Predicting rare DNA conformations via dynamical graphical models: a case study of the B→A transition

Area of Science:

• Computational biology
• Biophysics
• Machine learning

Background:

• DNA conformation is crucial for biological functions like protein binding.
• Traditional methods (MD, Markov state models) face computational challenges in capturing rare DNA states.
• Understanding local DNA conformational preferences is key to predicting global structures.

Purpose of the Study:

• To develop a novel AI framework for predicting unseen DNA conformational transitions.
• To overcome the limitations of traditional methods in sampling rare states.
• To accurately predict sequence-dependent DNA structural preferences.

Main Methods:

• Developed a generative machine learning framework using dynamical graphical models (DGMs).
• Trained DGMs on equilibrium molecular dynamics (MD) data.
• Leveraged local DNA interactions to generate a comprehensive transition matrix.

Main Results:

• DGMs successfully predict DNA conformational transitions not observed in MD ensembles.
• The framework captures thermodynamic and kinetic properties of unsampled states.
• Accurate prediction of sequence-dependent A-DNA preferences, validated against umbrella sampling simulations.

Conclusions:

• DGMs offer an efficient method for predicting rare DNA conformations and sequence-structure relationships.
• This AI approach reduces the need for extensive computational sampling.
• DGMs have potential applications in DNA sequence design and optimization.