Nano Trees: Machine Learning Enhanced Signal Processing of Nanopore Data

Abstract

Solid-state nanopores are molecular-size holes in ultra-thin dielectric membranes used for single-molecule sensing. Nanopore sensors are operated by immersing them in a salt solution and applying an electric potential to drive the electrophoretic motion of ions and the capture of biomolecules. In a solid-state nanopore experiment, as a biomolecule passes through the nanopore, it produces fluctuations in the electrical current measured across the nanopore. These fluctuations can be used to determine various properties of the biomolecule and the nature of this translocation. As the complexity of these experiments increases, analysis of the resulting electrical signals to determine biomolecular details becomes a challenge. Piecewise constant function fitting is an important part of this analysis which is used to extract sublevel details from these signals. In an ideal world with perfect measuring devices, sublevel transitions would be instantaneous and perfectly captured without any noise in the signals. However, in practical scenarios, measurement imperfections and noise necessitate sophisticated fitting techniques to accurately interpret the sublevel transitions and deduce the underlying biomolecular properties.

Current techniques for this task perform poorly when fitting this data prompting the development of more robust and efficient methods. In this thesis, I address this challenge through a novel Nano Trees algorithm for fitting piecewise constant functions. Nano Trees leverages machine learning algorithms to provide accurate fits to the noisy piecewise constant data characteristic of nanopore signals, producing accurate fits on events as short as twice the response time of the measurement system. I analyze the performance of our algorithm on several real and synthetic datasets. These results produced using Nano Trees underscore the generalizability and accuracy of this approach in the regime of fast molecular event fitting.

Finally, I provide several approaches to understanding and fine-tuning various hyperparameters of Nano Trees. These include debugging procedures, flowcharts, isolated hyperparameter tuning instructions, default values for hyperparameters, and more. As the algorithm is novel and complex, I also discuss various automation and implementation strategies aimed at improving its practical applicability, usability, and overall effectiveness in real-world settings.