2026 CMP Academic Projects

There are a few kinds of hearing technologies that help people with hearing loss to hear. These include hearing aids that acoustically amplify sound entering someone’s ear. While hearing aids are normally prescribed and programmed by a healthcare professional, many over-the-counter hearing aids have recently become available as well, and even Air Pods contain hearing-assistive programs. However, these types of devices do not help those with more severe hearing loss. A more complex hearing device involves bypassing the outermost parts of the auditory system and are more like bionic ears: these are cochlear implants.

A cochlear implant is neural prosthetic that provide people with a profound hearing impairment with auditory perception by directly electrically stimulating the auditory nerve. It requires a surgical operation wherein a small string of electrodes is inserted into the patient’s inner ear. These electrodes are then controlled using a speech processor with a microphone that sits behind the patient’s ear. Most recipients can perceive speech well with their implant, and some enjoy music. They are arguably the most successful auditory prosthetic in existence today, with over 1 million users globally.

However, many struggle to understand speech with their implants, especially in challenging listening conditions with background noise. This may be due to the fact that individual cochlear implant users lose their hearing for different reasons, and the cochlear implant software and settings are not optimized for their unique pattern of hearing loss. We have developed a diagnostic tool called the Panoramic ECAP method that is designed to provide patient-specific indicators of the interaction between a patient’s implant and their auditory system. It involves applying a non-linear optimisation algorithm to electrophysiological measurements of neural activity in the cochlea. These measurements are called ‘Electrically Evoked Compound Action Potentials’ or ‘ECAPs’ for short. It is our hope that this tool can be used in a clinical setting to personalise and optimize cochlear-implant software and enable patients to achieve their hearing potential with their device.

The Panoramic ECAP (PECAP) method provides two estimates that describe the interaction between a cochlear implant patient’s inner ear (a.k.a. cochlea [1]) and their implant: the health of the auditory nerves and the spread of electrical current. The two PECAP estimates are quantified for each electrode of the cochlear implant. Areas of poor neural responsiveness or wide current spread reduce the efficiency of delivering auditory information from the neuroprosthetic device to the brain. The PECAP method estimates the neural responsiveness and the current spread at each electrode separately, using a nonlinear optimisation algorithm based on sequential quadratic programming whose structure is based on a theoretical framework [2]. We have conducted experiments to validate that these estimates are accurate and indeed separate from each other. While the results of our previous work are encouraging for clinical translation, long neural-response recording times and noisiness of some recorded data are barriers to translating this research to the clinic.

This project will have two facets that must be balanced with each other:

1. Building on a previous CMP student's work, determine if we can record less neural data (i.e. requiring a shorter amount of time to record) to accurately reproduce the PECAP algorithm's current-spread and neural-responsiveness estimates
2. Evaluate the potential and effectiveness of a recently-published denoising technique [3] to reduce measurement noise and thereby increase the percentage of cochlear-implant patents with whom the PECAP technique can be used (i.e. in cases of noisy neural data).

Two previous CMP students - Taren Rughooputh (2018) and Scott Hislop (2023) - conducted similar projects in our laboratory in the past whose work contributed to academic publications, for which they are on the author lists. (Rughooputh [1], Hislop [Garcia, et al, 2024, DOI: 10.1007/s10162-024-00966-x])

[1] Background Coursework from Duke University describing the peripheral auditory system:
https://gb.coursera.org/lecture/medical-neuroscience/peripheral-auditory-mechanisms-part-1-SQ6H7.
I recommend watching the first 3 videos of week 7 for a review of the human auditory system

[2] The publication describing the ‘PECAP’ algorithm that this project is based on:
Garcia, C., Goehring, T., Cosentino, S. et al. The Panoramic ECAP Method: Estimating Patient-Specific Patterns of Current Spread and Neural Health in Cochlear Implant Users. JARO 22, 567–589 (2021).
https://doi.org/10.1007/s10162-021-00795-2

[3] The publication describing the de-noising technique:
Kung F. J. (2025). An Integrated Spatial-Spectral Denoising Framework for Robust Electrically Evoked Compound Action Potential Enhancement and Auditory Parameter Estimation. Sensors (Basel, Switzerland), 25(11), 3523.
https://doi.org/10.3390/s25113523

Since the landmark detection of GW170817, gravitational wave astronomers have been searching for additional "bright sirens" - gravitational wave events with electromagnetic counterparts. However, LIGO's frequentist methodology using false alarm rates may have missed signals buried just below detection thresholds. These subthreshold signals could provide crucial cosmological information, including independent measurements of the Hubble constant to address the Hubble tension.

