End-to-end learning of multiple sequence alignments with differentiable Smith-Waterman

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning

Background:

• Multiple sequence alignments (MSAs) are crucial for understanding sequence-structure-function relationships and evolutionary histories.
• Current MSA generation is often a separate step, lacking integration with downstream applications like structure prediction.

Purpose of the Study:

• To develop a method for jointly learning MSAs and downstream machine learning systems in an end-to-end manner.
• To introduce SMURF (Smooth Markov Unaligned Random Field) for unsupervised contact prediction using learned alignments.
• To demonstrate the utility of differentiable alignment modules in improving protein structure prediction pipelines.

Main Methods:

• Implemented a smooth and differentiable version of the Smith-Waterman algorithm for joint MSA learning.
• Developed SMURF, integrating differentiable alignment with Markov Random Fields for contact prediction.
• Connected the differentiable alignment module to AlphaFold2 to optimize MSAs for structure prediction.

Main Results:

• SMURF achieved mild improvements in contact prediction for protein and RNA families.
• Jointly learned MSAs improved AlphaFold2 structure predictions compared to initial MSAs.
• Alignments that enhanced structure prediction were found to be self-inconsistent, suggesting adversarial properties.

Conclusions:

• Differentiable dynamic programming offers a powerful approach to enhance neural network pipelines reliant on sequence alignments.
• Optimizing protein sequence predictions using current methods requires careful consideration due to potential adversarial effects.
• The developed method facilitates end-to-end learning for MSA-dependent biological applications.