FR-BINN: Biologically Informed Neural Networks for Enhanced Biomarker Discovery and Pathway Analysis

Abstract

Chronic inflammation plays a pivotal role in human health, with certain inflammatory conditions significantly increasing the risk of cancer, while others do not. However, the molecular mechanisms underlying this divergent risk remain poorly understood. In this study, we propose FR-BINN, a biologically informed neural network framework for disease prediction and interpretability. Incorporating Fenton reaction (FR)-related biological priors and leveraging multiple interpretability methods, FR-BINN identifies key genes driving cancer-prone and non-cancer-prone chronic inflammatory diseases. The experimental results demonstrate that FR-BINN achieves superior classification performance while offering biologically interpretable insights. Moreover, attribution results derived from different explainable techniques show high consistency, and intra-method results exhibit distinct patterns across disease categories. We further combine large language models with feature attributions to identify candidate biomarkers, and independent datasets confirm the robustness of these findings. Notably, genes such as NCOA1 and SDHB are identified as being associated with cancer susceptibility. The framework further reveals distinct patterns in energy metabolism, oxidative stress, and pH regulation between cancer-prone and non-cancer-prone inflammatory diseases. These insights enhance our understanding of inflammation-associated tumorigenesis and contribute to the identification of potential biomarkers and therapeutic targets.

Keywords: biologically informed neural network; biomarker; chronic inflammation; explainable artificial intelligence.

Figure 1

Overview of FR-BINN framework. (A) Integration of transcriptomic datasets, disease category definitions (derived from statistical indicators), and FR-associated biological prior knowledge. (B) Biologically informed neural network encoding hierarchical FR-related knowledge for classification. The framework further provides multi-level explanations (e.g., key genes, pathways) and utilizes a LLM for semantic reasoning and interpretation.

Figure 2

Performance evaluation. (A) Disease categorization into CP-CIDs and NCP-CIDs based on epidemiological statistical thresholds. (B) Comparative classification performance of FR-BINN versus five baseline models across categories. (C) Density distributions of model-predicted probabilities aligned with clinical disease stages.

Figure 3

Validation of attribution methods. (A) Top attributed gene identification: Top 10 genes identified by the IG method that distinguish CP-CIDs from NCP-CIDs. (B) Expression validation: Gene expression heatmap illustrating the discriminatory power of the union of the top 10 genes identified by both IG and SV methods for cancer-prone versus non-cancer-prone inflammatory diseases. (C) Attribution method robustness: High concordance between IG and SV attribution results for top-ranked genes across both two inflammatory disease categories, indicating the robustness of the identified features. (D) Performance of the logistic regression model with refined genes. (E) Predictive accuracy on independent datasets.

Figure 4

Analysis of key attributed genes. (A) Top 10 candidate genes ranked by combined IG and SV attribution scores. (B) Venn diagram illustrating, in the top 10 candidate genes, the overlap between 7 causal genes and 4 genes encoding proteins related to hydrogen ion production. (C) Gene expression level of NCOA1 in CP-CIDs, NCP-CIDs, diseases, and control (********** denotes p = $1.47\times {10}^{-32}$). (D) NCOA1 expression profiles across various cancer types of TCGA. (E) Kaplan–Meier survival analysis for NCOA1 in LIHC.

Figure 5

Pathway-level analysis. (A) Patterns in CP-CIDs and NCP-CIDs. (B) GO enrichment analysis. (C) Energy metabolism pathway enrichment results of two categories. (D) GSVA scores of the energy metabolism pathways. (E) Mitochondrial and cytosolic protein damage.