• analyzing structure and function of proteins is important → **designing proteins from primary-tertiary structure is hard, primarily because predicting and testing function and generating designs takes time (**also known as protein folding problem)
• Using GANS or generative models for generating protein structures for faster de novo protein design

"The paper encodes protein structures based on pairwise distances between alpha-carbons on protein backbone"

• Think of the alpha-carbons on the protein backbone as a molecule where we can compute the pairwise distances → prevents the model to learn translational and rotational symmetries
  • 𝛼-carbon is adjacent to "carbonyl" or C=O → distance is 3 carbons away
  • when we refer to backbone, it means the hydrogen bonds of the amino-acids and their respective 𝛼-carbon
  • the proteins are folded based on generated pairwise distance maps → solve this using Alternating Direction Method of Multipliers
• effectiveness/cost-function is used by predicint completions of corrupted protein structures and show that method is capable of quickly producing structured solutions
• to control structure and function of engineered proteins is key in practice create proteins de novo → finding new, novel folds and designs/elements that can be used for designing proteins is a fundamental, unsolved question
• folding proteins is dependent on heuristics → requires subjective expertise for optimizing design and improving computer responses
• deep generative modelling for fast generation of new, viable protein structures was done using GANs, tested models by predicting missing sections from corrupted/denatured protein structures (reversing denaturing effects)
• the data is represented based on pairwise distances of alpha-carbons (reduces the dimensionality of the data significantly), model performs and learns to generate new structures whil einfering solutinos for completing corrupted structures

Alternating Diection Method of Multipliers: algorithm used to solve convex optimization problems by breaking them down into smaller, more manageable problems

• algorithm used to fold 2D pairwise distances into 3D cartesian coordinates → this also reduces the complexity of orientation (rotation) and translation so the final product stays the same → the focus is to learn protein design and folding with mixes of generative and reinforcement learning models

  

  • Proteins are made of amino acid chains and other secondary units including alpha helices (folding shape) and beta sheets → pairwise distances of a-carbons between amino-acids in angstroms gives protein maps with atomic resolution
  • The DCGAN generates the pairwise map, while the ADDM folds the map into a 3D trace structure → protein backbone generated on top of the trace-out, Rosetta used to fold it into 3D structures