Feature engineered embeddings for classification of molecular data.

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning

Background:

• Molecule classification is crucial for drug discovery.
• Current deep learning methods using structural data are computationally intensive and slow to train.
• Molecular data exists as both structural and sequence/text information.

Purpose of the Study:

• To develop a computationally efficient and reproducible method for molecule classification.
• To explore the use of natural language processing (NLP) techniques for feature engineering molecular text data.
• To create fast, chemical text-data-dependent molecular embeddings for machine learning models.

Main Methods:

• Feature engineering using NLP techniques: count vectorisation, term frequency-inverse document frequency (TF-IDF), word2vec, and Latent Dirichlet Allocation (LDA).
• Application of these techniques to FASTA sequence data and Simplified Molecular Input Line Entry Specification (SMILES) data.
• Evaluation of the generated embeddings' performance in machine learning classification tasks.

Main Results:

• NLP-based feature engineering created robust and reproducible molecular embeddings.
• The developed embeddings are fast to implement and solely dependent on chemical text data.
• These embeddings demonstrated excellent performance for molecule classification tasks when applied to FASTA and SMILES data.

Conclusions:

• Natural language processing techniques offer an efficient alternative to traditional deep learning for molecule classification.
• The proposed method provides a fast, computationally inexpensive, and effective approach for generating molecular embeddings from text data.
• This approach has significant implications for accelerating drug discovery and other applications requiring molecule classification.