EvoRator2: Predicting Site-specific Amino Acid Substitutions Based on Protein Structural Information Using Deep...

Area of Science:

• Computational biology
• Structural biology
• Molecular evolution

Background:

• Multiple sequence alignments (MSAs) infer tolerated amino acids but are limited for proteins with few homologs.
• Orphan and de novo designed proteins lack sufficient homologous sequences for traditional MSA-based analysis.

Purpose of the Study:

• To develop a deep learning algorithm (EvoRator2) that predicts tolerated amino acids using only protein structural information.
• To address the challenge of analyzing proteins with limited or no homologous sequences.

Main Methods:

• EvoRator2 was trained on over 15,000 protein structures.
• The algorithm predicts tolerated amino acids based on structural information from atomic coordinate files.
• Performance was evaluated using position-weighted scoring matrices (PSSM) and deep mutation scanning (DMS) experiments.

Main Results:

• EvoRator2 showed satisfying results in predicting PSSMs.
• It achieved near state-of-the-art performance in predicting mutation effects in DMS experiments, outperforming existing methods on certain targets.
• Combining EvoRator2 with MSA-based methods improved prediction accuracy and stability for DMS experiments.

Conclusions:

• EvoRator2 effectively predicts tolerated amino acid substitutions using protein structure alone.
• The tool is valuable for studying orphan and de novo designed proteins.
• The EvoRator web server is available for broader application.