ToxDL

ToxDL predicts protein toxicity from amino acid sequences using convolutional neural networks and domain-specific embeddings to distinguish toxic versus non-toxic proteins.

Key Features:

• Deep Learning Architecture: A convolutional neural network (CNN) tailored for variable-length input sequences processes diverse protein sequences to extract predictive features.
• Domain Embeddings (domain2vec): A domain2vec module generates embeddings for protein domains to capture structural and functional characteristics.
• Toxicity Classification: An output module integrates CNN outputs and domain2vec embeddings to classify proteins as toxic or non-toxic.
• Performance Superiority: Independent testing on animal proteins and cross-species transferability assessments involving bacterial proteins showed performance superior to homology-based methods and other machine-learning techniques.
• Saliency Maps: Saliency maps visualize network-learned features and known toxic motifs and support directed in silico sequence modifications to alter predicted toxicity.

Scientific Applications:

• Crop safety assessment: Predicts potential toxicity of proteins introduced into genetically engineered food crops to inform safety evaluations.
• Therapeutic protein evaluation: Assesses toxicity risk of candidate therapeutic proteins during drug development.
• Regulatory and research support: Provides predictive evidence for GMO development and safety regulation decision-making.

Methodology:

The model is trained on protein sequences with a CNN that extracts sequence features, domain2vec that embeds domain-specific information, an output module that integrates these representations for toxic/non-toxic classification, and saliency maps to visualize learned motifs.