2017 CMP Industrial Project Proposals

Gabrielsson, J., & Hjorth, S. (2015). Pattern Recognition in Pharmacodynamic Data Analysis. The AAPS Journal. http://doi.org/10.1208/s12248-015-9842-5

Gabrielsson, J., & Peletier, L. A. (2014). Dose-response-time data analysis involving nonlinear dynamics, feedback and delay. European Journal of Pharmaceutical Sciences, 59(1), 36–48. http://doi.org/10.1016/j.ejps.2014.04.007

Marusic, M., Bajzer, Z., Vuk-Pavlovic, S., & Freyer, J. P. (1994). Tumor growth in vivo and as multicellular spheroids compared by mathematical models. Bull Math Biol, 56(4), 617–631. http://doi.org/10.1007/BF02460714

Xu, X. S., Wang, H., & Vermeulen, A. (2011). Modeling delayed drug effect using discrete-time nonlinear autoregressive models: a connection with indirect response models. Journal of Pharmacokinetics and Pharmacodynamics, 38(3), 353–367. http://doi.org/10.1007/s10928-011-9197-1

The project is a first step towards a more sophisticated use of principal components analysis by the fund. In this context PCA is a valuable statistical tool for handling the datasets we encounter every day, essentially the technique is to project a high dimensional dataset onto a suitable low dimensional subspace which is chosen so as to maximise the information content preserved by the projection (in some sense which can be made more precise). A standard and typical example would be to use an eigen-decomposition of the covariance matrix of the data, choosing the most significant eigenvectors based on the

One obvious issue with eigen-techniques for PCA is that the results are extremely sensitive to outliers in the initial dataset, and in particular to noise. Another issue is that eigen-techniques do not scale well to large dimensions, but work most efficiently on small to medium scale problems. Recent developments in the field have led to techniques that are more scalable and more robust to noisy data.

The aim of this project is to review some of the recent literature on this area, to identify and implement a few of the more interesting techniques and to apply these algorithms to financial datasets of importance to the business.

The project will focus on a theme of future urban analysis, where large datasets capture the ebb and flow of normal population activity, often periodic. Within the datasets reside changing structures, symmetries and dynamics that provide a narrative of urban life at different levels of granularity. The main challenge of this student project will be to learn, apply and advance the understanding and theory of the global and local structure within large, undirected graphs using graphlet approaches [1, 2]. Additionally, the utility of graphlets to analyse the ‘role’ of nodes on graphs will be examined. Graphlets may be informally defined as small components of graphs that provide a structural-basis onto which large graphs may be decomposed. Formally, they are connected, induced, non-isomorphic subgraphs, with the edges between graphlet vertices only present if they are present in the large graph. Traditionally, the counting of graphlets within graphs has required the explicit counting of their embeddings on that graph. However, it is of interest to understand and exploit the group-theoretic symmetries (automorphism orbits) of graphlets to reduce the computational complexity of their counting. The project will seek to implement and advance graphlet analysis methods over a selected dataset, drawn from open sources [3] and using open source code [4, 5].

References

1. Hocevar T., Demsar J., Combinatorial algorithm for counting small induced graphs and orbits, 2016 Pre-print, Available from Website https://arxiv.org/abs/1601.06834?context=cs [accessed 10th January 2017].

2. Hocevar T., Demsar J., A combinatorial approach to graphlet counting, Bioinformatics, Vol. 30 no. 4 2014, pages 559–565., Available from Website http://bioinformatics.oxfordjournals.org/content/30/4/559.full.pdf+html [accessed 10th January 2017].

3. Stanford Network Analysis Project, Available from Website https://snap.stanford.edu/data/index.html [accessed 10th January 2017].

4. Graphlet C++ source code, Available from Website http://www.biolab.si/supp/orca/orca.html [accessed 10th January 2017].

5. SageMath, software package, Available from Website http://www.sagemath.org/ [accessed 11th January 2017].

Introduction: Patents have traditionally been modelled as important for economic growth. This is because economic growth has been modelled as being based on developments in science, which in turn have been modelled as being funded sub-optimally by the market. But whereas the empirical evidence shows that developments in science do indeed underpin economic growth, the existence or absence of patent laws seems not to affect rates of either scientific or economic growth (M Boldrin, D Levine, 2010, Against Intellectual Monopoly, CUP.) A new model for patents is therefore required.

