Machine Learning-Based Pipette Positional Correction for Automatic Patch Clamp In Vitro

Abstract

Patch clamp electrophysiology is a common technique used in neuroscience to understand individual neuron behavior, allowing one to record current and voltage changes with superior spatiotemporal resolution compared with most electrophysiology methods. While patch clamp experiments produce high fidelity electrophysiology data, the technique is onerous and labor intensive. Despite the emergence of patch clamp systems that automate key stages in the typical patch clamp procedure, full automation remains elusive. Patch clamp pipettes can miss the target cell during automated experiments because of positioning errors in the robotic manipulators, which can easily exceed the diameter of a neuron. Further, when patching in acute brain slices, the inherent light scattering from non-uniform brain tissue can complicate pipette tip identification. We present a convolutional neural network (CNN), based on ResNet101, to identify and correct pipette positioning errors before each patch clamp attempt, thereby preventing the deleterious effects of and accumulation of positioning errors. This deep-learning-based pipette detection method enabled superior localization of the pipette within 0.62 ± 0.58 μm, resulting in improved cell detection success rate and whole-cell patch clamp success rates by 71% and 59%, respectively, compared with the state-of-the-art cross-correlation method. Furthermore, this technique reduced the average time for pipette correction by 81%. This technique enables real-time correction of pipette position during patch clamp experiments with similar accuracy and quality of recording to manual patch clamp, making notable progress toward full human-out-of-the-loop automation for patch clamp electrophysiology.

Keywords: CNN; automated; deep learning; electrophysiology; machine learning; patch clamp.

Graphical abstract

Figure 1.

A, Schematic of the error nomenclature used in this work. B, Example images of the CNN identifying the pipette tip over a brain slice. C, Error distribution of neural network testing dataset n = 300 images (red) compared with the true pipette error after moving to the cleaning bath for n = 32 images (blue). D, Convergence of true pipette error magnitude using the CNN as the measurement feedback in the (top) xy-plane and (bottom) z direction. The black dotted lines indicate appropriate error ranges for patch clamp experiments, at 2.5 mm in the xy-plane (half the diameter of a typical cell) and 3 mm in the z direction. The box width indicates the first and third quartiles, the white line indicates the median, and the whiskers of the box plot indicate the most extreme, non-outlier data points. E, Spatial representation of pipette tip locations after the second iteration of using the CNN for correction in the (top) xy-plane and (bottom) z direction. Black dotted lines indicate the range of one and 2 SDs.

Figure 2.

Comparing (A) pipette detection, cell detection, and whole-cell success rates and (B) time required for cross-correlation and CNN methods (n = 32 and = 36, respectively). The box width indicates the first and third quartiles, the white line indicates the median, and the whiskers of the box plot indicate the most extreme, non-outlier data points. Using Fischer's exact test; *p ≤ 0.05, ***p ≤ 0.001.

Figure 3.

A, Example image of a pipette on a neuron during a whole-cell recording. B, Distributions of access resistance, membrane capacitance, and membrane resistance for n = 23 successful whole-cell patch clamp recordings using the CNN. The white lines indicate the median, the width of the boxes indicates the first and third quartiles, and the whiskers indicate the range of the data. C, Representative current clamp trace with current injection. D, Representative voltage clamp trace.