Imperial College London

Programme Overview

MSc in Biomedical Engineering with specialisation in Computational Bioengineering at Imperial College London. The programme bridges engineering, computing, and biology, covering everything from how organs work as physical systems to building machine learning models for drug discovery and neuroscience.

Autumn Term

Systems Physiology

• Modelling organ systems mathematically: kidney, heart, lungs, muscles, digestive system, endocrine system, nervous system
• Engineering and biophysics perspective rather than descriptive anatomy

Data Analysis for Research

• Statistical methods for bioengineering research, implemented in Python
• Hypothesis testing, ANOVA, Bayesian inference, generalised linear models, bootstrapping, Monte Carlo, PCA
• Focus on choosing the right method and understanding its assumptions

Digital Biosignal Processing

• Discrete-time signal processing applied to biomedical signals
• Sampling theory, DFT, z-transform, FIR/IIR filter design, power spectral density estimation

Reinforcement Learning for Bioengineers

• Markov decision processes, Bellman equation, dynamic programming, tabular RL, deep RL
• Continuous and high-dimensional action spaces; implemented in Python and PyTorch
• Applications in robotics and neuroscience contexts
• Final project: ProteinTuneRL: A Reinforcement Learning Perspective on Antibody Design

Spring Term

Brain Machine Interfaces

• Recording technologies (ECoG, EEG, MEG), their trade-offs, and decoder theory
• Clinical applications: restoring movement in paralysed patients, deep brain stimulation for Parkinson's
• Lab sessions decoding neurophysiological signals in MATLAB

Artificial Intelligence for Drug Discovery

• Drug discovery pipeline, chemoinformatics, and bioinformatics fundamentals
• Where AI goes wrong: performance misestimation, unrealistic benchmarks, class imbalance, sparse and high-dimensional data
• Critical evaluation of AI claims in the biomedical literature

Image Processing

• Image transforms, neighbourhood operators, segmentation, registration, and image synthesis
• Data-driven approaches; applications across clinical imaging and research

Biomimetics

• How biological solutions inspire engineering, from molecular to systems scale
• Bio-inspired adhesives, materials, sensing systems, structural colour, and parallels between biological and artificial vision

Medical Device Certification

• EU regulatory framework, CE marking, and device classification
• Safety and hazard analysis, risk management, product development lifecycle, intellectual property

MSc Journal Club

• Weekly critical review and discussion of current research papers
• Presenting, critiquing, and debating primary literature as a group

Individual Project

Physics-Informed Operator Learning for Microbial Dynamics
Supervisor: Prof. Reiko Tanaka · Thesis submission: September 2026

Microbial ecosystems like the gut microbiome involve dozens of species interacting through complex nonlinear dynamics. The standard approach for modelling them with physics-informed neural networks (PINNs) requires retraining from scratch for every new dataset, which is expensive and doesn't scale. This project asks whether a single model can instead learn to infer the hidden biological parameters of any given microbial system directly from observed trajectories, then generalise to systems it has never seen.

The underlying physics is the generalised Lotka-Volterra (gLV) system, which describes how species grow, compete, and respond to perturbations like antibiotics. The project compares three approaches: a Baseline PINN retrained per dataset as a reference, PI-DION (a physics-informed deep inverse operator network trained once on synthetic gLV data that infers growth rates, interaction strengths, and perturbation sensitivities per system), and QPINN (an exploratory quantum-circuit variant implemented in PennyLane).

Generalisation is tested on real mouse gut microbiome data from Stein et al. (2013): nine trajectories across three populations, eleven interacting taxa, sparse and irregular sampling, and an antibiotic-induced C. difficile bloom in one population that makes it structurally unlike the others. Early results show the inverse mechanism works well on synthetic data (near-perfect growth rate recovery), but a synth-to-real transfer gap remains, which is what PI-DION's operator learning architecture is designed to address.