ToxDL: deep learning using primary structure and domain embeddings for assessing protein toxicity.

Area of Science:

• Biotechnology
• Bioinformatics
• Computational Biology

Background:

• Assessing the toxicity of engineered proteins is crucial for food safety and therapeutic development.
• Current methods for protein toxicity assessment often rely on traditional homology-based approaches.

Purpose of the Study:

• To develop a deep learning-based approach for predicting protein toxicity solely from amino acid sequences.
• To create a tool that can identify toxic protein motifs and guide sequence modification for reduced toxicity.

Main Methods:

• Developed ToxDL, a deep learning model integrating a convolutional neural network for sequence analysis and a domain2vec module for protein domain embeddings.
• Trained and tested the model on animal and bacterial protein datasets to evaluate its predictive performance.
• Utilized saliency maps for visualizing learned toxic motifs and enabling directed sequence modification.

Main Results:

• ToxDL accurately predicts protein toxicity from sequence data, outperforming traditional homology-based methods and current machine learning techniques.
• The model demonstrated strong performance in cross-species transferability tests for bacterial proteins.
• Saliency map visualizations confirmed that ToxDL learns known toxic motifs and can guide sequence alterations.

Conclusions:

• ToxDL provides an effective in silico method for predicting protein toxicity, essential for safety assessments in biotechnology and drug development.
• The tool's ability to identify and modify toxic motifs offers a novel approach to engineering safer proteins.