Geometric Deep Learning for Structure-Based Drug Discovery

Humans and molecules are living in the same, three-dimensional world. Drug discovery and development is a great work of managing our internal and enemies’ (like virus’) external system by a precisely and optimally elaborated molecule. Molecule-molecule interaction like PPI needs to be analyzed in a three-dimensional world for deep recognition, hence prediction and design of the drug structure. Geometric deep learning,1) we would think, blazes a trail for drug discovery.

Deep learning (DL) and artificial intelligence (AI) have been producing paradigm shifts enormously in every industrial field and also our daily lives. Character recognition, voice recognition and image recognition are the representative technologies that have changed the way of our live. Generative AI is also a technology of invaluable impact. Everyone knows ChatGPT as an example of a readily-usable tool.

Current AI-based technologies are basically in the Euclid’s world. Euclidian data is relatively easy to process in machine learning. Text, voice and image are ready for one- or two-dimensional presentation. This is one of the reason these recognition systems were developed very rapidly. From the other perspective, drug design has been difficult just by applying the well-investigated Euclidian data-based models.

Geometric deep learning is DL techniques to expand and generalize deep neural models to non-Euclidian world, including graphs and manifolds, for example. Molecular systems match this group of DL/AI technologies because molecule-mole interactions are in the three-dimensional field. The technologies are enabling precise prediction of molecular property, binding site, recognition interface, binding pose generation, molecular docking as well as de novo design.

This review summarizes the current state of art of geometric deep learning from the viewpoint of concepts and methods.2) The author recognizes the incorporation of symmetry information (geometric priors) into neural network is the primary direction of the current geometric DL.

Euclidean E(3), SE (3) and permutation are generally-used symmetry groups in geometric DL. E(3) is the basic three-dimensional symmetry group that consists of translation, rotation and reflection. Reflection is excluded in the case of SE(3) and it enables the model to distinguish chiral molecules, or enantiomers.
The permutation is a symmetry group of node (atom) numbering on neural network, which allows replacing atom numbering.

The molecular representation has a big impact on the molecular feature because it controls the nature of symmetry. 3D grids, 3D surfaces and 3D graphs are often used in geometric DL, and the features are summarized below.

Representation Feature
3D grids voxel-based Euclidean data structure translated into voxels in 3D
3D surfaces mesh-based 3D arrangement of coordinates (polygons)
3D graphs node-based non-Euclidean data structure of nodes and edges

These representations could solve any of structure-based drug discovery issues. Representative examples were shown in every category in three ways in the review. We don’t get into each research result here, but it is still a novel application of 3D-oriented DL and the golden rule has not been established yet.

Comprehensive evaluation and comparison of current models for every drug discovery task could cultivate a fast and reliable path for the candidate. It is because geometric DL models are aware of third dimension, like us.

Geometric DL’s concept is to approach humans in the context of spatial recognition. It also hits humans’ instincts for molecule-molecule interactions. Generally usable framework in developing3) and there would be a huge growth in this field in a couple of years.

We are eager to seek the opportunity for emerging and novel DL approaches. Please contact us to figure out the way to generate a synergy between you and us.


Scroll to top