Data-driven discovery of digital twins in biomedical research

Area of Science:

• Systems Biology
• Computational Biology
• Biomedical Informatics

Background:

• High-throughput biological data enables digital twin development for personalized medicine.
• Current digital twin creation relies on manual data integration, necessitating automated methods.
• Biology presents unique challenges for data-driven discovery compared to physics.

Purpose of the Study:

• To review and evaluate methodologies for automated inference of biological digital twins from time-series data.
• To assess algorithms against key biological and methodological challenges.
• To propose future directions for developing robust digital twin inference methods.

Main Methods:

• Systematic review of 177 methodologies for inferring digital twins.
• Evaluation of algorithms based on eight challenges: data integration, multiple conditions, prior knowledge, latent variables, high dimensionality, unobserved variable derivatives, library design, and uncertainty quantification.
• Comparison of sparse regression, symbolic regression, deep learning, and large language models.

Main Results:

• Sparse regression generally outperformed symbolic regression, especially with Bayesian frameworks.
• Deep learning and large language models show potential for integrating prior knowledge but require reliability improvements.
• No single method currently addresses all challenges effectively.

Conclusions:

• Hybrid and modular frameworks are essential for advancing digital twin development.
• Future progress requires combining mechanistic grounding (chemical reaction networks), Bayesian uncertainty quantification, and deep learning's generative capabilities.
• Standardized benchmarks are needed to evaluate methods across all identified challenges.