Example Projects

The aim of my doctoral project is to improve the prediction of antibody structures and optimise the use of existing antibody data. As a starting point, we developed ABlooper [1], a tool that predicts the position of backbone atoms for CDR loops in antibodies. This was the first method to incorporate equivariant deep learning into antibody structure prediction. However, ABlooper has some limitations: it cannot model side-chain atoms and sometimes generates unphysical structures, which need to be corrected using computationally expensive methods. 

Our second iteration for antibody structure prediction, ABodyBuilder2 [2], addresses some of these challenges. Inspired by the AlphaFold2 [3] model, ABodyBuilder2 directly predicts side chain atoms, thus enhancing the accuracy of their modelling. By incorporating hard-coded rules for bond angles and distances within a residue, it also reduces the number of unphysical structures predicted. ABodyBuilder2 is as accurate as AlphaFold-Multimer [4] while achieving over a hundred-fold increase in speed. At the time of release, it stands as the most accurate of all antibody-specific methods. When we trained the model exclusively on T-Cell Receptors (TCRs) and nanobodies, comparable results were obtained, leading to the creation of TCRBuilder2 and NanoBodyBuilder2.

Concurrently, we introduced two innovative tools: KA-Search [5] and PLAbDab [6], designed to effectively harness publicly available antibody data. KA-Search facilitates rapid and exhaustive searches in vast NGS antibody databases based on sequence identity. PLAbDab compiles patent and paper-derived antibody sequences into a searchable database and makes them searchable either by sequence or structure. We anticipate that these tools, alongside our improved structural modelling, will aid in the process therapeutic antibody design. 

All methods developed during this PhD are freely available either as web servers (https://opig.stats.ox.ac.uk/webapps/sabdab-sabpred/sabpred) or on GitHub (https://github.com/oxpig/).  

[1] Abanades, Brennan, et al. “ABlooper: Fast accurate antibody CDR loop structure prediction with accuracy estimation”. In: Bioinformatics 38.7 (2022), pp. 1877–188 
[2] Abanades, Brennan, et al. “ImmuneBuilder: Deep-Learning models for predicting the structures of immune proteins.” In: Communications Biology 6.1 (2023), p. 575 
[3] Jumper, John, et al. "Highly accurate protein structure prediction with AlphaFold." Nature 596.7873 (2021), pp. 583-589. 
[4] Evans, Richard, et al. "Protein complex prediction with AlphaFold-Multimer." bioRxiv (2021) 
[5] Olsen, Tobias H., et al. “KA-Search: Rapid and exhaustive sequence identity search of known antibodies”. In: Scientific Reports 13.1 (2023), p. 11612. 
[6] Abanades, Brennan, et al. “The Patent and Literature Antibody Database (PLAbDab): an evolving reference set of functionally diverse, literature-annotated antibody sequences and structures” In: bioRxiv (2023). 

David’s research focus is the study of treatment response variability and modelling approaches to mitigate it. Treatment response variability across patients is a common phenomenon in clinical practice. For many drugs this inter-individual variability does not require much (if any) individualisation of dosing strategies. However, for some drugs, including chemotherapies and some monoclonal antibody treatments, individualisation of dosages are needed to avoid harmful adverse events. Model-informed precision dosing (MIPD) is an emerging approach to guide the individualisation of dosing regimens of otherwise difficult-to-administer drugs.

Several MIPD approaches have been suggested to predict dosing strategies, including regression, reinforcement learning (RL) and pharmacokinetic and pharmacodynamic (PKPD) modelling, but a unified framework to study the strengths and limitations of these approaches was missing. David developed a framework for the simulation of clinical MIPD trials [1].

In the context of this project, David developed the open-source Python package Chi which can be used to implement and simulate treatment response models [2]. The package simplifies the reproduction of clinical trial simulation and enables researchers to implement their own clinical trial models. In addition, Chi integrates with the open-source Python package Pints which makes the estimation and inference of model parameters from data possible [3].

A central research theme in David’s work is the quantification of model uncertainty, including parametric and structural uncertainty. In a recent publication, David introduces a novel method for the estimation of the parameters of nonlinear mixed effects models from snapshot measurements [4]. In another publication, David discusses the limitations of information criteria-based model selection for the estimation of structural uncertainty and studies alternative approaches [5].

