YOLO-PeakDetect: A Convolutional Neural Network for Automatic Analysis of Irregular Bands in Gel Electrophoresis

Gel electrophoresis (GE) is a critical tool in the fields
of molecular
biology and biopharmaceuticals. However, the current analysis software
requires extensive manual adjustments and cannot accurately determinate
particle-like bands of virus vector aggregates and monoclonal antibody
(mAb) or protein precipitates in GE. Herein, we developed a convolutional
neural network of YOLO-PeakDetect based on the You-only look-once
(YOLO) network architecture. Comparative results demonstrate that
on the simulated data set, the YOLO-PeakDetect network attains the
average precision 50–95 (AP50_95) of 0.9717, which is remarkably
superior to those obtained by the traditional algorithm and the existing
CNN-based peak detection model. In the experimental data set evaluated
against manually precisely annotated GE peak profiles, YOLO-PeakDetect
outperforms the traditional algorithm for both single peaks and overlapping
peaks. Meanwhile, in GE experiments employing three standard proteins,
the proposed network elevates the average linear correlation coefficient
from 0.9883 (achieved by ImageJ) to 0.9952. Particularly, the network
could detect the aggregate ratio of particle-like adeno-associated
virus (AAV) band in the sample well, full and empty AAV capsid from
the crescent-like vector bands, and the precipitate ratio of particle-like
bands of instable mAb accumulation. All the data showcased that the
network model achieves significant improvements in the accuracy, noise
resistance, and automation level of irregular band detection, providing
a reliable and intelligent solution for particle-like bands of virus
vector aggregates and instable mAb precipitates as well as regular
bands of proteins in GE.