A hitchhiker's guide to deep chemical language processing for bioactivity prediction.

Area of Science:

• Computational chemistry
• cheminformatics
• artificial intelligence in drug discovery

Background:

• Deep learning accelerates drug discovery using chemical language processing (CLP).
• CLP models learn from molecular string representations like Simplified Molecular Input Line Entry Systems (SMILES) and Self-Referencing Embedded Strings (SELFIES).
• Training predictive CLP models involves complex methodological considerations.

Purpose of the Study:

• To analyze key elements of chemical language processing (CLP).
• To provide guidelines for training predictive CLP models for drug discovery.
• To offer practical recommendations for researchers, from newcomers to experts.

Main Methods:

• Evaluation of three neural network architectures.
• Comparison of two molecular string representations (SMILES and SELFIES).
• Assessment of three embedding strategies across ten bioactivity datasets for classification and regression tasks.

Main Results:

• Identification of critical methodological decisions in CLP model training.
• Demonstration of the impact of different architectures, representations, and embedding strategies.
• Highlighting the importance of hyperparameter optimization for predictive accuracy.

Conclusions:

• Provides a comprehensive 'hitchhiker's guide' to CLP for drug discovery.
• Equips researchers with practical recommendations for selecting optimal methods.
• Aims to improve the efficiency and effectiveness of deep learning in accelerating drug discovery.