deepFPlearn

deepFPlearn predicts associations between chemical structures and biological effects at the gene and pathway level.

Key Features:

• Deep learning architecture: Combines a deep autoencoder and a deep feedforward neural network to compress molecular-structure representations and predict chemical–biological associations.
• Autoencoder-based feature reduction: Uses a deep autoencoder to reduce dimensionality of molecular structure data while preserving essential chemical information.
• Large unlabeled chemical inventory: Leverages a vast inventory of unlabeled chemical data to capture a wide range of chemical structures.
• High prediction quality: Demonstrates robust prediction of meaningful and experimentally verified chemical–biological associations on unseen data.
• Efficiency and scalability: Classifies hundreds of thousands of chemicals in seconds to enable large-scale assessments.
• Customizable and retrainable: Supports model customization and retraining for different application settings.

Scientific Applications:

• Toxicology: Predicts chemical–gene and chemical–pathway interactions to support hazard identification and mechanistic toxicology.
• Pharmacology: Aids prediction of target effects and off-target interactions for drug safety assessment at the gene/pathway level.
• Environmental risk assessment: Enables screening and prioritization of environmental chemicals to inform risk assessment and regulatory decision-making.

Methodology:

A deep autoencoder performs feature reduction on molecular-structure representations, followed by a deep feedforward neural network that predicts chemical–gene/pathway associations using training that leverages large unlabeled chemical datasets.