DCell

DCell models eukaryotic cell structure and function using a deep neural network whose visible hierarchical architecture maps to 2,526 cellular subsystems to enable genotype-to-phenotype simulation.

Key Features:

• Visible Neural Network (VNN): A deep neural network structured from cell biology knowledge that renders internal components interpretable as biological subsystems.
• Hierarchical structure: Embeds 2,526 subsystems representing cellular components and processes for multilevel simulation and analysis.
• Training on genotypes: Trained on several million genotypic data points to simulate cellular growth with accuracy comparable to laboratory observations.
• Genotype–phenotype associations: Infers patterns of subsystem activity from input genotypes to investigate molecular mechanisms underlying genotype-phenotype associations.
• Boolean mechanisms: Identifies unexpected mechanisms, including relationships governed by Boolean logic.
• Predictive component importance: Determines that approximately 80% of importance for predicting cell growth is captured by 484 subsystems (21%).

Scientific Applications:

• Decoding disease genetics: Simulates cellular processes to analyze the genetic basis of disease phenotypes.
• Drug resistance studies: Investigates cellular mechanisms underlying drug resistance.
• Synthetic biology: Predicts effects of genetic changes on cellular function to inform design in synthetic biology.

Methodology:

Implements a deep neural network as a visible neural network (VNN) with a hierarchical embedding of 2,526 subsystems; trained on several million genotypic data points; infers subsystem activity patterns from genotypes; detects Boolean-style regulatory relationships; quantifies subsystem importance for predicting cell growth (484 subsystems account for ~80% of importance).