HiMolformer: Integrating graph and sequence representations for predicting liver microsome stability with SMILES

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Accurate prediction of metabolic stability is critical for early-stage drug discovery.
• Existing deep learning models often rely on single data formats (2D graphs or 1D sequences), limiting performance.
• Integrating diverse molecular representations can enhance predictive accuracy.

Purpose of the Study:

• To develop a novel hybrid deep learning model for predicting metabolic stability.
• To combine the strengths of graph-based and sequence-based molecular representations.
• To establish a new benchmark for predicting metabolic stability using a single input format.

Main Methods:

• Developed a hybrid model, HiMolformer, integrating a graph neural network (HiMol) and a Transformer model (Molformer).
• Utilized pre-trained models for molecular feature extraction from both 2D topological and 1D sequential data.
• Applied a regression task for prediction using a dataset of 3,498 molecules from the Korea Chemical Bank (KCB).

Main Results:

• HiMolformer demonstrated superior predictive performance compared to existing models.
• The model successfully predicted metabolic stability using mouse liver microsome (MLM) and human liver microsome (HLM) data.
• This represents the first reported instance of MLM and HLM prediction models using regression with a single SMILES input.

Conclusions:

• The hybrid HiMolformer model effectively leverages combined molecular representations for enhanced metabolic stability prediction.
• This approach offers a promising strategy for improving early-stage drug discovery pipelines.
• The open-source availability of the code facilitates further research and application.