Researchers from TU Delft, its spinoff company SoundCell and Reinier Haga MDC have shown that graphene “nanodrums” combined with machine learning can identify bacteria and determine their antibiotic susceptibility from the nanomotion of single cells within a couple of hours. The approach unifies bacterial identification and antimicrobial susceptibility testing (AST) in one label-free measurement at the single-cell level.
Each nanodrum consists of a bilayer graphene membrane less than 1 nanometer thick, suspended over an 8 micrometer-wide cavity that can host a single bacterium. When a living cell adheres to the drum, its intrinsic motions drive nanoscale vibrations of the graphene, which are read out optically as a time-dependent signal. This configuration avoids ensemble averaging and captures the mechanical behavior of individual bacteria.
Instead of analyzing noisy time-domain traces, the nanomotion signals are converted into time-frequency spectrograms that preserve both spectral and temporal information. These spectrograms are used as inputs to machine learning models - Convolutional Neural Networks (CNNs) and Support Vector Machines (SVMs) - which learn discriminative features directly, without manual feature engineering. In total, 456 measurements were used for species identification and 347 for susceptibility testing, with performance assessed using accuracy, sensitivity, specificity and ROC-based metrics.
Using this framework, the researchers successfully differentiated Escherichia coli, Staphylococcus aureus and Klebsiella pneumoniae while simultaneously distinguishing resistant from susceptible strains. Species identification reached accuracies up to 88%, and susceptibility testing for meropenem-resistant versus susceptible E. coli achieved up to 98.6% accuracy, corresponding to 98% precision for resistant/susceptible classification. CNNs performed best for susceptibility profiling, while SVMs offered more stable generalization across species for identification.
Because the models operate on single-cell nanomotion, the method delivers both ID and AST in 1-2 hours, without additional culturing after loading bacteria onto the graphene drums. With further development - such as cartridge-based, parallel sensor arrays and training datasets enriched in resistant strains - this graphene-ML platform could evolve into a powerful clinical tool for rapid, label-free diagnostics and for probing antimicrobial resistance and cellular biophysics at the single-cell level.
