Derivative-free neural network for optimizing the scoring functions associated with dynamic programming of...

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Profile-comparison methods using Position-Specific Scoring Matrices (PSSMs) are accurate alignment techniques.
• Current scoring functions (cosine similarity, correlation coefficients) may not be optimal as they fail to capture nonlinear relationships between PSSMs.
• A need exists for a more suitable scoring function for profile alignment methods.

Purpose of the Study:

• To discover a novel scoring function for profile comparison methods using neural networks.
• To develop a learning system capable of optimizing scoring functions in derivative-free problem settings.

Main Methods:

• Developed a novel derivative-free neural network by integrating a conventional neural network with an evolutionary strategy optimization method.
• Optimized a scoring function for aligning remote sequence pairs using this system.
• Implemented the novel scoring function into a pairwise-profile aligner.

Main Results:

• The novel scoring function significantly improved both alignment sensitivity and precision compared to existing functions.
• The developed aligner (Nepal) demonstrated enhanced adaptation to remote sequence alignments.
• The scoring function increased the expressiveness of similarity scores.

Conclusions:

• A novel derivative-free neural network and aligner (Nepal) were developed for optimizing sequence alignments.
• The novel scoring function can be easily integrated into other aligners and potentially improve homology detection and multiple-sequence alignment.
• The developed optimization method is useful for derivative-free problems common in practical applications.