Bayesian model comparison offers a principled framework for assessing whether multi-messenger information (such as gamma-ray burst sky localizations) can help identify gravitational wave signals that would otherwise be classified as noise. Recent advances in GPU-accelerated inference using JAX-based tools like BlackJAX and JimGW make this computationally tractable.

The student will develop GPU-accelerated nested sampling pipelines to search for subthreshold gravitational wave signals by conditioning on sky localizations from gamma-ray bursts. Specific tasks include:

1. Learning the fundamentals of gravitational wave data analysis and Bayesian nested sampling
2. Implementing GPU-accelerated inference using JAX, BlackJAX, and JimGW 
3. Generating realistic mock gravitational wave data using BILBY and injecting signals at various signal-to-noise ratios
4. Comparing Bayesian evidence between unconditioned and sky-conditioned models
5. Applying the pipeline to real LIGO data coincident with short gamma-ray bursts from the Fermi GRB catalogue

A successful outcome would be a validated pipeline capable of identifying candidate subthreshold signals and quantifying the evidential support for their astrophysical origin. The project uses mathematical skills in probability theory, Bayesian inference, and high-dimensional integration (nested sampling). Programming skills in Python and familiarity with numerical computing are essential.

The Hubble tension - the 4-6σ disagreement between early-universe (CMB) and late-universe (supernovae) measurements of the cosmic expansion rate H₀ - is one of the most pressing problems in modern cosmology. Gravitational wave "standard sirens" offer a completely independent method to measure H₀, as the luminosity distance is encoded directly in the gravitational wave signal amplitude.

GW170817, the first binary neutron star merger with an electromagnetic counterpart, provided the first standard siren measurement of H₀. However, standard parameter estimation pipelines require thousands of CPU hours, limiting rapid multi-messenger follow-up. GPU-accelerated inference using JAX can reduce this to minutes, enabling real-time cosmological constraints from future detections.

The student will develop and validate a GPU-accelerated pipeline for joint inference of gravitational wave source parameters and the Hubble constant. Specific tasks include:

1. Understanding the theoretical framework connecting gravitational wave observations to cosmological distance measurements
2. Implementing joint parameter estimation for GW source properties and H₀ using JAX-based tools
3. Validating the pipeline on the benchmark event GW150914
4. Reproducing the H₀ posterior from GW170817 and comparing with published LVK collaboration results
5. Investigating prior sensitivity by comparing direct sampling approaches with reweighting techniques
6. Benchmarking GPU vs CPU performance

A successful outcome would be a validated, fast pipeline for H₀ inference that reproduces published results while providing insight into statistical methodology. The project uses mathematical skills in Bayesian inference, cosmology, and high-dimensional sampling. Strong programming skills in Python are essential.

Transmissible cancers are infectious tumours that spread through animal populations as clonal allografts [1]. Three are known to affect mammals: Canine Transmissible Venereal Tumour (CTVT) [2], and two forms of Tasmanian Devil Facial Tumour disease (DFT1 and DFT2) [3,4]. Several more forms have been discovered as disseminated neoplasias in marine bivalves [5].

One common feature of transmissible cancers is that they have outlived the progenitor that founded them. Subsequently they have been exposed to new, temporary hosts of diverse genotype and microbiota. DNA from one such host has been permanently incorporated into a transmissible cancer genome on at least one occasion [6].

However, we don’t know the extent to which microbial and viral material has been included. This project aims to assess this through the analysis of several high-coverage, short read sequencing data derived from CTVT, DFT1 and DFT2 tumour samples, and from normal tissue obtained from their matched hosts.

