A natural language processing approach based on embedding deep learning from heterogeneous compounds for quantitative...

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning in drug discovery

Background:

• Advancements in biological and chemical technologies generate vast datasets requiring sophisticated analytical methods.
• Deep learning models are increasingly applied in computer-aided drug discovery due to their success in diverse fields.

Purpose of the Study:

• To propose a natural language processing (NLP) approach for quantitative structure-activity relationship (QSAR) modeling.
• To transform molecular representations into semantic vectors for toxicity prediction.

Main Methods:

• Utilized deep neural networks for natural language processing to convert Simplified Molecular Input Line Entry System (SMILES) strings into word embedding vectors.
• Employed supervised machine learning algorithms, including convolutional long short-term memory neural networks, support vector machines, and random forests, for QSAR model development.
• Applied the models to predict toxicity data, specifically for the ciliate Tetrahymena pyriformis (IGC50) and rat median lethal dose (LD50).

Main Results:

• The NLP-based deep learning approach demonstrated efficient prediction of chemical compound activities on toxicity datasets.
• Quantitative structure-activity relationship models were successfully built using transformed molecular representations.

Conclusions:

• The proposed NLP and deep learning strategy offers an effective method for predicting chemical compound activities, particularly toxicity.
• This approach facilitates efficient computer-aided drug discovery by leveraging semantic representations of molecular structures.