ETLD: an encoder-transformation layer-decoder architecture for protein contact and mutation effects prediction.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning in Biology

Background:

• Multiple sequence alignments (MSAs) of homologous proteins contain latent features crucial for understanding protein function and evolution.
• Machine learning, both supervised and unsupervised, has been extensively applied to extract insights from MSAs over the past 30 years.
• Existing methods often require significant feature engineering or lack interpretability.

Purpose of the Study:

• To introduce a novel self-supervised model, the encoder-transformation layer-decoder (ETLD) architecture.
• To capture latent features directly from MSAs for predicting protein properties.
• To improve upon existing machine learning models for analyzing protein sequence data.

Main Methods:

• Developed the encoder-transformation layer-decoder (ETLD) architecture, a self-supervised model.
• ETLD incorporates a transformation layer to learn inter-site couplings within MSAs.
• The model encodes and decodes protein sequences, predicting amino acid probabilities at each site.

Main Results:

• ETLD successfully captures latent protein sequence features directly from MSAs.
• The model can derive residue-residue contact maps and predict mutation effects.
• ETLD demonstrated superior performance compared to established models like GEMME, DeepSequence, and EVmutation in mutation effect prediction.

Conclusions:

• ETLD is a highly interpretable, unsupervised model for analyzing protein MSAs.
• The architecture shows significant potential for further development and combination with supervised methods.
• ETLD offers a powerful new tool for predicting protein residue contacts and mutation effects.