ICON: an adaptation of infinite HMMs for time traces with drift

TL;DR

This paper introduces an adaptation of the infinite hidden Markov model (iHMM) that effectively accounts for drift in single molecule time traces, enabling automatic, unbiased analysis without pre-processing or manual model selection.

Contribution

The authors develop a Bayesian method coupling iHMM with a continuous control process to handle drift, allowing simultaneous learning of drift, states, and model parameters across multiple traces.

Findings

Successfully models data with drift without pre-processing.

Eliminates user-dependent trace selection and post-processing.

Automatically learns the number of states and drift parameters.

Abstract

Bayesian nonparametric methods have recently transformed emerging areas within data science. One such promising method, the infinite hidden Markov model (iHMM), generalizes the HMM which itself has become a workhorse in single molecule data analysis. The iHMM goes beyond the HMM by self-consistently learning all parameters learned by the HMM in addition to learning the number of states without recourse to any model selection steps. Despite its generality, simple features (such as drift), common to single molecule time traces, result in an over-interpretation of drift and the introduction of artifact states. Here we present an adaptation of the iHMM that can treat data with drift originating from one or many traces (e.g. FRET). Our fully Bayesian method couples the iHMM to a continuous control process (drift) self-consistently learned while learning all other quantities determined by the iHMM (including state numbers). A key advantage of this method is that all traces -regardless of drift or states visited across traces- may now be treated on an equal footing thereby eliminating user-dependent trace selection (based on drift levels), pre-processing to remove drift and post-processing model selection on state number.