Analyzing protein dynamics from fluorescence intensity traces using unsupervised deep learning network

Abstract

We propose an unsupervised deep learning network to analyze the dynamics of membrane proteins from the fluorescence intensity traces. This system was trained in an unsupervised manner with the raw experimental time traces and synthesized ones, so neither predefined state number nor pre-labelling were required. With the bidirectional Long Short-Term Memory (biLSTM) networks as the hidden layers, both the past and future context can be used fully to improve the prediction results and can even extract information from the noise distribution. The method was validated with the synthetic dataset and the experimental dataset of monomeric fluorophore Cy5, and then applied to extract the membrane protein interaction dynamics from experimental data successfully.

Toward high-throughput oligomer detection and classification for early-stage aggregation of amyloidogenic protein

“…Classification time for a given trace was on a sub-millisecond timescale, opening the door to real-time analysis of oligomer distributions during highthroughput experiments. In comparison to the literature study (Xu et al, 2019;Yuan et al, 2020), the accuracy of our deep learning methods is similar, but we have demonstrated that our approach is applicable to even larger oligomers. Furthermore, our approach is quite flexible since photobleaching and imaging noise are physical phenomena well understood, making simulation a valid way to train a model and obtain training parameters before applying them to experimental photobleaching traces.…”

“…We note that the well-known Chung-Kennedy algorithm ( Chung and Kennedy, 1991 ) and the Hidden Markov model (HMM) analysis ( Messina et al, 2006 ) are computationally expensive. The application of machine learning approaches in single-molecule analysis has gained considerable interest recently ( Meng et al, 2022 ; Xu et al, 2019 ; Yuan et al, 2020 ; Thomsen et al, 2020 ; White et al, 2020 ; Li et al, 2020 ). A deep learning algorithm named DeepFRET has been demonstrated to reach classification accuracy on ground truth data by over 95%, not only outperforming human operators but also reducing the computation time by two orders of magnitude ( Thomsen et al, 2020 ).…”

“…Most notably, a convolutional and long-short-term memory deep learning neural network (CLDNN) used by Xu and co-workers ( Xu et al, 2019 ) achieved much higher accuracy and two orders of magnitude improvement in efficiency compared to HMM for small oligomers up to tetramer. Built upon the success of CLDNN, an unsupervised learning framework of the discriminator-generator network (DGN) was developed which outperforms both HMM and PIF ( Yuan et al, 2020 ). We adopted supervised learning approaches based on lightweight artificial neural networks ( Yu et al, 2019 ) as well as RBF-kernel support vector machines ( Huang et al, 2018 ), and these approaches demonstrated excellent capability to classify oligomers up to 19-mer, well beyond the capability of trace idealization algorithms (PIF, HMM etc.).…”

Deep learning in single-molecule imaging and analysis: recent advances and prospects

“…(b) The training and performances of DGN on both photobleaching event counting and dynamic finding with fluorescence intensity traces. 100 Copyright © 2020, The Author(s). 104 Copyright © 2020, The Author(s)…”

“…Copyright © 2019, American Chemical Society. (b) The training and performances of DGN on both photobleaching event counting and dynamic finding with fluorescence intensity traces 100. Copyright © 2020, The Author(s).…”