This is an AI for Health MSc project. Students are elgible to receive a monthly reimbursement of €500,- for a period of six months. For more information please read the requirements.
Clinical Problem
Deeper understanding of the immune system’s intricacies has led to clinical breakthroughs of personalized cancer vaccines in eliminating tumors in advanced-stage cancer patients 1-4 . Formulated with fragments from a patient’s tumor DNA, cancer vaccines train a patient’s own immune system to recognize a patient’s mutated cancer proteins as ‘foreign’ and wage a lethal attack against tumors (see key concepts in Box 1). The major puzzle in this field is: which of a patient’s hundreds of tumor mutations can trigger the immune system to attack tumors ? Complementary to costly and time-consuming wet-lab screenings (e.g., Sipuleucel-T was priced at $93,000 5 ), predictive algorithms that can quickly pinpoint neoantigens from a patient’s tumor DNA are urgently needed, if personalized cancer vaccines are to be applied on a large scale. We aim to predict 3D peptide:MHC structures using a neural network simulator in this project. 3D peptide:MHC structures bear crucial information for predicting effective cancer vaccine candidates, the major challenge of this field. Our overall aim is to improve the efficacy, safety and development time of existing T cell based cancer vaccine approaches.
Solution
General-purpose AI algorithms predicting 3D structures of proteins, l ike AlphaFold2, are computationally expensive and do not perform well on proteins without MSA (multiple sequence alignment), an i mportant source of evolutionary i nformation. Here, we will explore using a dynamic graph neural network (GNN, https://arxiv.org/abs/2102.09844) for modelling 3D peptide:MHC complexes (peptides do not have MSA). Protein complexes have different shapes and sizes presenting unique challenges to conventional machine l earning algorithms, which take fixed-size i nput. Graphs are natural representations of protein complexes. We will further i ntegrate domain knowledge (e.g., where are protein i nterfaces, experimentally measured distance restraints, and so on) i nto the GNN model to restrain the structure predictions thus speeding up the computation. This project pioneers i ntegrating domain knowledge to guide AI-based modelling, and thus has a wide range of applications on antibody:antigen, TCR:pMHC, and general protein-peptide interactions.
Training such GNN requires at l east hundreds of thousands of 3D structures, which i s a big data challenge. This project i s built upon DeepRank, our deep l earning framework for data mining very l arge sets (up to millions) of protein i nterfaces. The resulting software will be i ntegrated i nto our DeepRank platform, distributed on GitHub, and disseminated through user-community workshops. The software will be designed generally applicable to structural biology, chemistry, human genetics and beyond. We specifically define the following tasks:
Task 1 : Train a GNN on a single well-studied peptide:MHC complex. The goal here is to find a suitable architecture and the correct hyperparameters before proceeding to the bigger task of making a general peptide:MHC complex predictor. For the test complex, we plan to use: PDB ID: 3MGT, a crystal structure of a H5-specific CTL epitope variant derived from H5N1 influenza virus in complex with HLA-A*0201. We will use the PANDORA software to generate 50,000 models, which will serve as training and validation datasets for the GNN.
Task 2 : Train the GNN from task 1 (but with more parameter capacity) on a bigger dataset containing a great variety of peptides and MHC alleles. This is to accomplish our final desired goal of predicting peptide conformations for any peptide:MHC complex. Here again we will use the Pandora software package to generate/simulate the “true” models as labels for training (dataset 3 below). For testing, we will use the experimentally derived peptide conformations from dataset 1, which will be excluded from the train and validation sets to test the generalization capacity of the final GNN.
Task 3: Test the generalization capabilities of the trained GNN from task 2 on non-peptide loop predictions. The GNN will, by design, iteratively update the peptide loop conformation until convergence. Since the updates are based on local interactions, which should be general for all loop formations, we will test whether or not the trained GNN can generalize to also model TCR-CDR3 loops. We will test on a set of bound and unbound TCR:peptide:MHC structures. We will take unbound CDR3 loops as the starting point of GNN and ask GNN to model the bound conformation of CDR3 loops. We will also compare the final results with HADDOCK simulation results, a flagship physics-based docking software for flexible modeling protein-protein complexes.
Data
We will use our PANDORA software to generate 3D peptide:MHC models and their energy values. We have shown that PANDORA is able to reproduce near-native peptide:MHC structures and other plausible conformations.
Dataset 1: X-ray pMHC-I structures and their energy values. We have downloaded 841 X-ray pMHC-I structures from the IMGT database. Dataset 2: 3D models for dataset 1 and their energy values. For each entry in dataset 1, we will use PANDORA to generate N (100, to be optimized) models. PANDORA also outputs the energy values for each model. Dataset 3: 3D models for 147,326 strong MHC-I binders from biochemistry assays and their energy values. We collected 147,326 strong MHC-I binders from the MHCflurry paper, which are biochemistry binding affinity data and mass spectrometry (MS) eluted data. For each entry, we will generate 100 PANDORA models along with their energies.
Embedding
You will be embedded in the Center for Molecular and Biomolecular Informatics at Radboudumc. We provide access to CPU and GPU clusters, direct supervision by PhD students and regular meetings with the PIs of this project.
Requirements
- Students Artificial Intelligence, Data Science, Computer Science, Bioinformatics in the final stages of their Master education.
- You should be proficient in python programming and have a theoretical understanding of deep learning architectures.
- Basic biological / biomedical knowledge is preferred.
Information
- Project duration: 6 months
- Location: Radboud University Medical Center
- More information can be obtained from Li Xue (mailto: li.xue@radboudumc.nl)
