Foundation Model for Biological Temporal Data Dynamics with Experimental Validation.

Area of Science:

• Computational Biology
• Dynamical Systems
• Machine Learning

Background:

• High-dimensional biological and environmental time series data are often noisy, incomplete, and heterogeneous.
• Learning stable continuous-time models for analysis and intervention is challenging.
• Existing models struggle with cross-dataset transferability and diverse downstream tasks.

Purpose of the Study:

• To introduce a reusable latent-dynamics backbone for robust temporal modeling.
• To demonstrate the model's effectiveness across heterogeneous biological and environmental datasets.
• To unify forecasting, adaptation, counterfactual analysis, and interpretable AI in time series analysis.

Main Methods:

• Coupling a mask-aware variational autoencoder with a latent neural ordinary differential equation.
• Developing a latent-dynamics backbone as a transferable foundation model for temporal data.
• Evaluating the framework on electroencephalography (EEG), air quality, and gene-expression datasets.

Main Results:

• The backbone improved open-loop forecasting and supported counterfactual rollouts on EEG and air quality data.
• It enabled data-efficient subject adaptation in EEG and interpretable intervention screening in air quality.
• In Drosophila gene expression, it facilitated mechanistic model supervision, yielding regulatory insights.

Conclusions:

• A shared latent-dynamics backbone can unify diverse temporal modeling tasks, including forecasting, adaptation, and interpretable AI.
• This approach offers a transferable solution for analyzing heterogeneous biological and environmental time series.
• The framework advances mechanistic analysis and intervention strategies through interpretable AI.