Unsupervised Drift Compensation based on Information Theory for Single-Molecule Sensors

Abstract

Single-molecule sensors record molecular dynamics in the form of time series representing a stochastic trajectory in the space of molecular conformational states. One of the main challenges in modeling this trajectory is to suppress the sensor baseline drift, especially in high-throughput and long acquisitions that are characteristic of single-molecule field-effect transistor sensors (smFETs). In this work, a multiscale signal compression technique based on the minimum description length principle, combined with an adaptive piece-wise cubic interpolation, is implemented to address the problem of baseline modeling through a blind source separation framework. Tests on simulated single-molecule traces over a large space of parameters (kinetics, noise, drift) show that the proposed algorithm accurately estimates the sensor baseline, including for signals with high and mixed noises, concept drifts, and various shapes and rates of baseline drifts, without prior knowledge of the sensor parameters or molecular kinetics.