Sign in
Ctrl K

3D-Vac

Personalized cancer vaccine design through 3D modelling boosted geometric learning

Photo credit – Shutterstock

Personalized cancer vaccines have recently achieved clinical breakthroughs in eliminating advanced tumors. Tailored to a patient’s individual cancer mutations, cancer vaccines train the immune system to attack tumors. However, the translation of cancer vaccine to a routine therapy is hindered by one key challenge: Which of a patient’s many tumor mutations serve best as vaccine candidates? To be good candidates, mutated tumor peptides must bind MHC proteins (forming peptide-MHC complex), which ship the peptides to the cell surface to be visible to T-cells.

This project combines the power of physics-based 3D modelling and data-driven geometric deep learnings (GDLs). GDLs are recent advances of deep learning that are specially designed for 3D data. We train GDL networks on 3D models of peptide-MHC to predict the binding of the two. Rotation invariant GDLs allow efficient usage of enriched structural information in 3D models, which is not exploited by hitherto used sequence-based methods. Our approach naturally deals with the challenge caused by the high length variance of peptides binding MHC class II, which dominates T-cell responses upon cancer vaccinations. Efficient training of GDLs on millions of 3D atomic-level models and associated experimental binding data requires efficient parallel data handling, and massive HPC and GPU use.

The resulting software will be integrated into our DeepRank platform, distributed on GitHub, and disseminated through user-community workshops. A user-friendly web server with a user-support forum will be developed for the biomedical community. The GDL software will be designed generally applicable to structural biology, chemistry, human genetics, and beyond.

On top of the main budget, we also obtained a Software Sustainability budget, aimed at improving DeepRank platform UX by providing a flexible interface that will attract more users, and combining the original DeepRank, DeepRank-GNN and DeepRank-Mute packages into DeepRank2, which will become the standard version for developers and users.

In addition to our core budget, we have obtained another funding specific for software sustainability. This initiative is dedicated to improving the user experience of our DeepRank platform by introducing a more adaptable interface intended for a wider audience. It also aims to consolidate the original DeepRank, DeepRank-GNN and DeepRank-Mut packages into DeepRank2, intended to become the standard for developers and users.

Participating organisations

Life Sciences
Life Sciences
Netherlands eScience Center
Radboud University Medical Center

Output

Team

Giulia Crocioni
Giulia Crocioni
eScience Research Engineer
Netherlands eScience Center
LX
Li Xue
Principal investigator
Radboud University Medical Center
Pablo Lopez-Tarifa
eScience Coordinator
Netherlands eScience Center
Dani Bodor
Dani Bodor
eScience Research Engineer
Netherlands eScience Center
DM
Dario Marzella

Related projects

DeepMolGen

Deep learning for molecule generation with optimal properties

Updated 7 months ago
Finished

DeepRank

Scoring 3D protein-protein interaction models using deep learning

Updated 12 months ago
Finished

Related software

DeepRank

DE

Deep learning framework for data mining protein-protein interactions using CNN

Updated 5 months ago
60 8

DeepRank2

DE

DeepRank2 is an open-source deep learning framework for data mining of protein-protein interfaces or single-residue missense variants. This package is an improved and unified version of three previously developed packages: DeepRank, DeepRank-GNN and DeepRank-Mut.

Updated 1 week ago
7 11

DeepRank GNN

DE

DeepRank-GNN is the graph neural network of our DeepRank package. DeepRank GNN allows to train graph neural networks to classify protein-protein interface

Updated 21 months ago
37 2