1. Ní Leathlobhair M, Lenski RE. Population genetics of clonally transmissible cancers. Nat Ecol Evol. 2022;6: 1077–1089.
2. Murchison EP, Wedge DC, Alexandrov LB, Fu B, Martincorena I, Ning Z, et al. Transmissible [corrected] dog cancer genome reveals the origin and history of an ancient cell lineage. Science. 2014;343: 437–440.
3. Pye RJ, Pemberton D, Tovar C, Tubio JMC, Dun KA, Fox S, et al. A second transmissible cancer in Tasmanian devils. Proc Natl Acad Sci U S A. 2016;113: 374–379.
4. Hawkins CE, Baars C, Hesterman H, Hocking GJ, Jones ME, Lazenby B, et al. Emerging disease and population decline of an island endemic, the Tasmanian devil Sarcophilus harrisii. Biol Conserv. 2006;131: 307–324.
5. Metzger MJ, Villalba A, Carballal MJ, Iglesias D, Sherry J, Reinisch C, et al. Widespread transmission of independent cancer lineages within multiple bivalve species. Nature. 2016. Available: http://www.nature.com/nature/journal/vaop/ncurrent/full/nature18599.html
6. Gori K, Baez-Ortega A, Strakova A, Stammnitz MR, Wang J, Chan J, et al. Horizontal transfer of nuclear DNA in transmissible cancer. Proc Natl Acad Sci U S A. 2025;122: e2424634122.

This project will investigate whether we can detect the inclusion of foreign DNA in the genomes of CTVT, DFT1 and DFT2. In the case of CTVT, we are already aware of a novel retrovirus that has implanted in the genome, and that papillomavirus is a common coinfection. The project will initially focus on detecting and characterising these viruses, before expanding to look for evidence of acquisition from any source.

Sequence that has inserted into the genome is of particular interest; a challenging part of the project will be to determine whether foreign DNA is integrated into the tumour genome, or is instead associated with the cellular environment. The project uses high-coverage, short read sequencing data derived from CTVT, DFT1 and DFT2 tumour samples, and from normal tissue obtained from their matched hosts. It will develop skills in bioinformatics, especially in handling genomic data, running sequence assembly, and comparing and annotating sequence contigs. Sequence analysis is an inherently mathematical topic. Assembling contigs is a longest common substring problem: by representing short sequence fragments as nodes in a graph in which the edges represent their overlaps, this amounts to finding the Euler path through the graph [7]. Sequence evolution models changes in sequence as a continuous time Markov process operating over a tree [8]. The large amount of data produced by sequencing projects necessitates use of statistics to simplify and identify patterns in the data. This project, though biological, has scope for the application and development of mathematical techniques, and would be of interest to a mathematician.

7. Pevzner PA, Tang H, Waterman MS. An Eulerian path approach to DNA fragment assembly. Proc Natl Acad Sci U S A. 2001;98: 9748–9753.
8. Felsenstein J. Evolutionary trees from DNA sequences: a maximum likelihood approach. J Mol Evol. 1981;17: 368–376.

The biological motivation of the project is the study of the formation and maintenance of the apical hook, including its response to biochemical perturbations [4], as well as the gravitropic behavior of shoots and roots [5,6]. From a mathematical perspective, the project aims to develop and implement numerical methods for integrating continuum morphoelastic models of rod-like plant organs.

The project will focus on the design and implementation of a numerical scheme to integrate the model proposed in [7]. By assuming morphoelastic roles for morphogens, the model directly links axial growth and bending to spatial morphogen distributions within a quasi-static formulation of growth mechanics. The organ is represented as an array of concentric cylindrical morphoelastic shells connected in parallel, yielding a mathematically rich framework that couples elasticity, growth, and geometry.

To relate this continuum description to cellular processes, gene regulatory networks, and active morphogen transport, the model will be discretized and implemented within an existing custom numerical solver for plant morphodynamics written in C++ [8].

The student will be hosted within the Jönsson group, led by Professor Henrik Jönsson, Director of the Sainsbury Laboratory. The group offers a collaborative and intellectually stimulating research environment, with weekly group meetings in which members present and discuss ongoing research.

The student will receive regular one-on-one supervision through weekly meetings with Dr. Amir Porat, with additional guidance and support available as needed throughout the internship.

The Sainsbury Laboratory also provides a vibrant and welcoming environment for summer students, hosting a range of social events and activities designed to foster interaction among students from different programs and research backgrounds.

Plant morphogenesis arises from the interplay of growth, mechanics, and genetic regulation. A central mathematical challenge is understanding how spatially distributed biochemical morphogens generate stable growth patterns. This project focuses on plant organs exhibiting self-similar growth, where global geometry is preserved as cells grow and divide.

Such systems exhibit approximately stationary material flows in suitable coordinates, making them analytically tractable [1]. We have developed a novel continuous geometric framework for self-similar growing manifolds, in which growth is described by the Lie derivative of a Lagrangian metric [2,3]. Under the assumption of shearless growth along orthogonal curvilinear coordinates, the Lie derivative simplifies, providing a compact and powerful description of tissue deformation and growth kinematics.

