LMProtein: a protein language model based framework for protein structural property prediction.

Area of Science:

• Computational biology
• Machine learning in structural biology

Background:

• Machine learning and deep language models advance protein structure prediction.
• Current methods often rely on Multiple Sequence Alignments (MSAs), which are computationally intensive and fail for proteins without homologs.

Purpose of the Study:

• To develop a fast and accurate framework (LMProtein) for predicting protein structural properties using only the primary sequence.
• To overcome limitations of MSA-dependent methods.

Main Methods:

• LMProtein combines the unsupervised pretrained language model ESM-2 with Convolutional Neural Networks (CNNs), Long Short-Term Memory networks (LSTMs), and Multilayer Perceptrons (MLPs).
• The framework predicts secondary structure, dihedral angles, fluorescence, and stability landscapes.

Main Results:

• LMProtein outperforms recent MSA-based and single-sequence models.
• Achieved ~74% accuracy for eight-state secondary structure (SS8) prediction.
• Obtained mean absolute errors of 19° (Phi) and 29° (Psi) for dihedral angles.
• Spearman's correlation coefficients of 0.69 for fluorescence and 0.79 for stability.

Conclusions:

• LMProtein offers a computationally efficient and accurate alternative for predicting protein structural properties.
• The framework has significant potential for accelerating protein engineering and drug target identification, especially for proteins lacking homologous sequences.