Deep Neural Networks for Predicting Single-Cell Responses and Probability Landscapes

Area of Science:

• Synthetic Biology
• Computational Biology
• Machine Learning

Background:

• Accurate prediction of cellular responses is crucial for engineering biology.
• Challenges include biochemical stochasticity, cell variability, and incomplete biological process knowledge.
• Machine learning (ML) offers a powerful approach to model complex biological systems without prior mechanistic insight.

Purpose of the Study:

• To explore the application of ML for predicting gene expression dynamics using time-series data.
• To computationally simulate single-cell responses under various noise conditions and genetic circuit designs.
• To evaluate and improve ML model performance for predicting cellular behavior.

Main Methods:

• Development and training of deep neural networks (DNNs) on simulated single-cell response data.
• Incorporation of noise sources (measurement noise, biochemical stochasticity) and diverse genetic circuit designs (cascaded, bistable auto-activation).
• Iterative refinement of network architecture to handle multimodal dynamics, predicting future state distributions.

Main Results:

• DNNs successfully inferred cellular response dynamics despite noise and stochasticity.
• Prediction quality was influenced by training set size and input data history length.
• An updated network architecture accurately predicted bimodal expression distributions for bistable circuits.

Conclusions:

• ML, particularly DNNs, provides a robust framework for predicting and controlling biological circuits in synthetic biology.
• The developed methods are adaptable to various prediction tasks in biological engineering.
• Addressing multimodal dynamics requires specialized network architectures capable of predicting state distributions.