Decoding of Inconsistent Biological Data: A Critical Step toward Enhanced AI Predictivity in Drug Discovery.

Area of Science:

• Computational chemistry
• Drug discovery
• Biochemistry

Background:

• Combining bioactivity data from diverse sources introduces noise for machine learning (ML) models.
• Assay protocol variations (buffer composition, experimental setup) significantly impact data reliability.
• Protein targets like enzymes and viral surface proteins exhibit conformational changes influenced by extrinsic factors.

Purpose of the Study:

• To highlight the impact of assay protocol variations on bioactivity data.
• To discuss strategies for mitigating noise in computational drug discovery.
• To explore the potential of deep learning (DL) and large language models (LLMs) in addressing these challenges.

Main Methods:

• Analysis of assay protocol variations and their effect on protein targets.
• Review of strategies for handling enzyme inhibitors/binders data.
• Discussion on the utility of deep learning (DL) models.
• Exploration of current limitations in computational protein-ligand interaction studies.
• Interview with an expert on large language models (LLMs) and agentic AI.

Main Results:

• Assay protocol differences are a major source of noise in bioactivity datasets.
• Conformational changes in enzymes and viral proteins complicate data integration.
• Deep learning (DL) models show promise but have limitations.
• LLMs and agentic AI offer potential advancements for drug discovery.

Conclusions:

• Standardizing assay protocols is essential for reliable ML model training in drug discovery.
• Advanced computational methods, including DL and LLMs, are needed to overcome data heterogeneity.
• Further research is required to fully leverage AI in predicting protein-ligand interactions.