AI-Driven molecule generation and bioactivity prediction: A multi-model approach combining VAE, graph and...

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Oncology

Background:

• Cancer remains a leading global cause of death, with traditional drug discovery being lengthy and expensive.
• Deep learning (DL) is revolutionizing drug design and prediction by leveraging molecular data.
• Current methods face challenges in speed and cost-effectiveness for novel drug development.

Purpose of the Study:

• To explore and compare various deep learning models for anticancer small molecule design and bioactivity prediction.
• To present a novel multi-model integrating generative and predictive DL approaches.
• To enhance the efficiency and accuracy of anticancer drug discovery pipelines.

Main Methods:

• A Variation Autoencoder (VAE) model was fine-tuned to generate novel anticancer molecules.
• A meta-model employing averaging and stacking ensemble methods was developed for activity prediction.
• Graph Neural Networks (GNNs) including Graph Attention Networks (GATs), Graph Convolutional Networks (GCNs), and Message Passing Neural Networks (MPNNs) were utilized.
• A pre-trained ChemBERTa model based on the attention mechanism was incorporated.

Main Results:

• The multi-model demonstrated superior performance in predicting anticancer molecule activity against breast cancer cell lines.
• The stacking ensemble method achieved a Pearson's correlation coefficient of up to 83%.
• The generative VAE model successfully created new molecules with drug-like properties.

Conclusions:

• The proposed deep learning multi-model significantly advances anticancer drug discovery and prediction.
• This approach offers a faster and more cost-effective alternative to traditional drug development methods.
• The integration of VAEs, GNNs, and ensemble techniques shows great promise for future pharmaceutical research.