ToxMVA: An end-to-end multi-view deep autoencoder method for protein toxicity prediction

Area of Science:

• Biotechnology
• Computational Biology
• Drug Discovery

Background:

• Protein toxicity prediction is vital for efficient drug discovery, aiding in screening and cost reduction.
• Existing machine learning methods show promise but can be hindered by direct feature concatenation, potentially introducing noise.
• There is a need for advanced methods to integrate diverse protein features effectively for improved toxicity prediction.

Purpose of the Study:

• To develop a novel end-to-end deep learning method, ToxMVA, for accurate protein toxicity prediction.
• To address limitations of previous methods by intelligently integrating multi-view protein features.
• To enhance the efficiency and reduce the cost of early-stage drug discovery.

Main Methods:

• Construct comprehensive protein feature profiles from primary sequences, encompassing sequential, physicochemical, and contextual semantic information.
• Employ an autoencoder network to integrate these multi-view features, generating a concise and accurate representation.
• Utilize an end-to-end deep learning architecture for toxicity prediction.

Main Results:

• ToxMVA demonstrated superior performance in protein toxicity prediction across three independent datasets.
• The method showed enhanced robustness compared to existing approaches.
• The integrated feature representation proved more effective than simple concatenation.

Conclusions:

• ToxMVA offers a significant advancement in predicting protein toxicity.
• The autoencoder-based multi-view integration effectively overcomes limitations of previous feature handling methods.
• This approach holds substantial potential for accelerating drug discovery and development pipelines.