Abstract
Atomic force microscope (AFM) indentation allows high-resolution spatial characterization of biomechanical properties of cells and tissues. Rapid, reproducible, and quantitative analysis of AFM force curves has been challenging due to several technical limitations, such as excessive noise and uncertainty associated with contact-point determination. Here, we propose a novel machine-learning algorithm composed of convolutional bidirectional long short-term memory neural networks called Convolutional Bidirectional Recurrent Architecture (COBRA) that can reliably process raw AFM elastography data, triage poor-quality curves, and accurately identify the contact point without any a priori knowledge of underlying material properties. Using over 5000 manually curated force curves on seven different healthy and diseased cell types, we trained several regression and classification algorithms to compare their utility. In contrast to classical analytical or semi-quantitative techniques and other machine-learning methods, the COBRA approach identified low-quality or anomalous indentation events better, with an area under the curve of 0.92, and it estimated the contact point with the minimal absolute error of 28 ± 3 nm and pointwise elastic modulus with mean absolute percentage error of 5.3% ± 0.7%. The method was also successful in identifying the contact point in independently acquired AFM data from the literature with divergent probes and substrates. In conclusion, our method can rapidly filter low-quality AFM force curves and automatically process raw indentation data with the lowest error levels, allowing high-throughput analyses with increased precision and reproducibility.