Efficient detection of opening and closing events in recordings of small numbers of ion channels is difficult. The identification of such events is an example of the more general statistical problem of the detection of changepoints in a time series. The simplest methods for tackling this involve filtering and thresholding, but these perform badly when current levels are closely spaced, when the intervals between events are short, or when there is substantial noise. More complex methods such as hidden Markov modelling have been successfully used in the analysis of patch-clamp recordings, but these impose quite detailed assumptions about the underlying physical process and have, so far, only been implemented using the assumption that successive data points are independent events, i.e. that the data has not been filtered.
We have adopted an alternative strategy. The signal is modelled as a sequence of filtered transitions between constant levels with no further assumptions about the process underlying the transitions. The investigator must specify the spectrum of the noise, the characteristics of the data acquisition filter, and initial guesses (strictly prior distributions) of the frequency of events and their magnitudes. This information is sufficient for a Bayesian statistical analysis of the data. Locations of changepoints and currents are chosen using Markov chain Monte Carlo sampling of their posterior distribution. By fitting this type of model to the data one achieves a better trade-off between sensitivity and reliability of detection compared to thresholding methods. Examples will be displayed of estimation of closely spaced current levels and the analysis of records with flicker type events.