DeepGANnel: Synthesis of fully annotated single molecule patch-clamp data using generative adversarial networks

• DOI: 10.1371/journal.pone.0267452
• Publication Date: 2022-05-10T17:24:27Z
• Authors: Sam T. M. Ball; Numan Celik; Elaheh Sayari; Lina Abdul Kadir; Fiona O’Brien; Richard Barrett-J...
• Keywords: Leverage (statistics); Sample (material); Generative model; Synthetic data

Development of automated analysis tools for “single ion channel” recording is hampered by the lack available training data. For machine learning based tools, very large sets are necessary with sample-by-sample point labelled data (e.g., 1 sample every 100microsecond). In an experimental context, such human supervision, and whilst this feasible simple analysis, it infeasible to generate enormous datasets that would be a big approach using hand crafting. work we aimed develop methods simulated channel free from assumptions prior knowledge noise underlying hidden Markov models. We successfully leverage generative adversarial networks (GANs) build end-to-end pipeline generating unlimited amount small, annotated “seed” record, needs no theoretical dynamical properties. Our method utilises 2D CNNs maintain synchronised temporal relationship between raw idealised record. demonstrate applicability 5 different sources show authenticity t-SNE UMAP projection comparisons real synthetic The model easily extendable other time series requiring parallel labelling, as ECG signals or nanopore sequencing