DE-EA: A double-encoded evolutionary algorithm for mining dynamical network biomarkers in higher-order interaction

Complex diseases often undergo abrupt transitions from healthy to pathological states, necessitating biomarkers that can signal these critical tipping points before irreversible damage occurs. While dynamic network biomarkers (DNBs) have emerged as promising early-warning indicators, existing approaches are confined to pairwise gene interactions-a simplification that overlooks the group-level cooperative behaviors pervasive in biological systems. Higher-order interactions (HOIs), where multiple genes jointly regulate cellular processes, offer a more realistic representation yet remain unexplored in DNB identification.

We present DE-EA, a double-encoded evolutionary algorithm that simultaneously identifies DNBs and their associated higher-order interaction structures from time-series transcriptomics data. By encoding both gene sets and hyperedge topologies within a unified chromosome framework, DE-EA navigates the combinatorial complexity of hypergraph search spaces through specialized crossover and mutation operators. Benchmarking across five independent GEO datasets demonstrates that DE-EA consistently outperforms pairwise-based methods in tipping-point detection accuracy, while yielding substantially smaller gene signatures that enhance interpretability for therapeutic targeting. Ablation studies confirm that explicit modeling of HOIs, rather than algorithmic tuning, drives these performance gains.

DE-EA establishes a computational paradigm for early disease warning that transcends pairwise network assumptions. By revealing how DNBs reorganize within higher-order structures during state transitions, our framework bridges the gap between hypergraph theory and precision medicine, offering a scalable tool for biomarker discovery and critical-point prediction in complex disease dynamics. The code for this method is available at: https://github.com/yhwangWH/DEEA.