Structure-free drug-target affinity prediction using protein and molecule language models.

Area of Science:

• Computational chemistry and cheminformatics
• Bioinformatics and computational biology
• Artificial intelligence in drug discovery

Background:

• Accurate drug-target affinity (DTA) prediction is vital for efficient drug discovery.
• Traditional DTA methods often depend on handcrafted features or structural data, limiting their scope.
• Developing scalable and generalizable DTA prediction models remains a significant challenge.

Purpose of the Study:

• To develop a novel, sequence-centric approach for DTA prediction.
• To leverage pretrained large language models (LLMs) for encoding protein and molecule sequences.
• To improve DTA prediction accuracy and scalability without relying on structural information.

Main Methods:

• Utilized pretrained LLMs (ChemBERTa and ESM2) to generate sequence embeddings for proteins and molecules.
• Developed a customized Residual Inception architecture for integrating sequence embeddings.
• Employed multi-scale convolutions and residual connections for efficient feature integration.
• Evaluated the model on benchmark datasets: Davis, KIBA, and BindingDB.

Main Results:

• Achieved state-of-the-art performance on benchmark DTA datasets.
• Reported Mean Squared Error (MSE) of 0.182 and Concordance Index (CI) of 0.915 on Davis.
• Reported MSE of 0.135 and CI of 0.902 on KIBA.
• Reported MSE of 0.467 and CI of 0.888 on BindingDB.

Conclusions:

• Sequence-based DTA prediction using LLMs offers a scalable and accurate alternative to structure-based methods.
• The proposed method provides robust DTA predictions, even in data-sparse scenarios.
• This approach enhances drug discovery pipelines by offering efficient and adaptable DTA prediction frameworks.