Generative modelling aims to accelerate the discovery of novel chemicals by directly proposing with desirable properties. Recently, score-based, or diffusion, generative models have outperformed previous approaches. Key to their success is the close relationship between score and physical force, allowing the use of powerful equivariant neural networks. However, the of the learnt score is not yet well understood. Here, we analyse the score by training an based diffusion model for molecular generation. We find that during the generation the score a restorative potential initially and a quantum-mechanical force at the end. In between
the two endpoints, it exhibits special properties that enable the building of large molecules. Using from the trained model, we present Similarity-based Molecular Generation (SiMGen), a new for zero shot molecular generation. SiMGen combines a time-dependent similarity kernel descriptors from a pretrained machine learning force field to generate molecules without any training. Our approach allows full control over the molecular shape through point cloud and supports conditional generation. We also release an interactive web tool that allows users generate structures with SiMGen online (https://zndraw.icp.uni-stuttgart.de).
Description
[2402.08708] Zero Shot Molecular Generation via Similarity Kernels
%0 Generic
%1 elijosius24a
%A Elijošius, Rokas
%A Zills, Fabian
%A Batatia, Ilyes
%A Norwood, Sam Walton
%A Kovács, Dávid Péter
%A Holm, Christian
%A Csányi, Gábor
%D 2024
%K EXC2075
%R 10.48550/arXiv.2402.08708
%T Zero Shot Molecular Generation via Similarity Kernels
%X Generative modelling aims to accelerate the discovery of novel chemicals by directly proposing with desirable properties. Recently, score-based, or diffusion, generative models have outperformed previous approaches. Key to their success is the close relationship between score and physical force, allowing the use of powerful equivariant neural networks. However, the of the learnt score is not yet well understood. Here, we analyse the score by training an based diffusion model for molecular generation. We find that during the generation the score a restorative potential initially and a quantum-mechanical force at the end. In between
the two endpoints, it exhibits special properties that enable the building of large molecules. Using from the trained model, we present Similarity-based Molecular Generation (SiMGen), a new for zero shot molecular generation. SiMGen combines a time-dependent similarity kernel descriptors from a pretrained machine learning force field to generate molecules without any training. Our approach allows full control over the molecular shape through point cloud and supports conditional generation. We also release an interactive web tool that allows users generate structures with SiMGen online (https://zndraw.icp.uni-stuttgart.de).
@misc{elijosius24a,
abstract = {Generative modelling aims to accelerate the discovery of novel chemicals by directly proposing with desirable properties. Recently, score-based, or diffusion, generative models have outperformed previous approaches. Key to their success is the close relationship between score and physical force, allowing the use of powerful equivariant neural networks. However, the of the learnt score is not yet well understood. Here, we analyse the score by training an based diffusion model for molecular generation. We find that during the generation the score a restorative potential initially and a quantum-mechanical force at the end. In between
the two endpoints, it exhibits special properties that enable the building of large molecules. Using from the trained model, we present Similarity-based Molecular Generation (SiMGen), a new for zero shot molecular generation. SiMGen combines a time-dependent similarity kernel descriptors from a pretrained machine learning force field to generate molecules without any training. Our approach allows full control over the molecular shape through point cloud and supports conditional generation. We also release an interactive web tool that allows users generate structures with SiMGen online (https://zndraw.icp.uni-stuttgart.de).},
added-at = {2024-03-27T09:43:01.000+0100},
archiveprefix = {arXiv},
author = {Elijošius, Rokas and Zills, Fabian and Batatia, Ilyes and Norwood, Sam Walton and Kovács, Dávid Péter and Holm, Christian and Csányi, Gábor},
biburl = {https://puma.ub.uni-stuttgart.de/bibtex/2a36ffb4d073542470967688db997304a/simtech},
description = {[2402.08708] Zero Shot Molecular Generation via Similarity Kernels},
doi = {10.48550/arXiv.2402.08708},
eprint = {2402.08708},
interhash = {9f917b4a269e4d100d0406f9905afb0b},
intrahash = {a36ffb4d073542470967688db997304a},
keywords = {EXC2075},
month = feb,
primaryclass = {physics.chem-ph},
timestamp = {2024-03-27T09:52:01.000+0100},
title = {Zero Shot Molecular Generation via Similarity Kernels},
year = 2024
}