effect of altering specific cytokine-responsive gene modules by shifting their representations in the latent space. By observing the change in predicted RA risk (Δp), the framework can quantify the causal-like influence of individual cytokines.

Datasets

The study utilizes a synthetic gene expression dataset for proof-of-concept and initial validation. This dataset was carefully constructed to mimic key characteristics of real transcriptomic data, including:

  • Training/Validation Set: 240 RA patients and 120 healthy controls, with 15,000 gene features.
  • Independent Holdout Cohort: 100 RA patients and 50 healthy controls, also with 15,000 gene features.

The synthetic data was generated by simulating gene expression under Gaussian distributions, with deliberate shifts for differentially expressed genes (DEGs) in RA samples and organized into modules to mimic biological pathways. This synthetic approach allowed for rigorous model evaluation without the confounding factors present in real-world data. The authors mention plans to validate on publicly available RA datasets (e.g., Gene Expression Omnibus: GSE89408, GSE55235) in future work.

Key Insights and Ideas

  • Bridging Prediction and Mechanism: The most significant insight is GeneCytNet's ability to simultaneously achieve high diagnostic accuracy and provide interpretable biological insights. It moves beyond a "black box" model to offer testable hypotheses.
  • Emergent Biological Knowledge: The framework learned the hierarchy of cytokine influence (IL-6 > TNF-α > IL-1β) without being explicitly told which genes belonged to which cytokine pathway. This emergent understanding strongly validates the model's ability to capture genuine biological signals.
  • "Virtual Patient Simulator": The in silico perturbation capability allows GeneCytNet to act as a virtual simulator, enabling researchers to explore the potential impact of modulating specific biological pathways on disease risk.
  • Synergistic Architecture: The ablation study clearly demonstrated that both the VAE (for feature compression) and the GAT (for network modeling) are crucial for optimal performance, highlighting the power of their integration.

Key Results

  • Superior Classification Performance: GeneCytNet achieved a test AUC of 0.962 ± 0.005, significantly outperforming all baseline models, including classical ML algorithms (like XGBoost) and simpler neural networks. Accuracy was 0.914 ± 0.007, and F1-score was 0.915 ± 0.006.
  • Robustness: Cross-validation confirmed the model's robustness with a mean AUC of 0.957 ± 0.006.
  • Mechanistic Insights:
    • SHAP analysis identified IL6R and STAT3 as the most influential gene-level biomarkers.
  • These key genes were co-localized in Module 27, which was identified as the most influential gene module.
  • Module 35, containing TNFSF11 (RANKL), was also a significant contributor.
  • Cytokine Hierarchy: In silico perturbation experiments revealed a clear hierarchy of cytokine influence on RA risk prediction:
  • IL-6: Upregulation increased RA probability by +0.12 ± 0.03 (most significant effect).
  • TNF-α: Upregulation increased RA probability by +0.08 ± 0.02.
  • IL-1β: Upregulation increased RA probability by +0.06 ± 0.02.

Limitations and Future Work

  • Synthetic Data Validation: The primary limitation is that the model was trained and validated on synthetic data. While this allowed for controlled proof-of-concept, real-world validation with inherent technical and biological variability is the essential next step.
  • Simplified Cytokine Biology: The in silico perturbation model simplifies the complex, dynamic nature of cytokine signaling, which involves feedback loops, cross-talk, and cell-type-specific effects.
  • Module Annotation Uncertainty: While gene modules were defined by correlation and informed by gene ontology, there's an inherent degree of uncertainty in their precise biological annotation.

Future work will focus on:

  • Real-world Data Validation: Applying GeneCytNet to multi-center RA transcriptomic datasets (e.g., from GEO) to assess its clinical translatability.
  • Multi-omics Integration: Incorporating proteomic and epigenomic data to provide a more comprehensive view of disease mechanisms.
  • Single-cell RNA-seq Analysis: Dissecting cell-type-specific signaling networks within the RA synovium.
  • Precision Medicine Applications: Exploring how cytokine module sensitivity could stratify patients for personalized treatment strategies, potentially predicting treatment response (as illustrated in Figure 5).

Summary updated

Abstract

Background Rheumatoid arthritis (RA) is a heterogeneous autoimmune disease where cytokine-driven dysregulation of gene networks poses a significant challenge for accurate diagnosis and targeted therapy. While transcriptomic data hold immense promise, most machine learning models lack the interpretability to decipher the underlying biological mechanisms, particularly the specific roles of key cytokines. Methods We developed GeneCytNet, a novel deep learning framework that integrates a Variational Autoencoder (VAE) for nonlinear feature compression with a Graph Attention Network (GAT) to model gene-gene interactions. The model was developed on a synthetic cohort of 240 RA and 120 healthy control samples, with an independent holdout cohort of 100 RA and 50 controls, each with 15,000 gene features, designed as a robust proof-of-concept. Performance was benchmarked against classical models, and generalizability was assessed via cross-validation and the independent holdout. Crucially, we introduced in silico cytokine perturbation experiments to simulate the effect of modulating IL-6, TNF-α, and IL-1β responsive gene modules on RA risk prediction. Results GeneCytNet achieved superior classification performance, with a test AUC of 0.962 ± 0.005, accuracy of 0.914 ± 0.007, and an F1-score of 0.915 ± 0.006, outperforming all baseline models. Cross-validation confirmed robustness (mean AUC = 0.957 ± 0.006). The perturbation experiments provided mechanistically interpretable insights, revealing that the IL-6–responsive module had the most significant effect on RA probability (+0.12 ± 0.03), followed by TNF-α (+0.08 ± 0.02) and IL-1β (+0.06 ± 0.02). This hierarchy of cytokine effect sizes aligns with established clinical evidence. Conclusion GeneCytNet demonstrates that advanced, interpretable deep learning can simultaneously achieve high diagnostic accuracy and generate testable biological hypotheses. By functioning as a virtual patient simulator , our framework bridges the gap between prediction and mechanism, offering a powerful tool for precision diagnostics, biomarker discovery, and the design of cytokine-targeted therapies in RA and other complex diseases.
BibTeX
@inproceedings{chen2026genecytnet,
  author = {C. L. Philip Chen and Dagang Li and Lujia Xu},
  title = {{GeneCytNet: an interpretable deep learning framework for rheumatoid arthritis classification and in silico cytokine perturbation modeling}},
  booktitle = {Frontiers in Immunology},
  year = {2026},
  doi = {10.3389/fimmu.2026.1738625},
}
Figure