DeepMolGen

Deep learning for molecule generation with optimal properties

Image credits: Shutterstock

Designing new molecules is at the core of chemistry, but in practice, creating molecular structures with useful properties is a difficult and time-consuming task, simply because there are so many possibilities of connecting different atoms together to form new molecules, most of which are useless. In this project, we use a new class of machine learning techniques, called diffusion models, that is able to learn how to generate new interesting molecules. Diffusion models have already proven to be very useful in generating artistic images, movies, music, and speech, and recently also for generating pharmaceutical drugs. We will develop and apply diffusion models to (1) recover the atomistic details from low-resolution simulations of polymers, (2) optimize enzyme structures, and (3) design better catalyst materials for CO2 conversion into useful chemicals.

The goal of the project was to make use of recent developments in generative AI, in particular diffusion denoising algorithms, to generate the atomic positions of molecular and material structures conditioned with desired molecular properties. An implementation of such an algorithm would be tremendously helpful for chemists interested in making new molecules, for example for catalyst materials or medicinal drug compounds. Currently, making new specialized molecules is a painstakingly slow process of trial and error. Clearly, if generative AI can propose interesting candidates for desired molecules this could speed up such molecule design enormously.

Unfortunately, our main objective was not reached, mainly because the development of denoising diffusion models for the purpose of molecule generation is more intricate and time consuming than we could anticipate. Nevertheless, we made good progress with developing the first scaffold of such a denoising diffusion model implementation, and our ambitions are not diminished so that our next plan is to find resources and man power to continue our aim to further develop and finalize a working software for the generation of molecule and material structures with desired properties.

The target audience of this software is in the first place our (and other) computational chemistry group(s) interested on in “in silico” design of new molecules and materials, but ultimately researchers that design and make chemicals, including synthetic chemists, pharmaceutical chemists, and material scientists.

Participating organisations

University of Amsterdam
Netherlands eScience Center
Natural Sciences & Engineering
Natural Sciences & Engineering

Team

BE
Bernd Ensing
JD
Jacobus Dijkman
PhD student
University of Amsterdam
FH
Ferry Hooft
PhD student
University of Amsterdam
YM
Ying Ying Ma
MSc student
University of Amsterdam
Rena Bakhshi
Programme Manager
Netherlands eScience Center

Related projects

GEMDAT

Developing a Generalized Molecular Dynamics Analysis Tool

Updated 3 months ago
In progress

3D-Vac

Personalized cancer vaccine design through 3D modelling boosted geometric learning

Updated 11 months ago
Finished