Protein modeling with hybrid Hidden Markov Model/neural network architectures

Area of Science:

• Computational molecular biology
• Bioinformatics
• Machine learning in biology

Background:

• Hidden Markov Models (HMMs) are widely used for modeling and aligning protein families in computational molecular biology.
• First-order HMMs face limitations including numerous unstructured parameters and difficulty handling long-range dependencies.

Purpose of the Study:

• To introduce and evaluate hybrid Hidden Markov Model/Neural Network (HMM/NN) architectures for improved protein family modeling.
• To address the limitations of traditional HMMs by leveraging neural networks for parameter computation and constraint incorporation.

Main Methods:

• Developed hybrid HMM/NN architectures where neural networks compute HMM parameters, enabling flexible model complexity and constraint integration.
• Tested the hybrid approach on the immunoglobulin protein family, training a model and deriving a multiple alignment.
• Explored larger hybrid model classes using multiple HMMs modulated by NNs to capture complex dependencies.

Main Results:

• The hybrid HMM/NN model achieved multiple alignment of the immunoglobulin family using less than a quarter of the parameters required by previous single HMMs.
• Demonstrated the effectiveness of neural network reparametrization for reducing model complexity.
• Showcased the capability of larger hybrid models to capture dependencies by modulating HMM parameters via NNs.

Conclusions:

• Hybrid HMM/NN architectures offer a powerful and parameter-efficient alternative to traditional HMMs for protein family modeling and alignment.
• This approach significantly enhances the ability to model complex biological data, particularly for capturing long-range dependencies.
• The integration of neural networks provides a flexible framework for advancing computational biology tools.