The project is split into two complementary tracks, with the option to focus primarily on one while exploring the other as time permits.

Track 1: 3D Cellularization of Self-Similar Growth (Numerical Modelling)
Extending an existing 2D C++ implementation of self-similar growth in the shoot apical meristem (SAM) [4] to three dimensions. This involves modelling cellular growth and division, developing algorithms to match simulated cell shapes and sizes to live-imaging data [5], and using the resulting geometries to analyse gene patterning and tissue mechanics.

Track 2: Geometric Perturbations of Self-Similar Growth (Analytical Modelling)
Investigating geometric perturbations of self-similar growth, motivated by phyllotaxis - the periodic formation of leaves and flowers on the flanks of the SAM [6]. Starting from the Lie derivative of the Lagrangian metric in self-similarity, we wish to examine how combined perturbations of the velocity and metric fields can generate biologically relevant growth patterns. The work begins with interpreting custom axisymmetric shearless perturbations of shells using the Lie derivative formalism, followed by periodic or more general perturbations, with the aim of establishing a new geometric framework for studying shape dynamics and morphogenesis.

The student will be hosted within the Jönsson group, led by Professor Henrik Jönsson, Director of the Sainsbury Laboratory. The group offers a collaborative and intellectually stimulating research environment, with weekly group meetings in which members present and discuss ongoing research.

The student will receive regular one-on-one supervision through weekly meetings with Dr. Amir Porat, with additional guidance and support available as needed throughout the internship.

The Sainsbury Laboratory also provides a vibrant and welcoming environment for summer students, hosting a range of social events and activities designed to foster interaction among students from different programs and research backgrounds.

Our general long-term aim is to develop an ‘inference engine’ using machine learning methods that extracts interpretable mechanistic laws directly from biological data to bridge the gap between modern large biological datasets and mechanistic biophysical models. This process includes two stages: (i) learning a ‘meaningful’ representation of the data (e.g. using unsupervised learning [1, 2]) and (ii) inferring connections and ultimately physical models from this learned representation [3, 4]. We recently demonstrated an example of this approach (LMRL25, ICLR workshop [5]). Depending on the student’s interests this project can focus on either or both of these complementary strands:

Strand 1: Data Alignment. Biological data is captured across different modalities (morphology vs. gene expression). This strand focuses on the "representation" problem: using unsupervised methods (e.g. Matrix Factorisation [1] or Autoencoders [2]) to project high-dimensional data into a unified, low-dimensional "latent manifold" that represents the tissue's state.

Strand 2: Differentiable Simulations & Model Inference. This strand focuses directly on the "inference" problem. Given a representation of the tissue state and its dynamics, we will use differentiable simulations (e.g. using Graph Neural Networks [6]) combined with identifiability techniques [3, 4] to infer possible governing equations.

[1] Argelaguet, Ricard, et al. "MOFA+: a statistical framework for comprehensive integration of multi-modal single-cell data." Genome biology 21 (2020).
[2] Yang, Karren Dai, et al. "Multi-domain translation between single-cell imaging and sequencing data using autoencoders." Nature communications 12.1 (2021).
[3] Maddu, Suryanarayana, et al. "Learning physically consistent differential equation models from data using group sparsity." Physical Review E 103.4 (2021).
[4] Brunton, Steven L., Joshua L. Proctor, and J. Nathan Kutz. "Discovering governing equations from data by sparse identification of nonlinear dynamical systems." Proceedings of the national academy of sciences 113.15 (2016).
[5] Zardilis, Argyris, Alexandra Budnikova, and Henrik Jönsson. "Learning a mechanical growth model of flower morphogenesis." LMRL25, ICLR workshop.
[6] Choi, Jeongwhan, et al. "Gread: Graph neural reaction-diffusion networks." International conference on machine learning. PMLR, 2023.
Refahi, Yassin, Zardilis, Argyris et al. "A multiscale analysis of early flower development in Arabidopsis provides an integrated view of molecular regulation and growth control." Developmental Cell 56.4 (2021).
Wang, Hanchen, et al. "Scientific discovery in the age of artificial intelligence." Nature 620.7972 (2023): 47-60.
Villoutreix, Paul. "What machine learning can do for developmental biology." Development 148.1 (2021): dev188474.