Headline Summary

Master's dissertation at Imperial College London — built Transformer-based sequence classifiers for real-time pathogen detection from LAMP signals, achieving 94.72% accuracy on amplification curves (outperforming KNN baseline of 91.33%) and 100% on melt curves across 5 respiratory pathogens including SARS-CoV-2.

What I Built

An end-to-end classification pipeline for multiplex LAMP (loop-mediated isothermal amplification) signals, comparing feature-engineered classical ML baselines against Transformer sequence models. The work demonstrates that Transformers can classify raw time-series biosensor data without manual feature extraction, with practical implications for point-of-care diagnostic devices.

Tech stack: Python · PyTorch · Transformers · Scikit-learn · NumPy

Problem

Multiplex LAMP enables testing for multiple pathogens in a single sample, but classification is difficult: curve shapes vary across targets and experimental conditions, signals are noisy, and classes partially overlap. Two signal types were studied — amplification curves (fluorescence vs. time, cheaper to collect) and melt curves (fluorescence vs. temperature, more discriminative but costlier).

Targets: Influenza A, Influenza B, SARS-CoV-2, Human adenovirus, Klebsiella pneumoniae

Approach

Classical Baselines

• Amplification curves → sigmoid parameter fitting (5 features) → KNN, Random Forest
• Melt curves → peak temperature extraction → Random Forest

Transformer Sequence Models

• Raw signal fed directly as a sequence — no manual feature extraction
• Positional encoding to preserve temporal ordering
• Hyperparameter-tuned (model dimension, attention heads, layers, dropout)