Comprehensive Research on Druggable Proteins: From PSSM to Pre-Trained Language Models.

Area of Science:

• Computational biology
• Drug discovery
• Bioinformatics

Background:

• Identifying druggable proteins is crucial for cost-effective drug discovery.
• Traditional experimental methods are slow, expensive, and labor-intensive for large-scale screening.
• Computational methods offer efficient alternatives for predicting protein druggability.

Purpose of the Study:

• To develop a fast and precise computational classifier for identifying druggable proteins.
• To evaluate the performance of protein language model (PLM) embeddings against traditional features.
• To explore the application of large language models (LLMs) for druggable protein recognition.

Main Methods:

• Utilized a protein language model (PLM) with fine-tuned Evolutionary Scale Modeling 2 (ESM-2) embeddings for classification.
• Compared the predictive performance of ESM-2 embeddings with Position-Specific Scoring Matrix (PSSM) features.
• Developed and tested an end-to-end model based on modified Generative Pre-trained Transformers 2 (GPT-2).
• Validated the models using a benchmark dataset and the Pharos dataset.

Main Results:

• Achieved 95.11% accuracy in identifying druggable proteins using ESM-2 embeddings.
• ESM-2 embeddings demonstrated superior accuracy and efficiency compared to PSSM features.
• The study represents the first deployment of the GPT-2 LLM for druggable protein recognition.

Conclusions:

• Protein language models, particularly ESM-2 embeddings, offer a highly accurate and efficient approach for identifying druggable proteins.
• The developed computational models can significantly accelerate the early stages of drug discovery.
• LLMs like GPT-2 show promise for advancing protein function prediction and drug target identification.