Software Projects
One of the unique aspects of the SABS R3 programme are the software projects our students work on. Every cohort is split into three or four groups, who work together on an open source project in each of our three research themes. The SABS R3 Research Software Engineer, Dr Martin Robinson, will work closely with you, helping to guide your work and reinforce the training received in the first module of the Programme.
The projects will begin in 3rd Week of Michaelmas (late October). We give an introductory talk on the projects on the Monday, and then you will meet your academic and industrial supervisors and be given introductory talks and background materials relevant for your project. The whole of 3rd Week is devoted to the projects, with the exact schedule determined on an individual project basis.
After this week, you will be able to work on the projects every Wednesday (except during the Cells and Systems module), with space being available at the DTC to work in person with your group. It will not be necessary to be at the DTC in person if you do not wish, and we will ensure that you are able to dial in remotely if necessary. You will also have another full week in 10th Week of Michalemas (mid-December), and 7th Week of Hilary (early March 2021), to work on the projects. You continue to meet during your first and second rotation projects, and once again have a week in late September to work on your projects together. In 0th Week of Michaelmas (start of October) of your Second Year, you will give a presentation of your work to our SABS Industrial Representatives.
2020-21 Software Projects
Project 1: Automated Data Extraction to Future-Proof Therapeutic and Natural Antibody Databases
Companies: UCB & GSK
Supervisors: Seb Kelm (UCB), James Snowden (UCB), Anna Vangone (Roche), Iain Moal (GSK), Professor Charlotte Deane, Claire Marks and Fergus Boyles
Databases play a central role in antibody informatics, pooling together previously dispersed sets of relevant amino acid/nucleotide sequences and solved structures into a consistent format with metadata to facilitate powerful statistical analyses. Since 2018, the Oxford Protein Informatics Group have created three new field-leading databases: the Observed Antibody Space (OAS) database — one of the largest repository of natural B-cell receptor repertoire sequencing data, Thera-SAbDab — the largest repository of sequence & structural information of all WHO-recognised therapeutic antibodies and nanobodies, and most recently CoV-AbDab — the only repository of sequence & structural information on coronavirus binding/neutralising antibodies. These databases were all manually curated and, though some have associated scripts to accelerate data formatting, none can harvest knowledge directly from the primary literature/other sources in an automated fashion. These tools are used extensively across academia and industry in the development of novel antibody therapeutics and increasing their usability and accelerating their updating systems will make them an even more powerful in these realms.
Throughout the year, your project will be to create frameworks that automate as much of the updating process as possible for each database. Thera-SAbDab is most likely to be completely automatable as its data sources are narrow and relatively consistent in format. Sources/input formats for OAS and CoV-AbDab, however, could prove too heterogenous for entirely automated data extraction. These subprojects may therefore also lead to Natural Language Processing assistants that draw papers/patents likely to contain relevant data to the attention of the curator. Throughout this project you can develop key skills in data science, including ‘scraping’ data from websites, database management, confidently interconverting between different data formats (including handling image inputs), and developing machine learning bots useful for targeted literature mining.
Project 2: The Drug Discovery Game
Company: Roche
Supervisors: Dr Torsten Schindler (Roche), Dr Rosa-Maria Rodriguez-Sarmiento (Roche), Professor Garrett Morris, Fergus Boyles
Drug discovery is a multi-parameter optimization problem where many properties must simultaneously be met in order for a new compound to fulfill the desired target compound profile (TCP) according to indication and target. The TCP compiles the characteristics the molecules must achieve in order to be of clinical interest.
Building upon the ideas behind The Drug Discovery Game [1] by Brian McGuiness and Robert Merrit, a Velcro-based game, we want to create an improved computer game that aims to mimic the initial workflow of a medicinal chemist on achieving a desired TCP and it will be based on drug discovery cases published in literature.
The game should cover several phases of the drug discovery process. Provided only with data from a high throughput screening (HTS), or by using an information driven approach starting with competitor compounds, the initial focus will be a hit expansion or a hit to lead identification phase. More complex processes of lead optimization and clinical candidate discovery will be added to the game in later stages.
