Event detection (nanopore)

Event detection is the step of a nanopore analysis that finds, in a recorded current trace, the intervals in which something occupies the pore, and measures where each begins and ends and how far the current falls. The output is a table with one row per event, and the later steps of classification, counting and statistics work on that table. A single pore sampled at 200 kHz, as the company's stream is, produces 720 million samples an hour.

The company's guide gives the stage its name when it tells viewers of the live molecular stream: "Each dip is a candidate event, something passing through the sensing region."^[1] A detector proposes candidates. What the candidate is belongs to classification, and candidates that nothing identifies are the subject of Unidentified Signal Producing Event.

An event is a departure from the open-pore current, so the first problem is knowing that current. It does not stay put: the pore resistance and the event rate change over the course of an experiment, so the baseline has to be estimated locally. A moving average of the trace follows the drift but is dragged down by the events it contains. Plesa and Dekker's remedy is to detect events, remove them, rebuild an ideal event-free baseline from what remains and recompute, and with repetition the moving average and its standard deviation converge to their correct values.^[2] A 2025 program adjusts the threshold inside a sliding window so that oscillations of the baseline do not fake or hide events.^[3]

Some discontinuities are the instrument's own doing. The firmware of the MR1 Nanopore Reader has a default pulse of 800 mV for 250 ms for clearing a clogged pore, and the baseline after it need not be the baseline before.^[4]

The oldest detector flags the samples that fall more than a chosen multiple of the noise below the baseline. Single-channel physiologists set the threshold halfway between the open and closed levels and measured durations between crossings, and their treatment of the statistics of those records is the standard reference.^[5] For Gaussian noise a threshold of 3.8 standard deviations is crossed by chance by 7.2 samples in 100,000. At 100,000 samples a second that is seven false crossings every second if each sample is tested alone, and about one in 32 minutes if two consecutive samples must cross, a figure that assumes independent noise, which a filtered trace is not.

For a pulse of known shape in white noise, the best linear detector is the matched filter, which for a rectangular blockade is a moving average over its duration and improves the signal-to-noise ratio by the square root of the number of samples averaged.^[6] The durations are not known beforehand, so a detector needs several filters or one filter and a compromise, and a filter that suppresses noise also lowers the depth of the shortest events, as described under Dwell time.

E. S. Page proposed the cumulative-sum test in 1954 for catching a change in the mean of a production process. It adds up the deviations from the baseline and signals when the sum passes a limit, and the time at which the sum last left zero estimates when the change began.^[7] The OpenNanopore software of 2012 applied it to the multi-level events of molecules that pass in stages, finding each internal step.^[8] The MOSAIC package pairs a refined version, CUSUM+, for events long enough to reach a steady current, with ADEPT, which models the response of the electronics to characterize short events that never do.^[9]

A hidden Markov model treats the trace as noisy readings of a hidden state, such as open, blocked or one of several sublevels, that switches at random times. It was applied to single ion channels in 1990, and the Viterbi algorithm recovers the most probable path of states through a trace.^[10] Nanopore groups used the same models to tell which base pair closed a DNA hairpin in 2003, and a general toolkit followed in 2015.^[11][12]

In sequencing, learning removed the step. Basecalling used to begin by cutting the raw signal into segments that each stood for a short run of bases. Its authors described Chiron, in 2018, as the first deep network to translate the raw signal to bases without that segmentation.^[13] For single molecules, a denoising network trained on pairs of noisy recordings revealed corrugation signals from nanoparticles that the raw curves and frequency filtering did not show.^[14]

Detection and classification can shade into each other. The algorithm that first read the DNA barcodes of Bell and Keyser had twelve steps and more than a dozen hand-tuned parameters, and it could use about a fifth of the events. The QuipuNet network that followed it began by removing events whose charge deficit was more than two standard deviations from the mean or whose current drop was more than 3.2 times the unfolded level, which dropped up to 30% of the detections, and classified the rest.^[15]

A real recording comes without a list of its events, so detectors are often tested on simulated traces. The host program of the MR1 Nanopore Reader has a mock mode that "generates baseline ionic current & DNA translocation events", and an event record in its specification holds the start time, duration, baseline and blockade currents, depth, fractional blockade, area and sublevel count.^[4]

• Current blockade
• Dwell time
• Nanopore noise
• Machine learning on nanopore signals
• Signal-to-noise ratio