scMix: learning temporal dynamics of gene expression under irregular time intervals.

Area of Science:

• Computational Biology
• Genomics
• Developmental Biology

Background:

• Temporal gene expression is crucial for understanding cellular development and differentiation.
• Single-cell datasets often have limited, unevenly spaced time points, posing challenges for predictive modeling.
• Existing methods struggle with error accumulation due to sequential prediction and lack of explicit time interval modeling.

Purpose of the Study:

• To introduce scMix, a language-model-based framework for predicting single-cell gene expression.
• To enable prediction from multiple historical time points and explicitly model time intervals.
• To improve the accuracy and robustness of temporal gene expression prediction.

Main Methods:

• Utilizing the Receptance Weighted Key Value (RWKV) architecture with a time-decay mechanism.
• Implementing a delta-time mechanism to bypass unmeasured time points and mitigate error accumulation.
• Incorporating a trend regularization strategy for enhanced temporal coherence of gene expression trajectories.

Main Results:

• scMix achieves state-of-the-art performance in predicting gene expression at unmeasured time points.
• The framework demonstrates superior accuracy and robustness compared to existing methods.
• scMix yields outstanding results on downstream tasks, validating its predictive power.

Conclusions:

• scMix offers a robust and accurate solution for modeling temporal single-cell gene expression, even with irregular time sampling.
• The novel delta-time mechanism and trend regularization significantly improve predictive performance.
• This framework advances the analysis of dynamic biological processes from single-cell data.