On one hand, the game will serve as a tool that stimulates the player’s interest in AI guidance and medicinal chemistry. It will also reinforce knowledge and instill in the user the benefits to use a systematic approach rather than random selections.
On the other hand, it will allow the training of artificial agents which can be used to assist learning medicinal chemists in making strategic decisions at different stages of drug discovery similar to an augmented intelligence system for medicinal chemists.
[1] McGuinness, Brian F. and Merritt, J. Robert: “Illustrating Medicinal Chemistry Through an Interactive Demo: The Drug Discovery Game”, Poster presented at the 2016 Fall American Chemical Society Meeting. https://youtu.be/uH4DioPcbog
Project 3: Development of Open Source Epidemiology Modelling Software
Company: Dr Annabelle Lemenuel-Diot (Roche)
Supervisors: David Gavaghan, Ben Lambert, Martin Robinson and Robin Thompson
The current Covid-19 pandemic has highlighted the crucial role that mathematical and computational modelling can play in understanding the spread of infectious diseases, and in assessing the possible impact of any mitigation strategies that are put in place to control that spread. However, in shining a spotlight on this area of research, attention has inevitably turned to the manner in which the mathematical models and numerical algorithms have been instantiated in software. In common with almost all current academic codes, most of the software for modelling epidemic spread that was in existence at the start of the pandemic was developed principally to underpin the publication of academic papers. Code development therefore followed the standard norms in place within the discipline so that much of the subsequent criticism from professional research software engineers was perhaps unfair. A more valid criticism might be to argue that there is a very pressing need to develop robust open source modelling software that can be used in the future.
The goal of the proposed software project is to develop just such a robust software platform for the modelling of the transmission of infectious diseases. Software will be developed initially with the aim of providing a pedagogical tool for use in universities (and secondary schools) and in industrial settings in teaching both the modelling of epidemic spread and the importance of using robust software engineering principles. We will use the current Covid-19 pandemic as the key example due to the ready availability of extensive data. We will start with the simplest discrete and continuous models that simply look at the evolution in time of disease spread, and will extend this to consideration of spatial variation as the project develops. Particular attention will be paid to the key issues of the sensitivity of model outputs and predictions to model parameters, and to the inverse problem of estimating key parameters from data. Extensive tutorial examples (as Jupyter Notebooks) will be developed. Our ultimate goal will be to reproduce robust and sustainable versions of some of the key current codes that are being used to model the current pandemic.
To allow the platform to be used as a teaching tool, a subsidiary project will also develop a web app that allows interactive use of the models. Particular attention will be paid to the visualisation of outputs so that the material is accessible to secondary school students.
Much of the underpinning mathematical modelling, numerical algorithms, and statistical inference techniques will be common with the PKPD modelling project that was undertaken in collaboration with Roche in the last academic year (and that is still ongoing). This existing software will allow much more rapid progress to be made on this epidemiology modelling project.
The longer-term goal is to implement the Covid-19 transmission model currently being developed by Roche within the web-app. This model is an SEIR ODE-model with compartments for groups with differing infectivity. The app could be used to forward simulate epidemics across different countries and also be used for inference when given a data source.
Project 4: Extensible Clinical Imaging Quality Control Tool
Company: GE Healthcare
Supervisors: Chris Page (GE), Zak Catherall (GE), Professor Vicente Grau, Dr Martin Robinson
Imaging in clinical trials typically involves collecting data from multiple sites centrally for analysis. Data quality can vary considerably due to differences in technology and local procedures. Ingestion requires robust deidentification (to meet privacy and GCP regulations) and quality checks (to ensure adherence to the trial protocol) , both of which can be painstakingly manual. Initiatives providing data to researchers, such as UK Biobank or CSDR, perform further recoding on egress. These processes benefit from automatic checks and updates on image (meta)data according to user-defined rules.
Existing tools are typically proprietary or focussed on a specific modality or therapeutic area. We believe data quality is fundamental and such tools are precompetitive: a standardised, open, extensible approach benefits the research community.
We envision an in-stream device that receives DICOM data, accepts or rejects based on configurable criteria (specific to the trial, site, anatomy and/or modality), performs corrections or redactions, and finally exports to a repository.