Background: The development post-war of game theoretic models of science as a public good (see the references in T Kealey, M Ricketts, 2014, Modelling science as a contribution good, Research Policy, 43: 1014-1024, open access) provided a public-interest justification for patents, namely that under the market inventors cannot capture the full benefits of their inventions, so they will underinvest in R&D unless they are provided with the temporary monopoly of patents. But we have now modelled science as a contribution good (because the full benefits of science cannot be captured by free-riders, only by active scientists) under which investors may not need the incentive of patents before investing.

The project: Under the market, advances in science are translated into economic growth by first-mover advantage, which provides a temporary (if depreciating) monopoly of asymmetry of information and tacit knowledge. We would like to model that against a model of patents, to determine which - under the model of science as a public good - yields the socially-optimal outcome. The project will therefore involve modelling the incentive to do research under first mover advantage against modelling the incentive under patents.

The student: We envisage that the student would work with - or even at - the Department of Economics in Buckingham (which would supply accomodation for the length of the project). Last year we hosted a student under this scheme, Mr Sebastian Damrich, who has kindly agreed to discuss how he felt about working with us (sebastian.damrich@gmx.de). We are now writing up his work (on crowding out) as a paper, to be submitted to Research Policy, and though we'd hope to be as successful again, the project this year is more open-ended so we cannot sure it will translate standing-alone into a paper, we trust it will lay the basis for one.

The conceptual model for this project is the Magic Square. Many variants exist for this mathematical toy. We wish to devise a means of constructing a MS based around an initial set of “non-magical” values. For example, given an initial n x n grid of unknown values. Is it possible to construct an (n+b) x (n+b) grid around the initial data where any linear path traced through the grid has the same product as any other. To develop this problem further: a)the ultimate model is a volume of 500 – 2000 cubed. b) the linear paths through the 3D model can be from nearly any direction c) What are the limitations on value b.

X-ray absorption along a path of length t through a material with a linear absorption coefficient u, where u is dependent on the energy of the x-ray photon is given by I=Io exp(-ut). Where a sample is illuminated by a polychromatic x-ray beam the intensity at a point on the detector can be determined by integrating the absorption along the path from source to detector for each material and each x-ray energy.
This project will investigate means of determining the material properties or material that must be added to the extremities of a path such that any path generates the same or similar value for total x-ray absorption.

The successful candidate will develop proof-of-concept natural language processing (NLP) techniques aimed at extracting experimental parameters and biochemical contexts from PDF documents. Proof of concept work will be conducted on published research papers and will involve recovery of information stored in both raw text and tables. This information will be verified against curated public databases such as the Protein Data Bank or ChEMBL.

References:

Mallory, Zhang, Ré, and Altman, 2016, Large-scale extraction of gene interactions from full text literature using DeepDive, Bioinformatics, 32(1), 106-113; DOI: 10.1093/bioinformatics/btv476

McEntire et al., 2016, Application of an automated natural language processing (NLP) workflow to enable federated search of external biomedical content in drug discovery and development, Drug Discovery Today, 21(5), 826-835, DOI: 10.1016/j.drudis.2016.03.006

Multiple datasets of the received signals will be available from tests against different targets and using the same target in different orientations. The data will be divided between a group of data where the targets are identified and a group of data where the targets are blind. The datasets of identified targets are to be used to generate algorithms for target identification. The algorithms should then be used on the blind data group with the aim of outputting a prediction of the target type.

The scope of the project, size of datasets and required facilities will be finalised in discussion with the student.

If the work is undertaken at Dstl Fort Halstead the student will need to be a UK national and undergo security clearance. However, these restrictions will not be necessary if the student has access to suitable university computer facilities. Completion of the project at Cambridge is preferable due to the time required to process clearance.

For any of the chosen scientific questions, the project foresees both the theoretical exploration and understanding of the methods in form of literature review, which would justify the choice of specific approaches to the data, and the implementation of Python scripts to evaluate the methods on real data provided internally.

