Computationally predicting molecular shapes and structures plays an important role in drug discovery, as it enables the identification of drug properties and functionality without the need for initial experimentation. Many dimensional methods are available but are usually 1D or 2D based. A recent study has used a new 3D convolutional neural network, Drug3D-Net, to predict biochemical and molecular properties, outperforming current deep-learning models.
Accurately representing the 3D structure of a molecule is a constant struggle, with many dimensional data options available. Most researchers use 1D and 2D methods as these are less computationally demanding but they do result in the molecule conformation being ignored and a less accurate structural picture overall.
Conformation refers to the possible position of a molecule in space – think of how you would draw a Christmas tree, point up, this is the usual and most agreed upon conformation.
However, many molecules have infinite conformations due to rotating bonds and other factors. Therefore, determining which is the most likely needs to be considered. Like trying to fit a complicated puzzle piece into place when you don’t know what it or the puzzle looks like. This is where 3D structural information and 3D deep learning can come in.
1D to 2D to 3D
1D refers to the amino-acid sequence of a molecule, a string of letters from which you can infer the structure. However, it can’t consider sub-structures or the whole spatial context. Whereas 2D based methods, such as molecular fingerprinting, can convey molecular features and be used to extract a 3D model. Previous methods have combined different dimensional methods to great success.
3D-based methods, although more computationally demanding, can predict new binding sites or active molecules -which is unpredictable at best in 2D-based models.
(Figure 2 from Li et al., 2021) 2D – 3D -3D again of the same molecule.
Drug3D-Net: The new 3D Model
This study proposed Drug3D-Net with spatial-temporal gated attention module as a novel deep learning neural network, that represents a drug’s 3D structure based on its molecular geometry. The spatial-temporal gated attention module extracts the molecular geometry based on a 3D molecular grid descriptor. Whereas the Drug3D-Net architecture obtains the 3D molecular representation.
This two-part system outperformed current models in predicting molecular properties and biochemical activities. It was tested on over four molecular datasets: ESOL, FreeSolv, Tox21 and HIV.
This system can also be used to predict drug-drug interactions, which is important to understand as they can lead to serious adverse reactions. The researchers also suggested that it could be applied to larger molecules, such as proteins, whose structures are notoriously difficult to decipher – which has previously only been done from the 1D structure.
Although more computationally demanding, the 3D molecular structure holds much more information than that of the 2D or 1D versions. Models such as Drug3D-Net can be used to better predict drug properties, increasing the success rate of new drugs being developed. Although as with all models, it is only as good as the data we feed it.