The focus will be on checking, updating and tracking metadata according to a defined schema (real-world examples will be provided). The software shall have a plugin architecture, allowing e.g. image QC modules to be added in future.
GE will provide domain expertise, documented user requirements and regular feedback. We also propose “day-in-the-life” sessions, giving a flavour of our roles in software engineering and pharmaceutical R&D. Beyond a grounding in software engineering, students will gain expertise with the DICOM medical imaging standard (data format and validation, networking) and best practice in the pharma industry (software validation, regulation, data integrity).
2019-20 Software Projects
Project 1: Pharmacokinetic-Pharmacodynamic (PK-PD) Modelling
Supervisor: Dr Ken Wang, Roche
Utilization of mechanistic or semi-mechanistic mathematical model to quantify pharmacokinetics (PK) and pharmacodynamics (PD) have gained increasing importance in pharma industry throughout the discovery/development phases. The demand for PK and PK-PD modelling are high in pharma for both pre-clinical (discovery & translational) and clinical projects. Currently such “modelling” tasks requires skills in mathematical modelling (e.g. knowledge on model fitting, parameter estimation and uncertainty) and basic programming. The goal of this project is to develop a software package with a simple and user-friendly interface coupled with fast ODE solving and robust parameter estimation/ inference. This would be an ideal solution to promote PK-PD modelling to much wider community in pharma, enabling application of PKPD modelling to more complex projects (i.e. really moving away from the simplistic index or threshold PKPD approach).
Project 2: Expansion and Improvement of the Fragalysis Platform for Follow-Up Chemistry in Early Stage Drug-Discovery Projects
Supervisors: Prof. Frank von Delft and Dr Rachel Skyner (Diamond Light Source)
The cloud-based Fragalysis (https://fragalysis.diamond.ac.uk) platform has the potential to lead the way as an integrated platform for early-stage drug discovery and should guide the user from fragment hits to more potent binders, quickly and cheaply. The code base is designed as open-source, so is deployable at other facilities/institutions for internal use and development and can serve as a repository for diverse computational chemistry algorithms from the community. The project is linked with the EU Open Science Cloud and IRIS/the Ada Lovelace Centre (ALC).
14/19 listed SABS partners (e.g. Exscientia, Oxford Drug design) research in areas where the project is scientifically applicable. The current incarnation of the platform allows the user to review protein-ligand structures from their screening campaign, view the protein-ligand complex with an indication of interactions in the binding site, and interact with ‘vectors’ representing an elaboration point on the molecule. These vectors query a graph-network and provide the user with collections of commercially available compounds that they can purchase for a follow-up screen. This project will develop a python API that will allow users to upload their own data and will provide access to the underlying algorithms in an open source fashion, with direct access to compound libraries and open-source screening project data such as the Target Enabling Package (TEP) data from the SGC.
Project 3: MRI Brain Segmentation Optimized for Elderly Brains for Use in Neurodegenerative Disease
Supervisors: Dr Elisabeth Grecchi and Dr Christopher Buckley
Accurate anatomical segmentation of structure of interest is critical in the neurodegenerative disorder field. Quantitative analysis of medical images is often an important endpoint in research and clinical trials. Atlas-based MRI brain segmentation tools are widely available and used in research settings, however, brains of individuals suffering from Alzheimer’s Disease or Mild Cognitive Impairment are often characterised by localised or widespread atrophy, and pathologies are characterised by enlarged lateral ventricles when compared to a healthy population. Traditional atlas-based MRI segmentation tools have often fallen short in the segmentation of these specific regions, considering they rely on the original segmentation of healthy rather than atrophic brains. The goal of this project is to develop a brain MRI segmentation tool that provides accurate robust segmentation of problematic brain regions across the neurodegenerative spectrum. Furthermore, the methodology should be generalisable to perform well with the typical variance in MRI acquisition parameters and other factors that influence image contrast. The tool should be available and usable by the broad community without constraints due to necessary ancillary software or hardware. Accuracy and computation time should be comparable with commonly available methods.
