PaleAle 6.0: Prediction of Protein Relative Solvent Accessibility by Leveraging Pre-Trained Language Models (PLMs).

Area of Science:

• Computational biology
• Bioinformatics
• Structural biology

Background:

• Predicting protein relative solvent accessibility (RSA) is crucial for understanding protein structure and function.
• Deep learning, particularly natural language processing (NLP) integration, has advanced protein research.
• Homology transfer limitations necessitate novel approaches for RSA prediction.

Purpose of the Study:

• To leverage pre-trained language models (PLMs) for enhanced protein RSA prediction.
• To develop a deep neural network architecture for analyzing protein sequence interactions.
• To introduce PaleAle 6.0 as a predictor for real-value and discrete RSA classifications.

Main Methods:

• Utilized a deep neural network combining bidirectional recurrent neural networks and convolutional layers.
• Employed ESM-2 encoding for protein sequence analysis.
• Developed PaleAle 6.0 for predicting real-value, two-state, and four-state RSA.

Main Results:

• PaleAle 6.0 achieved over 82% accuracy for two-state RSA (RSA_2C) and 59.75% for four-state RSA (RSA_4C) on the 2022 test set.
• A Pearson correlation coefficient (PCC) of 77.88 was obtained for real-value RSA prediction on the 2022 test set.
• On the 2024 test set, PaleAle 6.0 demonstrated strong performance with 79.74% accuracy (RSA_2C), 55.30% (RSA_4C), and a PCC of 73.08, outperforming existing predictors.

Conclusions:

• PaleAle 6.0 effectively predicts protein RSA using PLMs and a hybrid deep learning architecture.
• The model shows robust performance across different RSA classification schemes and datasets.
• This work advances computational methods for protein structure prediction and functional analysis.