Two sequence- and two structure-based ML models have learned different aspects of protein biochemistry.

Area of Science:

• Computational biology
• Protein engineering
• Machine learning in bioinformatics

Background:

• Deep learning models, including large language models (LLMs) and 3D Convolutional Neural Networks (CNNs), are increasingly used for predicting protein mutational effects.
• LLMs utilize transformer architectures on protein sequences, while 3D CNNs process voxelized protein structures.

Purpose of the Study:

• To systematically compare the predictive performance and generalization capabilities of sequence-based LLMs and structure-based 3D CNNs for protein mutations.
• To identify specific strengths and weaknesses of each model type in predicting amino acid properties and locations.

Main Methods:

• Comparison of two LLMs and two 3D CNNs on protein mutation prediction tasks.
• Analysis of prediction accuracy correlations between sequence-based and structure-based models.
• Evaluation of model performance on predicting different types of amino acid residues (buried vs. solvent-exposed, hydrophobic vs. polar/charged).

Main Results:

• Overall prediction accuracies between sequence- and structure-based models are largely uncorrelated, indicating distinct predictive capabilities.
• Structure-based 3D CNNs excel at predicting buried aliphatic and hydrophobic residues.
• Sequence-based LLMs demonstrate superior performance in predicting solvent-exposed polar and charged amino acids.

Conclusions:

• Different deep learning architectures (LLMs vs. 3D CNNs) capture distinct aspects of protein biochemistry and mutation effects.
• A hybrid approach combining predictions from sequence- and structure-based models significantly enhances overall prediction accuracy.
• Integrating diverse model predictions offers a promising strategy for advancing protein engineering and functional prediction.