PROJECT A. Determining the phase transition for human-to-human transmissible and non-human to-human transmissible respiratory pathogens
PROJECT B. Learning Which Constraints Matter
PROJECT C. Active Brownian Particles as Models for Motile Robotics Systems

---

PROJECT A: Determining the phase transition for human-to-human transmissible and non-human to-human transmissible respiratory pathogens
Internal: Prof Matt KeelingLink opens in a new window and Dr Ed HillLink opens in a new window
External: Dr Tom Finnie (UKHSA)Link opens in a new window

Influenza is a significant pathogen to humans and substantial morbidity and mortality each year and, should a strain evade the immunity found in the human population, poses a substantial risk of causing a global disease pandemic. Influenza comprises four viral species (Influenza A – D) of which the ‘A’ strain is the focus of this project as, in addition to the human transmissible variants there exist a substantial number of animal variants which circulate freely in populations of wild animals. These rarely infect humans as the nature of the biology and in particular the location of the cells that a wild-life virus may invade within the respiratory system means that it is physically very difficult for these viruses to both attack human cells and simultaneously be able to move from one human to another. This project aims to quantitatively explore this trade-off between the cell type that a virus can invade, and the ability of that virus to move from one human to the next. It may also prove possible to study the genetic profile of the required mutations to adapt to human physiology and so become a ‘human’ virus.

In this RSG project, we would be interested to explore the following:

1. Data track: A review of the literature on influenza A virus receptor binding specificity. Building upon an internal UKHSA literature review (see Reference [1]), the two purposes are to: (i) establish the biological processes involved and, (ii) get an idea of the modelling approaches that have been typically used to date, given the data available.
2. Within-host modelling track: What is the trade-off between infectiveness and ability of pathogen to be transmitted out of the lung? To study this problem, initially two respiratory, within-host compartmental ODE models are to be coded. One model will focus on the upper respiratory track and the second on the lower respiratory track. These models can be used to explore the trade-off between infectiveness and ability to be transmitted out of the lung. A natural extension would then consider a multi-patch model. That modelling approach can be used to study how much expression of receptors at various stages of respiratory tract are necessary to get certain amount of infection in the lower- or upper- respiratory track of a susceptible individual.
3. Genetic track: How many mutations are we away from a non-human-to-human transmissible influenza strain becoming human-to-human transmissible? The investigation of this question will require the development of a mutations model (see Reference [2]).

References

[1] Briefing note: Influenza A virus receptor binding specificity

[2] PROTECT Phase 3 Final report of Work-package 2.1.3: Within host and dose response models of SARS-CoV-2 infection, transmission and susceptibility
URL: https://warwick.ac.uk/fac/sci/mathsys/courses/msc/ma932/protect3_ukhsa_final_report.pdf

---

PROJECT B: Learning Which Constraints Matter
Internal: Prof Juergen Branke (WBS)Link opens in a new window
External: Juan Ungredda, ESTECO SpA Link opens in a new window

ESTECO SpA is a software company specializing on simulation process automation and design optimization in engineering. Many engineering design problems are black box and computationally expensive, in the sense that evaluating a possible design involves running an expensive simulation model. Running such a model often takes minutes or even hours on a supercomputer, and it doesn’t provide gradient information. This is challenging for optimisation, as it limits the number of potential designs that can be evaluated.

One very promising approach to tackle engineering design problems is Bayesian optimisation (BO). BO builds a surrogate model based on the data it has collected so far, and then iteratively decides which additional evaluation would be most informative, explicitly balancing exploration (trying something new) and exploitation (trying high quality designs).

If the problem has constraints, they can be incorporated by building separate surrogate models for each constraint, and then additionally taking into account a design’s probability of being feasible when choosing the next design to be evaluated. For a recent constraint handling approach developed at Warwick, see https://arxiv.org/pdf/2105.13245.pdf.

However, when the problem has many constraints, and these constraints come from different simulation models, computing all the constraint values and updating all the surrogate models becomes computationally burdensome. On the other hand, in practice, only few of the constraints are usually binding at the optimum, so perhaps not all constraints have to be evaluated every time.

The idea of this project is thus to develop a Bayesian optimisation algorithm that can decide, for each solution, which constraints should be evaluated, and in what sequence (e.g., one may want to check the constraint first that is most likely to be violated, or if the constraints have different cost, one may want to check the least expensive constraint first, and one may not want to evaluate constraints at all that are likely to be feasible).

The project would start by getting an overview of existing approaches in this space, coming up with some new ideas, and then benchmarking different alternatives on a few test problems. For a subsequent PhD project one could for example extend this to multi-objective problems or high-dimensional problems where BO usually struggles.

---

PROJECT C: Active Brownian Particles as Models for Motile Robotics Systems
Internal: Prof Matthew TurnerLink opens in a new window and Dr Gareth AlexanderLink opens in a new window
External: Dr Michael Riedl, MPI Dresden, GermanyLink opens in a new window

Motion involving many individual particles is ubiquitous in nature generally and in human societies in particular. Examples of collective motion in nature include the motion of flocks of birds, swarms of insects or shoals of fish. In human societies we observe the behaviour of crowds or the motion of many vehicles on a transport network. More recently, algorithms to control driverless cars or drone swarms have become industrially important. The use of simple robots as model systems has become popular. One such experimental system has been developed by our external partner Dr Michael Riedl now working at the MPI in Dresden, Germany. This consists of motile spherical robots that can move by rolling. He is working to introduce sensory input and a dynamic programmed response into his system. We believe that now is an opportune moment to study these systems using computer simulation. Arguably the simplest system is that of “Active Brownian Particles” (ABPs) that self-propel with an orientation that is persistent over short timescales but which re-orientates over longer timescales due to rotational diffusion. These particles interact by a short-ranged repulsion.

This project will involve students undertaking the following steps:

1. A review of the literature on collective motion: The students will review the state-of the-art in modelling using ABPs and other agent-based models. The agents are usually taken to be identical and hence these systems are leaderless or internally coordinated, rather than externally controlled. What features are already known to emerge from ABP systems? Start thinking about how this might be modified by the presence of an internal oscillator (as in the experimental system) or active control in response to local sensory input, e.g. physical stress.
2. A coding exercise: a simple ABP algorithm is to be coded from scratch or deployed, e.g. using hoomd-blue, https://glotzerlab.engin.umich.edu/hoomd-blue/. Simple ABP to be calibrated against known properties from the literature, e.g. the existence of clustering, known as MIPS, as a test that the code is working properly
3. An analysis of possible refinements of these models, e.g. Do collisions occur in these models? What is the physical signature of a collision? How could this, or other sensory information, be used to generate a response in the ABPs that might, e.g. lead to (i) coalignment or other types of order, (ii) synchronisation of the oscillators, (iii) reinforce or suppress MIPS? Could any of this be useful feature in autonomous robotics or interesting to study the underpinning experiments? Can simple algorithms be identified or developed to reduce collisions? Can this problem be expressed as a form of control theory and in what way might an individual’s optimal behaviour be different to the global optimal behaviour?
4. A review of how such algorithms might be useful, how they could be refined further or extended and an analysis of future prospects.