Mi3-GPU: MCMC-based Inverse Ising Inference on GPUs for protein covariation analysis

Area of Science:

• Computational Biology
• Statistical Physics
• Biophysics

Background:

• Inverse Ising inference models protein residue co-variation to understand their interactions.
• Accurate inference is crucial for applications in protein physics and sequence analysis.
• Existing methods often rely on approximations or are computationally intensive.

Purpose of the Study:

• Introduce Mi3-GPU software for efficient and accurate inverse Ising inference on protein sequence datasets.
• Address limitations of analytic approximations and finite-sampling issues in Multiple Sequence Alignments (MSAs).
• Develop a generative modeling approach to validate inferred models.

Main Methods:

• Parallel Markov-Chain Monte-Carlo (MCMC) sampling optimized for Graphics Processing Units (GPUs).
• Quasi-Newton parameter-update scheme incorporating Zwanzig reweighting.
• Tools for Multiple Sequence Alignment (MSA) preparation and analysis to mitigate finite-sampling errors.

Main Results:

• Mi3-GPU enables precise reproduction of MSA covariation patterns, surpassing approximate methods.
• The software generates synthetic MSAs that match observed statistics within finite-sampling limits.
• Demonstrated ability to handle large datasets (L ~ 300, 21 residue types) with millions of parameters in short runtimes.

Conclusions:

• Mi3-GPU provides a highly accurate and efficient computational tool for inverse Ising inference in protein science.
• The generative modeling capability allows for robust validation of inferred protein interaction models.
• Accelerated GPU-based MCMC sampling combined with advanced parameter updates significantly advances the field.