Enhancing Nanopore Sensor Performance Through Modern Neural Network Architectures : An Evaluation of Classification Performance

Sammanfattning

Nanopore-based sensors have revolutionized DNA-sequencing and are showing great potential in related fields. However, the sensors' high sensitivity often results in noisy readings, making signal processing challenging and leading to traditional methods falling short in handling the complexity and variability of the signals when performing classification. To address this, scientists are increasingly turning towards data-driven approaches, yet a knowledge gap remains regarding which approches to apply and how modern methods can provide deeper insights. A previous study attempted to bridge this gap by training a Convolutional Neural Network (CNN) by the name of QuipuNet, acheiving significant performance gains over traditional methods. Through the use of the publicly available dataset from said study, this thesis aims to enhance the performance of nanopore-based sensors further by testing modern neural network architectures. The project introduces and tests four promising architectures: Fully Convolutional Networks (FCN), Residual Networks (ResNet), Long Short-Term Memory FCN (LSTM-FCN), and Transformers. These models were optimized through random search and Bayesian optimization, with rigorous evaluation through cross-validation. The models were then assessed for Accuracy, Precision, Recall, and F1-score, as well as their efficiency in terms of compute time and memory utilization. Additionally, the inherent advantages of the models' architectures were discussed to provide a holistic comparison. The results showed that while QuipuNet still slightly outperforms the others in classification, ResNet and Transformer architectures offer significant versatility, handling variable input lengths and providing valuable interpretability through class activation and attention mapping. ResNet achieved an F1-score of 0.929, closely rivaling QuipuNet's 0.935, while the Transformer showed promising potential with an F1-score of 0.910. However, no statistically significant difference could be ascertained between those three models. Only the FCN and LSTM-FCN showed statistically significant lower classification performance compared to QuipuNet. The FCN and LSTM-FCN models, though showing lower classification performance, excelled in inference and training speed and memory efficiency. The study concludes that newer neural network architectures, particularly ResNet and Transformers, offer competitive performance and enhanced versatility for nanopore signal classification. These findings pave the way for future research to further optimize these models and explore their application in a wider range of nanopore-related tasks.