There are different ways to represent a chemical molecule in digital formats, like SMILES for 1D text formulas or 2D graphs. There is also a way to capture 3D nature of molecules by density tensors representations[1]. A molecule is placed on a virtual grid, and the density of atoms is calculated within the grid's nodes. Thus space is divided into 3D voxels -- analogous to 2D pixels.
While there is software for generating voxels (like libmolgrid and moleculekit), the inverse problem of inferring molecular structure by its density tensor has no analytic solution. There were attempts to solve it by neural networks[1] or optimization procedure[3], but it works in a complex pipeline only. The tensor_to_structure is the python library that solves the problem of inferring molecular structure by its density tensor.
The algorithm is computationally heavy, so I prepared two versions. The tensor_to_structure/v1 is a general-purpose library with minimal assumptions about the problem. The tensor_to_structure/v2 library was developed for a gaussian mixture model. Use the second only if you model an atom's density by a normal distribution, otherwise use tensor_to_structure/v1.
Warning: It didn't intend to be published. There are tutorials, and the code is documented, but I shared it for myself only and won't maintain it.
- Ragoza, M., Hochuli, J., Idrobo, E., Sunseri, J., & Koes, D. R. (2017). Protein–ligand scoring with convolutional neural networks. Journal of chemical information and modeling, 57(4), 942-957.
- Ragoza, M., Masuda, T., & Koes, D. R. (2020). Learning a Continuous Representation of 3D Molecular Structures with Deep Generative Models. arXiv preprint arXiv:2010.08687.