Medt5-bi: bidirectional translation between drug indications and molecular structures using a chemically-aware...

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Bioinformatics

Background:

• Drug discovery is complex, costly, and time-consuming.
• Novel computational methods are needed to streamline the process.
• Mapping natural language indications to molecular structures is a key challenge.

Purpose of the Study:

• To develop a novel bidirectional transformer-based architecture for drug discovery.
• To seamlessly map natural-language drug indications to Simplified Molecular Input Line Entry System (SMILES) encoded molecular structures.
• To enhance molecular generation from textual descriptions.

Main Methods:

• Developed MedT5-Bi, a bidirectional transformer model.
• Integrated Molecule-Aware Embeddings (MAEmb) using MolEmbedder and Graph Neural Networks (GNN).
• Implemented a Dynamic Attention Mechanism (DAM) for adaptive attention.
• Fine-tuned the model using reinforcement learning (RL) with a composite reward function.

Main Results:

• MedT5-Bi outperforms state-of-the-art models by 16.6-24.8% on standard benchmarks.
• Achieved superior performance in BLEU, ROUGE, Levenshtein distance, Morgan/Tanimoto similarity, and Text2Mol metrics.
• Demonstrated improved chemical validity, structural similarity, and generalization.

Conclusions:

• The proposed MedT5-Bi architecture significantly enhances molecular generation from textual indications.
• This AI-driven approach offers a promising computational strategy to reduce drug discovery costs and timelines.
• The model's ability to integrate sequential and topological features is key to its success.