This project follows on from work done in last year's PMP project by Will Boulton, and is in an adjacent area to a current PhD project being worked on by Ferdia Sherry looking at Changepoint techniques, the results of which will be available by the start of this PMP project.

The project is reasonably well-defined, with key steps seen to be:
1. Review existing methods
2. Review existing Artesis data and characterise data sets
3. Select a small set of methods that appear to be appropriate to the nature of the data
3. Categorise and collate data into appropriate sub groups for treatment methods
4. Try out the methods on the relevant data sets - probably involving some programming to apply the methods
5. Evaluate the outcomes - identify and recommend the preferred method
6. If time allows, encode this preferred method, and demonstrate how it produces prognosis of failure dates, with confidence ratings, and how these confidence figures change as time progresses

In Diamond, SPEM and EM work-flows enable the automatic on-line reconstruction of cryo-electron microscopy (cryo-ET) datasets. This aimed to process terabytes of images that are usually acquired in a typical cryo-ET session and to enable the e-Bic users to speed-up the processing required by the computationally intensive SPEM-routines.

Although, the aim is ambitious, there are many difficulties that limit the quality of the final reconstruction. Examples of such difficulties are the noisy 2D cryo-ET-images, accuracy of classification and image alignment. Among many, the accuracy of alignment of 2D-images is a very important factor that significantly limits the fidelity and the resolution of 3D-reconstruction. The current workflows are based on classical approach of image-alignment. However, it is recognised clearly by the SPEM-community that there are reasons to doubt the accuracy of the results achieved by current alignment procedures. Therefore, there are compiling reasons to look for more advanced methods and conduct a quality assessment study of the state of the art alignment methods that can be used to improve the accuracy of our SPEM and ET pipelines at Diamond.

The placement will investigate the use of the signature of ship signals in the time domain, for example using the temporal curvature of radar fields, to remove ship signals from the radar data. Other approaches to the problem would be considered if thought to be more appropriate.

Existing methods for calling mutations in TP53 do not perform well at low levels of detection owing to noise from sequencing enzymes, sample library preparation, as well as the overwhelming abundance of normal DNA which dilutes the signal from the tumour DNA. The objective of this project is to develop new and more accurate approaches for detecting TP53 mutations in circulating tumour DNA and low cellularity tumour biopsies using Bayesian and related statistical modelling approaches. In preliminary work we have established error models that are nucleotide specific within TP53 and have sequencing data from over 2000 cases from which further error models can be constructed.

This project will provide strong training in the application of computational methods to clinically relevant problems in cancer genomics and how they can be implemented in clinical practice.

In exploring this phenomenon, we have had some success with numerical simulations in the FE package COMSOL, and are able to simulate the time development of the volume magnetization fields in the archetype process of NMR spin echo formation, in irregular pore spaces.

This open-ended project is to benchmark the COMSOL simulations against the (rather few) known asymptotic results for simple geometries, and to establish their regions of validity. Many extensions are possible and limited only by available time.

Prior experience with the COMSOL package would be desirable but is not essential.

Compared with HES Admitted Patient Care, the HES outpatients’ dataset remains relatively unexplored in the cancer field.

The intern project will focus on identifying patterns of outpatient clinic attendance for breast, colorectal and lung cancer patients; three of the most commonly diagnosed cancers.

http://healthdatainsight.co.uk/cancer-data-internships-2017/

Secondary aim: To identify and automate a data management cataloguing process across the multiple datasets within the Cancer Datasets Database.

We would like an intern to explore different data linkage concepts which could be automated and used across the wide-ranging datasets housed in the Cancer Datasets Database.

http://healthdatainsight.co.uk/cancer-data-internships-2017/

United Kingdom

A key objective of the internship role is to link the power of data back to relatable solutions, actionable outcomes and business decisions.
Core responsibilities of the role are to bring insights and analytics to support and/or predict patterns of consumer behaviour to service our customers better. This includes the scoping, design, development, monitoring and assessment of complex data models and tools. It can also involve identifying opportunities to streamline processes, gathering requirements from stakeholders and reporting back the gap analysis found after the process is complete.

SAS/SQL programming hands-on experience with experience of manipulating complex datasets and a work

Experience/knowledge of Data Science tools such as R, Python, SQL, Java, Hive and SPARK