The use of artificial intelligence is revolutionising almost every aspect of our lives; among the most impactful advances in AI, generative models are getting days by days more important thanks to their abilities in generating text, video, images and many other data. In recent years, following this trend, there has been a growing interest in development of generative models focused on de novo design of molecules for drug or material discovery. Many models available in literature nowadays consider 3D information, in particular the position in the three-dimensional space. This can pose a problem for learning architectures which do not incorporate rotational invariance. In this thesis an attempt is made to consider the information about the arrangement of the molecule in the space, overcoming problems related to coordinates in 3D space, by considering the distance matrix between between each atom. This allows to exploit information about the relationship of each atom in the space and overcome the problem related to the rotation of the molecule .This is done by considering a diffusion model that generates atoms, bonds and distance between each atom.

The use of artificial intelligence is revolutionising almost every aspect of our lives; among the most impactful advances in AI, generative models are getting days by days more important thanks to their abilities in generating text, video, images and many other data. In recent years, following this trend, there has been a growing interest in development of generative models focused on de novo design of molecules for drug or material discovery. Many models available in literature nowadays consider 3D information, in particular the position in the three-dimensional space. This can pose a problem for learning architectures which do not incorporate rotational invariance. In this thesis an attempt is made to consider the information about the arrangement of the molecule in the space, overcoming problems related to coordinates in 3D space, by considering the distance matrix between between each atom. This allows to exploit information about the relationship of each atom in the space and overcome the problem related to the rotation of the molecule .This is done by considering a diffusion model that generates atoms, bonds and distance between each atom.

A diffusion generative model on atom distances for de novo molecular design

BALLARINI, MARCO
2023/2024

Abstract

The use of artificial intelligence is revolutionising almost every aspect of our lives; among the most impactful advances in AI, generative models are getting days by days more important thanks to their abilities in generating text, video, images and many other data. In recent years, following this trend, there has been a growing interest in development of generative models focused on de novo design of molecules for drug or material discovery. Many models available in literature nowadays consider 3D information, in particular the position in the three-dimensional space. This can pose a problem for learning architectures which do not incorporate rotational invariance. In this thesis an attempt is made to consider the information about the arrangement of the molecule in the space, overcoming problems related to coordinates in 3D space, by considering the distance matrix between between each atom. This allows to exploit information about the relationship of each atom in the space and overcome the problem related to the rotation of the molecule .This is done by considering a diffusion model that generates atoms, bonds and distance between each atom.
2023
A diffusion generative model on atom distances for de novo molecular design
The use of artificial intelligence is revolutionising almost every aspect of our lives; among the most impactful advances in AI, generative models are getting days by days more important thanks to their abilities in generating text, video, images and many other data. In recent years, following this trend, there has been a growing interest in development of generative models focused on de novo design of molecules for drug or material discovery. Many models available in literature nowadays consider 3D information, in particular the position in the three-dimensional space. This can pose a problem for learning architectures which do not incorporate rotational invariance. In this thesis an attempt is made to consider the information about the arrangement of the molecule in the space, overcoming problems related to coordinates in 3D space, by considering the distance matrix between between each atom. This allows to exploit information about the relationship of each atom in the space and overcome the problem related to the rotation of the molecule .This is done by considering a diffusion model that generates atoms, bonds and distance between each atom.
Generative model
Diffusion model
Molecular generation
Deep learning
File in questo prodotto:
File Dimensione Formato  
Ballarini_Marco.pdf

accesso aperto

Dimensione 8.73 MB
Formato Adobe PDF
8.73 MB Adobe PDF Visualizza/Apri

The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.12608/80878