Nanobiomechanical data classified by Deep learning based on convolutional neural networks

Abstract Nanobiomechanical data have an interest in biomedical research, but the capability of deep learning (DL) based on convolutional neural networks (CNN) has not been explored to classify such data. We propose to use these strategies to treat nanobiomechanical data acquired by atomic force microscopy (AFM) on Candida albicans living cells, an opportunistic pathogenic micro-organism of medical interest. Data, acquired by force spectroscopy, allowed us to generate force vs. distance curves (FD curves) which its profile is linked to nanobiomechanical properties of C. albicans . DL was applied to classify FD curves, considered as images, into 3 groups: adhesive nanodomains, non-adhesive domains or in between domains. We achieved a real multiclass classification with a validation accuracy, macro-average of F1, and the weighted average of 92%, without the need to perform the usual dropout or weight regularisation methods. Transfer learning with a pre-trained (PT) VGG16 architecture with and without fine tuning (FT) permitted us to verify that our model is less computationally complex and better fitted. The generalisation was done by classifying on other C. albicans cells with more that 99% of confidence, to finally predict 16,384 FD curves in less than 90 seconds. This model could be employed by a non-machine learning specialist as the trained model can be downloaded to predict the adhesiveness, within seconds, on C. albicans cells characterized by AFM.