HoloMoA: a holography and deep learning tool for the identification of antimicrobial mechanisms of action and the detection of novel MoA.

We propose an innovative technology to classify the Mechanism of Action (MoA) of antimicrobials and predict their novelty, called HoloMoA. Our rapid, robust, affordable and versatile tool is based on the combination of time-lapse Digital Inline Holographic Microscopy (DIHM) and Deep Learning (DL). In combination with hologram reconstruction. DIHM enables a label-free, time-resolved visualization of bacterial cell morphology and quantitative phase map to reveal phenotypic responses to antimicrobials, while DL techniques are powerful tools to extract discriminative features from image sequences and classify them. We assessed the performance of HoloMoA on ATCC 25922 treated for up to 2 hours with 22 antibiotics representing 5 conventional functional classes (i.e. Cell Wall synthesis inhibitors, Cell Membrane synthesis inhibitors, Protein synthesis inhibitors, DNA and RNA synthesis inhibitors). First, using reconstructed phase images as input to a Convolutional Recurrent Neural Network (CRNN), we detected the MoA of known antibiotics with 95% accuracy. Secondly, we showed how our CRNN model combined with a Siamese Neural Network architecture can be used for the novelty assessment of the MoA of candidate antibiotics. We successfully evaluated our novelty detector on a test set containing three unseen molecules - two belonging to the conventional functional classes and one molecule from an additional class (Folate synthesis inhibitors, herein represented by trimethoprim-sulfamethoxazole). We demonstrated that the DIHM and DL combination provides a promising tool for determining the MoA of antimicrobial candidates using a large image database for known antimicrobials.