https://paperswithcode.com/paper/a-3d-molecule-generative-model-for-structure

• problem statement: generating molecules that bind to specific protein binding sites
  • existing methods of drug design are mostyly string / graph based
  • limited by spatial information → unable to be applied to structure-based design tasks
    • models have no or little knowledge of how molecules interact with their target proteins in 3D space
  • searching for suitable molecule candidates in silico → massive computational efforts bc of economous space of synthetically feasible chemicals + conformational degree of freedom of compound and protein structures.
• proposol: 3D generative model that generates molecules given a designated 3D protein binding site
  • binding site = 3D context → model estimates the probability density of atom’s occurences in 3D space — positions that are more likely to have atoms will be assigned higher probability.
  • technically (high level)
    • to generate 3D molecules → an auto-regressive sampling scheme —atoms are sampled sequentially from the learned distribution until there is no room for new atoms
    • combined with this sampling scheme → model can generate valid and diverse molecules
      • applicble to various structre based molecular design tasks like molecule sampling and linker design → not applicable to generating models that fit to a specific protein structure (drug target)
• more technical details

  • machine learning approaches to problems in drug design → molecule generation

    • uses deep generative models to propose drug candidates by learning the underlying distribution of desiable molecules.

  • 3D generative model

    • modeling the distribution of atom occurance in 3D space of the binding site

    

    • In order to ensure the distribution is equivariant to C’s rotation and translation, they use rotationally invariant graph neural networks to perceive the context of each query coordinate.
    • creating valid + diverse molecules is still challnenging
      • drawing iid samples from the distrubution p(e,r|C) does not yield valid molecules bc atoms are depend on each toher
    • to tackle above problems → auto regressive sampling algorithm.
      • start with a context consisting of only protein atoms
      • iteratively sample one atom from the distribution at ach step + add to context used in the next step until theres no room for new atoms
        • simpler compared to other recent mehotds
        • no replying on post processing algoirhtms to infer atom placements from density
• the actual method yay

1. need to predict the probability of atom occurence in 3D space of the binding site
2. auto-regressive sampling algorithm for generating valid + multi modal molecules from the model
3. model learns to predict where should be placed and atoms and what type of atom should be placed

• 3D generative model design

  

  To generate atoms in the binding site, you need to mode the probability of atom occurring at some position r in the site. Formally, this is to model the density

  

  

  • context encoder

    • contextawareness: the representation of an atom should not only encode the property of the atom itself, but also encode its context.

    • **rotational and translational invariance:**the physical and biological properties of the system don’t change → the representations that reflect these properties should be invariant to rigid transforms as well.

    • how →

      • construct a k-nearest-neighbor graph based on inter-atomic distances, denoted as G = <C, A>, where A is the adjacency matrix.
      • We also denote the k-NN neighborhood of atom i as Nk(ri) for convenience. The context encoder will take G as input and output structure-aware node embeddings.

      

      • atoms are sampled sequentially. probability density changes as we place new atoms. sampling process naturally diverges and creates new samples.

• gaps (for me)

  • invariant graph neural networks?
  • autoregressive sampling strategies?