References:

[1] Augustin, David, et al. "Simulating clinical trials for model-informed precision dosing: Using warfarin treatment as a use case." bioRxiv (2023): 2023-07.
[2] Augustin, David (2021). Chi - An open source python package for treatment response modelling (Version 0.1.0) [Computer software]. https://github.com/DavAug/chi
[3] Clerx, Michael, et al. "Probabilistic Inference on Noisy Time Series (PINTS)." (2019).
[4] Augustin, David, et al. "Filter inference: A scalable nonlinear mixed effects inference approach for snapshot time series data." PLOS Computational Biology 19.5 (2023): e1011135.
[5] Augustin, David, et al. "Treatment response prediction: Is model selection unreliable?." bioRxiv (2022): 2022-03.

Alzheimer’s disease is a complex, multi-factorial disease about which relatively little is known. Working between Oxford and Roche, I am investigating the role of toxic proteins in driving the progression of Alzheimer’s disease. The project is very interdisciplinary, drawing mainly from methods in applied mathematic and also relying heavily on multi-modal brain imaging and statistical inference. During my PhD, I have had the opportunity to visit the Roche headquarters in Basel to learn more about their internal projects and how best to incorporate to my DPhil work into their analysis pipelines. I have also had the the opportunity to visit external collaborators and present my work internationally.

My work focusses on modelling toxic forms of tau-protein, a particularly nefarious protein that exhibits a unique spatiotemporal progression throughout Alzheimer’s disease. We have shown that simple dynamical systems embedded a brain connectivity network can describe accurately the spatiotemporal evolution of tau-protein. Furthermore, I have shown that these models can be calibrated to patient data using Bayesian inference, allowing us to investigate changes in protein dynamics across the disease timeline and make patient specific predictions that incorporate uncertainty about the observation and modelling processes.

Cancer, rheumatoid arthritis, and many infectious diseases such as SARS-CoV-2 can all be treated, to an extent, using antibody drugs. These therapeutics are incredibly powerful given antibodies’ high binding affinity and specificity to their disease targets. However, developing such treatments often takes approximately ten years and over $1 billion in investment. Speeding up and lowering the cost of this process is therefore essential to the development of more affordable, effective treatments in the coming decades.

My DPhil research focuses on developing novel computational advances to the initial ‘discovery’ stage of the drug development pipeline. This step can last approximately five years and involves testing and optimising tens of thousands of potential antibody-leads for their neutralising capability, developability, and immunogenicity. Working within the Oxford Protein Informatics Group (OPIG) in the Department of Statistics, I use advanced deep learning methods and other computational tools to enable researchers to move some of these in vitro discovery stage tests in silico, helping reduce development times and costs.

Recently, we published our antibody binding site (paratope) prediction tool – Paragraph (Chinery et al 2023. Bioinformatics). Paragraph uses the latest advances in graph neural networks to predict which amino acids of an antibody are likely to take part in binding its target. This knowledge is useful as different 3D orientations of an antibody-antigen complex can potentially lead to varying on- and off-target effects. For those wishing to better understand the methods used in this project, please take a look at our GitHub page - https://github.com/oxpig/Paragraph. 

Left image shows a Y-shaped antibody bound to its target antigen (grey) e.g. a viral protein. An antibody is comprised of four protein chains – two heavy chains (dark blue) and two light chains (light blue). The right image provides a more detailed representation of one of the binding ‘arms’ of the antibody, highlighting the complex 3D conformations the protein chains form. These unique shapes allow antibodies to bind antigens with high affinity and specificity, making antibodies powerful immune proteins and therapeutics. Figure created in PyMOL.

In medical imaging, a lack of high-quality labelled data is a common issue. My research focuses on how synthetically generated data can assist neural networks in tasks such as classification, segmentation, and localisation. Specifically, I utilise deep learning methods to aid in the acquisition and interpretation of transoesophageal echocardiograms.

 

I experiment with generative methods based on different principles, such as adversarial, contrastive, and diffusive to create synthetic data. I then use this synthetic data to enhance existing real datasets, aiming to improve performance on tasks like left-ventricle segmentation which is key to calculating physiological variables that have diagnostic use. Moreover, the ability to localise the ultrasound probe in 3D space is becoming increasingly valuable. Therefore, I also investigate how synthetic data can be employed to train a network capable of determining the probe’s position and the viewed object, thereby significantly reducing the expertise required to perform an examination.