Development of constrictional microchannels and the recurrent neural network in single-cell protein analysis

Introduction: As the golden approach of single-cell analysis, fluorescent flow cytometry can estimate single-cell proteins with high throughputs, which, however, cannot translate fluorescent intensities into protein numbers. Methods: This study reported a fluorescent flow cytometry based on constrictional microchannels for quantitative measurements of single-cell fluorescent levels and the recurrent neural network for data analysis of fluorescent profiles for high-accuracy cell-type classification. Results: As a demonstration, fluorescent profiles (e.g., FITC labeled β-actin antibody, PE labeled EpCAM antibody and PerCP labeled β-tubulin antibody) of individual A549 and CAL 27 cells were firstly measured and translated into protein numbers of 0.56 ± 0.43 × 10⁴, 1.78 ± 1.0⁶ × 10⁶ and 8.11 ± 4.89 × 10⁴ of A549 cells (n_cell = 10232), and 3.47 ± 2.45 × 10⁴, 2.65 ± 1.19 × 10⁶ and 8.61 ± 5.25 × 10⁴ of CAL 27 cells (n_cell = 16376) based on the equivalent model of the constrictional microchannel. Then, the feedforward neural network was used to process these single-cell protein expressions, producing a classification accuracy of 92.0% for A549 vs. CAL 27 cells. In order to further increase the classification accuracies, as a key subtype of the recurrent neural network, the long short-term memory (LSTM) neural network was adopted to process fluorescent pulses sampled in constrictional microchannels directly, producing a classification accuracy of 95.5% for A549 vs. CAL 27 cells after optimization. Discussion: This fluorescent flow cytometry based on constrictional microchannels and recurrent neural network can function as an enabling tool of single-cell analysis and contribute to the development of quantitative cell biology.